landslide susceptibility case study

• Reference Manager
• Simple TEXT file

People also looked at

Original research article, machine learning for high-resolution landslide susceptibility mapping: case study in inje county, south korea.

• 1 Department of Advanced Science and Technology Convergence, Kyungpook National University, Sangju, Republic of Korea
• 2 Faculty of Water Resources Engineering, Thuyloi University, Hanoi, Vietnam
• 3 Department of Forest Environment and Conservation, National Institute of Forest Science, Seoul, Republic of Korea

Landslides are a major natural hazard that can significantly damage infrastructure and cause loss of life. In South Korea, the current landslide susceptibility mapping (LSM) approach is mainly based on statistical techniques (logistic regression (LR) analysis). According to previous studies, this method has achieved an accuracy of approximately 75.2%. In this paper, we expand upon this traditional approach by comparing the performance of six machine learning (ML) algorithms for LSM in Inje County, South Korea. The study employed a combination of geographical data gathered from 2005 to 2019 to train and evaluate six algorithms, including LR, Gaussian Naive Bayes (GNB), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Random Forest (RF), and Extreme Gradient Boosting (XGB). The effectiveness of these models was measured by various criteria, such as the percentage of correct classification (PCC) score, F1 score, and Kappa score. The results demonstrated that the PCC and F 1 scores of the six models fell between [0.869–0.941] and [0.857–0.940], respectively. RF and XGB had the highest PCC and F 1 scores of 0.939 and 0.941, respectively. This study indicates that ML can be a valuable technique for high-resolution LSM in South Korea instead of the current approach.

1 Introduction

Landslides are a typical natural disaster that can occur anywhere on the earth and have devastating consequences, including loss of lives and harm to society and the economy ( Highland and Bobrowsky, 2008 ; Biswas et al., 2022 ). The Republic of Korea, hereafter referred to as South Korea, is particularly susceptible to landslide-induced catastrophes. In South Korea, landslides are primarily concentrated in the summer, when approximately 60% of the annual rainfall occurs due to the Asian Monsoon. Climate change has intensified this pattern, evidenced by the nearly doubled mean annual landslide-damaged area, from 276.6 ha/year in 1981–2000 to about 527.6 ha/year in 2001–2020 ( Korea Forest Service, 2021 ). Despite the great efforts, including installing check dams to prevent fatalities and property loss since 1986, several landslide disasters have often inflicted severe damage. One of the fatal landslides in 2011 killed 43 people due to landslide-induced debris flows in Seoul and Chuncheon, with an estimated damage of 824 ha ( Kim et al., 2011 ). Moreover, 2020 experienced a prolonged monsoon period of 54 days, the third-longest in the recorded weather history of South Korea, which resulted in 1,343 ha of damaged area and nine fatalities ( Lee et al., 2021 ).

Identifying landslide-prone areas and implementing measures to prevent or mitigate their impacts necessitates a practical landslide susceptibility assessment ( Huabin et al., 2005 ; Chikalamo et al., 2020 ). This evaluation includes an examination of numerous elements that might contribute to landslides, such as land use, rainfall, slope, rainfall, and geology ( Pourghasemi et al., 2018 ). The involvement of fluvial processes and the hydro-geomorphic characteristics they generate is one crucial component that might contribute to the danger of landslides ( Tran et al., 2022 ). These water-moving activities can change the terrain in ways that raise the danger of landslides. LSM is central to understanding the risk associated with landslides in various regions. By conducting an LSM, we can better understand the frequency and location of landslides and take steps to lessen the danger of these disasters ( Eker et al., 2015 ; Maes et al., 2017 ).

Over the past decades, researchers have developed a myriad of techniques to gauge the potential for landslides. However, constructing an effective LSM presents several challenges, primarily related to the completeness and quality of the incoming data ( Pardeshi et al., 2013 ; Chae et al., 2017 ). The effectiveness of an LSM significantly depends on the data quality it is based upon. Yet, such data might be marred by incompleteness, inaccuracies, or spatial inconsistencies, resulting in reduced map accuracy ( Thirugnanam et al., 2020 ; Caleca et al., 2022 ). Historically, diverse methods ranging from geomorphological assessments to statistical models have been employed to predict landslide-prone zones ( Cardinali et al., 2002 ). While pivotal, these methods have challenges, ranging from extensive ground truthing to dependency on consistent spatial data ( Lombardo et al., 2020 ; Shano et al., 2020 ). Due to their inherently dynamic nature, landslide-causing conditions can differ dramatically across temporal and spatial scales, making it arduous to pinpoint the likelihood of landslides in a given location ( Atkinson and Massari, 2011 ). There may also be a lack of resources and technology available to monitor areas continuously for potential landslides, particularly in remote or inaccessible regions ( Zêzere et al., 2017 ; Piciullo et al., 2018 ). This can make identifying the circumstances that could lead to a landslide on time challenging. Furthermore, the traditional statistical models used to construct LSMs are based on statistical relationships and assumptions that may not always hold in reality, resulting in uncertainty in the predictions made by the model ( Lee and Min, 2001 ; Huang et al., 2020 ).

LSM utilizes a variety of strategies, such as statistical methods and ML algorithms, to predict potential landslide zones ( Stanley et al., 2021 ; Rahman et al., 2022 ). One such statistical technique is LR analysis, which assesses the association between a dependent variable (e.g., occurrence or absence of landslides) and several independent elements (e.g., land use, slope, and geology) ( Lee, 2005 ; Woo et al., 2014 ). This technique is attractive due to its ability to handle multiple independent variables and its probabilistic outcome, offering a likelihood of occurrence for landslides in specific areas ( Reichenbach et al., 2018 ). Another frequently applied statistical method in LSM is bivariate statistical analysis ( Mersha and Meten, 2020 ). Bivariate analysis examines the relationship between two variables—landslides and slope inclination—to uncover patterns and trends. This analysis’s simplicity and ability to identify statistically significant associations between landslide occurrence and influencing factors make it a valuable tool in LSM ( Yalcin, 2008 ; Hong et al., 2019 ). These statistical techniques have garnered popularity due to their simplicity and capability to handle multiple variables, as they often operate under assumptions of linearity and independence among predictors. Such assumptions can restrict their predictive power in complex natural terrains ( Yalcin et al., 2011 ).

Recently, ML techniques have emerged as powerful tools in LSM due to their ability to analyze large volumes of data, identify hidden patterns and relationships that may not be evident to human analysts ( Lv et al., 2022 ; Wang et al., 2023 ). Common models employed for LSM include decision trees (DTs), RF, SVM, and neural networks ( Azarafza et al., 2021 ; Wang et al., 2021 ). DTs are a form of ML technique that includes building a tree-like model of decisions based on various variables or attributes. This model is designed to forecast the chance of landslides occurring across multiple territory regions ( Kadavi et al., 2019 ). RFs include ensembling a group of DTs and utilizing their aggregate forecasts to produce more accurate predictions about the risk of landslides happening ( Dou et al., 2019 ; Sahin et al., 2020 ). SVMs estimate the likelihood of landslides occurring in various parts of the region by locating the hyperplane in a high-dimensional space that maximum separates distinct classes ( Pham et al., 2016 ; Ye et al., 2023 ). These ML algorithms can be beneficial for detecting locations at high risk of landslides and developing strategies to mitigate this risk ( Youssef and Pourghasemi, 2021 ). Despite its efficacy, each algorithm presents its strengths and challenges. The success of these algorithms largely depends on the nature of the data and the research context ( Merghadi et al., 2020 ; Ado et al., 2022 ). Therefore, it is essential to compare the performance of several ML algorithms to identify the most effective approach for a given study area.

In South Korea, LSM relies on the statistical technique of LR analysis, which has an accuracy rate of approximately 75.2% ( Lee et al., 2015 ). Although LR has been helpful, it has certain limitations, mainly when modeling non-linear interactions. This has motivated the exploration of alternative ML techniques, each presenting unique strengths: GNB’s probabilistic approach, KNN’s adaptability to non-linear data boundaries, SVM’s prowess in high-dimensional spaces, and the ensemble capabilities of RF and XGB that efficiently capture complex data relationships. However, comprehensive comparative studies on these ML algorithms, especially in the South Korean setting, are limited due to past restrictions in accessing high-resolution data.

This study aims to fill this gap by investigating the performance of 6 ML algorithms for high-resolution LSM in Inje County, Gangwon, South Korea, an area characterized by steep terrain and a high frequency of landslides ( Lee et al., 2016 ). The novelty of this study lies in its comprehensive comparison of the performance of 6 ML algorithms, providing a more extensive analysis than previous research in the South Korean context. This study also leverages high-resolution data to provide a detailed evaluation of each algorithm’s performance, addressing previous limitations in data availability.

This article discusses the following important research questions: 1) What is the overall accuracy of the ML algorithms for predicting landslide susceptibility? 2) How do the ML algorithms perform for specific assessment criteria like recall, precision, and F 1 score? 3) How do the ML algorithms perform across different landslide susceptibility classes? 4) How does the effectiveness of the ML algorithms compare to that of other methods, such as the traditional statistic method? 5) What are the critical factors influencing the performance of the ML techniques in the research area?

By identifying the most effective ML algorithms for LSM in South Korea, this study aims to contribute to developing more accurate and reliable landslide risk assessment models. These can underpin informed decision-making in disaster management, ultimately mitigating the impacts of landslide disasters. The remainder of the article is structured as follows: The Material and Methods section presents the data sources, the selection and implementation of the ML algorithms, and the evaluation of algorithm performance. The Results section describes the research results, and the Discussion section discusses the implications of the findings. The Conclusion section summarizes the key findings and provides recommendations for further study. The schematic diagram of this study is illustrated in Figure 1 .

FIGURE 1 . Flowchart of this study.

2 Materials and methods

2.1 study area.

Landslides are ground movements caused by the movement of slope-forming elements comprising rock, soil, vegetation, and water downward and outward. These movements can range from slow creep to rapid and destructive slides, often activated by factors like precipitation, geological attributes, human activity, or a combination of these triggers.

This investigation is centered in Inje County, Gangwon Province, South Korea. Situated in the eastern segment of the Korean Peninsula, Inje County is distinctly characterized by its steep terrains, predisposing it to a heightened risk of landslides ( Yoo et al., 2009 ). These geophysical phenomena not only carry the potential to damage infrastructure severely but also to impede transportation networks and critically impact local community livelihoods ( Kadavi et al., 2019 ). Recently, Inje County has witnessed an uptick in the frequency and magnitude of landslide events, a pattern emerging from an intricate interplay of natural and anthropogenic influences. Intense meteorological events like heavy monsoons and shifts in land use patterns have amplified these landslide occurrences ( Lee et al., 2015 ). Notably, the county is especially susceptible to landslides triggered by heavy rainfall events, typical during the monsoon phase ( Lee et al., 2015 ). Figure 2 presents the study area.

FIGURE 2 . Location of the studied region.

For LSM, the input data quality and the selection of the appropriate modeling approach are among the components that have the most significant impact on the accuracy of the map ( Yalcin, 2008 ). Because landslides are the slope’s movement or the slope’s instability and their control parameters, the fundamental causes of landslides are typically connected to geology, hydrology, morphology, and anthropogenic activities ( Ayalew and Yamagishi, 2005 ). As a result, it is crucial to identify the causative conditions for landslides in prone areas. Despite this, there is no universally accepted technique or approach for identifying effectiveness variables in LSM ( Pourghasemi et al., 2018 ), and identifying these elements differs from one research to the next.

For this study, guided by recommendations from the National Institute of Forest Science, South Korea, we gathered data spanning ten distinct types selected as input attributes for the modeling process. These data sources included soil depth (SD), tree diameter (TD), forest condition (FC), bedrock (BR), curvature (CU), azimuth (AZ), topographic wetness index (TWI), catchment length (CL), catchment area (CA), and slope (SL). All variables have been compiled into a high-resolution raster grid format corresponding to a grid cell size of 10 m × 10 m. In addition, information on past landslide occurrences was also collected from the Korean Meteorological Administration database. All data were processed and analyzed using geographic information systems software. Figure 3 depicts a map of several conditions contributing to landslide occurrence in the region.

FIGURE 3 . Several factor maps cause landslides.

2.2 Data pre-processing

As previously highlighted, the ten factors collected in this research are rasterized with an impressive 10 m resolution. In addition to these datasets, we collected and processed data on landslide inventory for 2005–2019. It can be said that this is the most important dataset for studies on building LSM based on ML algorithms. This entire dataset was measured and analyzed by surveys at locations immediately after a landslide from aerial images, drones, and field trips. The pre-processing of this dataset for the ML models is marked by several integral phases. The landslide inventory dataset is first rasterized to match the grid cell size of the ten factors mentioned in this study, which is 10 m×10 m. This high resolution indicates that a single landslide event can be identified across multiple neighboring cells and the number of landslide-affected cells depending on the scale and size of the respective landslide event. Following rasterization, cell locations affected by landslides are carefully identified, inventoried, and labeled. Ten different data types are then stacked on the digitized landslide data. Once layered, data from all the superimposed layers is extracted for each grid cell. Based on their data attributes, these cells are systematically classified as landslides or non-landslides.

After processing and analyzing the data, it was found that approximately 7,188 grid cells recorded landslides, significantly lower than the number of grid cells that did not. The imbalance between these data classes is a typical problem in practice and can result in biased algorithms that perform poorly on the minority class ( Fernández et al., 2018 ). This occurs because ML models may learn to classify the majority class correctly while ignoring or incorrectly classifying the minority class ( Ma and He, 2013 ). To solve this problem and boost the effectiveness of the ML models, this study applied several solutions, such as the Tomek Links technique and the Near Miss algorithm, to resample the data. The following subsections go into further depth about these approaches. After processing, the final dataset had 14,376 samples, comprising 7,188 landslide samples and 7,188 non-landslide samples.

2.2.1 Tomek Links technique

The Tomek Links technique is a data pre-processing technique frequently utilized to improve the efficiency of ML algorithms. It is based on the idea that a pair of samples near one other and belonging to distinct classes are termed Tomek Links ( Tomek, 1976 ). This technique removes samples from the majority class for each pairing while expanding the gap between the two categories. This can be useful as noise and outliers often negatively affect the model’s accuracy and reliability. This approach could be applied to any ML issue, including classification, regression, and clustering. Figure 4 depicts the resampling procedure of the Tomek Links techniques.

FIGURE 4 . Illustration of Tomek Links technique.

2.2.2 NearMiss algorithm

The NearMiss algorithm is commonly employed to boost the effectiveness of ML algorithms, especially for imbalanced datasets. An imbalanced dataset is one in which the number of samples in one category differs considerably from the number of samples in the other. This frequently results in poor performance of ML algorithms, as the minority class is often underrepresented and may not be adequately learned by the model.

The NearMiss algorithm tackles this issue by selecting a subset of samples from the majority group that is similar to samples from the minority group based on a distance measure like Euclidean distance ( Zhang and Mani, 2003 ). The NearMiss algorithm has three variants: NearMiss1, NearMiss2, and NearMiss3. NearMiss1 chooses samples from the majority group nearest to the minority group. NearMiss-2 and NearMiss-3 select samples from the majority group that are furthest from the majority group and closest to the minority group, respectively. Figure 5 depicts the resampling procedure of the Near Miss algorithms.

FIGURE 5 . Illustration of NearMiss algorithm.

2.3 ML algorithms

This study used 6 ML algorithms to classify areas as prone to landslides or not prone to landslides based on a set of input features. Each algorithm has its assumptions, strengths, and limits, and we chose various algorithms to provide robust and reliable results. In addition, the cross-validation technique was utilized to analyze the performance of each algorithm and select the best-performing algorithm based on a range of evaluation criteria, for instance, PCC, recall, precision, and F 1 score. Before using ML algorithms to LSM, it is necessary first to gather a collection of labeled data, in which both the input characteristics and the landslide susceptibility (either “1″or “0″) are already established. ML models are then trained using the aforementioned labeled data via appropriate optimization algorithms.

LR is a simple and popular classification method using linear regression and the logistic function as its primary building blocks ( Cramer, 2002 ). In the context of LSM, LR is utilized to classify areas as either prone or not prone to landslides based on a collection of input features that are known to be associated with landslides. These elements may include geological and topographic characteristics, land use, land cover, and environmental factors like precipitation and slope angle. The LR objective is to find the line that best matches the data points and demarcates the two distinct groups of “landslide” and “non-landslide”. The output is a probability that ranges between 0 and 1, which is utilized to categorize the data into one of the two options. For instance, if the likelihood is higher than 0.5, the data is classified as potentially affected by landslides. Still, if the probability is lower than 0.5, the data is categorized as not having the potential to be affected by landslides. Figure 6A briefly illustrates the LR algorithm. The probability of the event occurring is expressed as:

FIGURE 6 . Illustration for (A) LR algorithm and (B) GNB algorithm.

Where β 0 , β 1 ,… are the coefficients and x 1 , x 2 ,… are the predictor variables

GNB classification algorithm is based on Bayesian probability theory and follows a Gaussian normal distribution, making it suitable for continuous data classification tasks. It is a straightforward and commonly adopted approach, mainly when the dataset is large. The method assumes that the input characteristics are independent of one other, called the “naive” assumption. This indicates that the probability of each factor is computed independently without considering the probabilities of the other characteristics. Through the use of the maximum likelihood estimate, it is possible to learn the possibility of the input characteristics given the various classes. The output is the likelihood that the data belong to each category, and the prediction is taken from the type with the highest probability. GNB can be sensitive to the assumptions of independence and normalcy of the features, even though it is a straightforward and efficient classification method that performs well on a wide range of datasets. Figure 6B briefly illustrates the GNB algorithm. Given a feature vector X = x 1 , x 2 , . . . , x n , the probability that a sample belongs to class C k is given by:

KNN is an ML algorithm used for classification tasks. The core of this technique is rooted in the principle of instance-based learning, wherein the categorization of a new data point is established based on its similarity to prior data points within the training dataset ( Cover and Hart, 1967 ). The technique uses a distance metric, for instance, Euclidean or Manhattan distance, to estimate the distance from the new data point to all points. The K closest points to the latest data are then selected, where K is a value determined by the user. The new data point is assigned to the category with the highest frequency among the K nearest points. KNN is a prevalent choice for classification jobs because of its simplicity and performance on various input formats. Figure 7A briefly illustrates the KNN algorithm. The distance between two points x and y in the Euclidean space is:

FIGURE 7 . Illustration for (A) KNN algorithm and (B) SVM algorithm.

SVM is a powerful and versatile classification algorithm that works by finding the hyperplane that maximally separates various classes in the feature space ( Cortes and Vapnik, 1995 ). The fundamental of SVM is finding a decision margin that maximally divides the data points into multiple classifications. A set of support vectors defines the decision boundary, the data points in the training dataset nearest to the hyperplane. The relative position of the support vectors determines the hyperplane position. The SVM output is a score that indicates the likelihood of the data belonging to one class or the other. SVM is a powerful and effective classification technique suited for a wide range of datasets, although it can be computationally expensive when the training set is large. Figure 7B briefly illustrates the SVM algorithm. The decision function is:

Where ω is the normal vector to the hyperplane and b is the offset.

RF is a classification algorithm incorporating numerous DT predictions to make more accurate and robust predictions ( Tin Kam, 1998 ; Breiman, 2001 ). A DT is a tree-shaped model in which each node represents an individual characteristic, each branch a particular choice, and each leaf node a distinct category or value. RF generates a collection of DT by bootstrapping a sample of the data and randomly picking a sample of the input data for every split. This means that each DT is constructed by drawing from a distinct sample of the data and a unique subset of the input variables. In classification tasks, RF determines the final classification by taking the unweighted average of the predictions made by individual DTs and selecting the most commonly predicted category. RF is a robust and efficient method resistant to overfitting and can be used on various datasets. Figure 8A briefly illustrates the RF algorithm. The classification decision is made by majority voting:

FIGURE 8 . Illustration for (A) RF algorithm and (B) XGB algorithm.

Where Y Ti is the prediction of the i th tree.

XGB is a well-known gradient-boosting technique that may be applied to classification and regression applications ( Chen and Guestrin, 2016 ). It is based on DT and is meant to form an ensemble of DTs to make predictions based on data that has not been seen before. An optimization approach that minimizes the loss function generates a set of decision trees in XGB. This means the DTs are formed to reduce the gap between the observed and estimated values. XGB is a robust and efficient technique for classification and regression work. It is a popular algorithm because of its strong performance on a diverse range of data formats and its capacity to deal with categorical and numerical characteristics. Figure 8B briefly illustrates the XGB algorithm. Given a differentiable loss function L(y, F(x)) , where y is the actual value and F(x) is the prediction, the prediction model is updated iteratively:

Where f i (x) is the prediction of the i th base leaner (typically a DT), and η is the learning rate.

2.4 Hyperparameter tuning and model validation

This study employed the grid search strategy for hyperparameter tuning, a widely used approach that can effectively enhance the accuracy and reliability of the resulting LSMs. As its applicability to small to medium-sized datasets, this method trains and evaluates an ML algorithm for each set of hyperparameters defined within a specified grid. For each model, the Python scikit-learn library facilitates the implementation of the grid search method.

The final processed dataset for this study contained 14,376 samples, with a balanced distribution of landslide and non-landslide instances (7,188 each). By maintaining a balanced dataset, potential model bias is mitigated, and the generalizability of prediction tasks is improved. The dataset is separated into training and testing subsets, representing 80% and 20% of the total dataset. Thus, the training set includes 11,500 samples (5,750 landslides and 5,750 non-landslides), while the testing set consists of 2,876 samples (1,438 landslides and 1,438 non-landslides). A 5-fold cross-validation is performed throughout the grid search procedure to fine-tune the models. This ensures that the findings are not overly optimistic and can be generalized to unseen data. The selected parameters for each ML model used in the grid search are detailed in Table 1 .

TABLE 1 . Parameter settings of each ML model used in the grid search.

2.5 Performance measures

Assessing algorithm performance is a crucial stage in using ML algorithms for LSM. It helps to identify the most accurate and reliable algorithm for a given study and can provide valuable insights into the strengths and limitations of the algorithms. This study used various evaluation criteria, including PCC, precision, recall, and F 1 score. PCC measures how well the model correctly classifies the data, whereas recall estimates the proportion of true positive (TP) predictions over total observed positive events. Precision measures how well the model avoids false positives (FP), and the F 1 score is the harmonic average of precision and recall. We estimated these evaluation metrics using a confusion matrix, which compares each case’s predicted and actual values. A confusion matrix comprises four main components: TP, FP, true negatives (TN), and false negatives (FN). In this case, TP and RN stand for the number of samples rightly categorized as positive and negative. In contrast, FP and FN refer to the number of samples wrongly classified as positive and negative. The following is a description of these criteria.

Where PCC means the percentage of correct classification (or accuracy), PV implies the probability of random agreement, and KC represents the Kappa coefficient.

3.1 Comparison of six ML algorithms

To compare the performance of 6 ML algorithms for LSM in Inje County, South Korea, we use several evaluation metrics, including PCC, F 1 score, and KC. The results of these evaluations are presented in Table 2 and Figures 9 , 10 below.

TABLE 2 . Performance statistics of six models.

FIGURE 9 . Performance of six algorithms in terms of classification.

FIGURE 10 . Classification performance of six models in terms of landslide class.

All algorithms generally have relatively high PCC, F 1 score, and KC values, indicating that they all perform well overall. However, the performance of the various models differs slightly. RF and XGB algorithms had the highest PCC values, with 0.939 and 0.941, respectively. These models can correctly classify landslide-prone areas with a high percentage, indicating that they can detect zones with an increased susceptibility to landslides. Regarding the F 1 score metric, these models can balance the precision and recall well, as the corresponding values for RF and XGB were 0.939 and 0.940, respectively. This is important as it will reduce the number of FNs and FPs. The KC metric also revealed a strong agreement between the predictions of these models and actual observations, and the models are not just labeling all areas as landslide-prone. The best values for this metric were 0.878 and 0.881, belonging to RF and XGB algorithms.

In contrast, the PCC and F 1 score values of the LR and GNB were lower than those of the other algorithms (0.869 and 0.87, respectively), and their KC values were also lower (0.738 and 0.739, respectively), which indicated that they might not perform as well as the other models. This might be because these models are simple, and they may be incapable of capturing the complex relationships between the input variables and the landslide sensitivity, as well as other models such as RF or XGB. It is worth noting that the KNN and SVM algorithms also performed pretty well, with PCC, F 1 score, and KC values of 0.921, 0.918, and 0.843 for the KNN, and 0.918, 0.915, and 0.836 for the SVM, respectively. These models were capable of recognizing regions with a high risk of landslides. However, they are not as good as RF and XGB regarding balancing the precision, recall, and agreement between the model predictions and the observations.

Figure 10 illustrates the classification performance for the landslide class of the ML models corresponding to the PCC, precision, and recall criteria. Similar to the previous evaluation, the LR and GNB models had lower accuracy and recall values, with the earlier having scores of 0.894 and 0.947, while the latter had numbers of 0.837 and 0.783, respectively. The KNN and SVM algorithms depicted a similar pattern, with precision and recall scores of 0.956 and 0.884 for KNN and corresponding numbers for SVM of 0.947 and 0.886, respectively. XGB had the highest precision value of 0.953 compared to other models, indicating its proficiency in identifying TPs and having a low FP rate. On the other hand, RF had the highest recall value of 0.929, demonstrating its ability to identify TP and low FN rates.

3.2 Robustness and sensitivity analysis

To offer a more thorough assessment of the robustness of RF and XGB models, an analysis of the confusion matrix and feature importance was performed and shown in Figure 11 , Figure 12 .

FIGURE 11 . Confusion matrix for RF and XGB.

FIGURE 12 . Variable importance of RF and XGB.

Figure 11 illustrates the confusion matrix details for both models. Out of 2,876 samples, the RF model yielded 1,365 TNs, 1,336 TPs, 73 FPs, and 102 FNs. In contrast, the XGB model resulted in 1,373 TNs, 1,332 TPs, 65 FPs, and 106 FNs. Regarding PCC, both the RF and XGB exhibited remarkable performance, registering classification accuracies of 93.9% and 94.1%, respectively. The precision and recall metrics underscore their efficacy, with RF showing 94.8% and 92.9%, respectively, whereas XGB showcased 95.3% and 92.6%.

Feature importance is a metric of how much each factor contributes to the predictions made by the model. According to the data presented in Figure 12 , the variable considered to be of the utmost significance in both the RF and the XGB models was SD (soil depth), which had a value of 0.246 and 0.285, respectively. This means that soil depth was the most influencing variable when determining landslide susceptibility in the research region. The second most significant variable in the RF algorithm was CU, with a value of 0.133, followed by TWI, with 0.144. In the case of the XGB model, the CA variable was the second most important variable, followed by the TWI variable, which had values of 0.144 and 0.125. BR was the component with the lowest importance score in both models, with a value of 0.032 in RF and 0.038 in XGB models.

3.3 LSM of Inje area

The LSM generated by the RF and XGB algorithms were displayed in Figure 13 , Figure 14 , respectively. These maps were based on the likelihood of landslides occurring for each grid cell, which ranges from 0 to 1. As seen in these Figures, the LSM is segmented into five levels following the categorization used by the National Institute of Forest Science in South Korea. Among them, level 4, with a range of probability [0.5–0.7], and level 5, with a range of probability [0.7–1.0], are considered the most dangerous warning levels. The visualized results of the two algorithms revealed a slight difference in the landslide probability estimation for grid cells. This distinction was particularly noticeable when comparing level 1 (0–0.1) and level 3. In contrast, there was not much difference between the other levels, notably levels 4 and 5, each with a significant possibility of landslides.

FIGURE 13 . LSM using RF algorithm.

FIGURE 14 . LSM using XGB algorithm.

3.4 Current approach for LSM in South Korea

The current approach for LSM in South Korea focuses primarily on using statistical techniques, particularly logistic regression analysis, to identify the association between input factors ( Woo et al., 2014 ). The following equations are used to calculate the probability of a landslide hazard at a specific site. The LSM for the Inje area was developed from these equations, as shown in Figure 15 .

FIGURE 15 . LSM using LR analysis.

According to Lee et al. (2015) , this methodology achieved an average accuracy of approximately 75.2%. When comparing this method to the ML approach presented in Section 3.1, the accuracy of ML models was about 92% on average, which was a significant improvement over the process that is currently being used. The PCC and F 1 scores of the 6 ML models range from 0.869 to 0.941 and 0.857 to 0.940, respectively. Among the six algorithms, RF and XGB performed best, with accuracy and F 1 values of [0.939–0.941]. The results of this study indicated that using ML algorithms for LSM can lead to more accurate and higher-resolution LSMs. It is essential to point out that one of the benefits of adopting an ML technique is that it is capable of handling complicated interactions between input variables and landslide susceptibility. These types of relationships may be difficult to capture using statistical methods. Additionally, the use of ML algorithms provides a valuable alternative to the current statistical approach in South Korea.

4 Discussion

The ability of ML algorithms to enhance LSM in contexts such as Inje County, South Korea, has been substantiated in this study. The RF and XGB models performed particularly well in correctly identifying landslide-prone areas, while KNN and SVM performed well. Nonetheless, algorithm selection for such applications should be based on performance metrics and factors like ease of implementation, model interpretability, and inherent limitations ( Zhou et al., 2021 ; Liu et al., 2023 ).

The poorer performance exhibited by LR and GNB resonates with prior literature on their inherent limitations. For instance, LR assumes a linear relation between predictor and response variables, which often is not true in complex natural phenomena like landslides ( Pham et al., 2016 ). Additionally, its sensitivity to outliers and imbalanced datasets hampers its ability to capture intricate non-linear associations ( Akinci and Zeybek, 2021 ). Similarly, while GNB’s simplicity and computational efficiency are well acknowledged, its assumption of input independence often leads to its suboptimal performance, consistent with our findings and those of Azarafza et al. (2021) .

The comparable performance of KNN and SVM echoes prior works. KNN, though simple in its construct and devoid of data distribution assumptions, suffers when processing power is limited, given its requirement to compute distances between every training and new data point. The sensitivity of KNN to the choice of ‘K' has been reported by studies such as Sameen et al. (2020) . SVM’s strength lies in its ability to grapple with high-dimensional data and intricate non-linear predictor-response relationships ( Yao et al., 2008 ). However, SVM’s sensitivity to kernel function choices and the regularization value, as corroborated in our study, is a limitation noted by Huang and Zhao (2018) .

According to the findings of this research, the 2 ML algorithms most useful for LSM are RF and XGB. Both algorithms have the advantage of being capable of handling high-dimensional data and non-linear correlations between input characteristics and output. RF’s resilience to data noise and outliers has been highlighted by Dou et al. (2019) , whereas XGB’s speed and robustness to missing data are mentioned in Biswas et al. (2022) . However, the potential computational intensiveness of RF and the sometimes opaque results of XGB, as observed in our study, are consistent with the findings of Merghadi et al. (2020) . Despite these limitations, both these models hold promise for LSM, and the choice between them should be reasonable, considering data characteristics and the study’s objectives.

Adopting the ML-based approach for LSM holds substantial practical implications for regions like South Korea. Enhanced predictive accuracy not only aids in more effective resource allocation for disaster prevention but also offers a robust foundation for policymaking at both local and national scales. Proactively identifying high-risk zones, authorities can prioritize infrastructure development, design better evacuation strategies, and optimize disaster response protocols. This research thus paves the way for an integrated, data-driven approach to landslide management, potentially safeguarding communities and infrastructures against the devastating impacts of unforeseen landslides.

Despite the promising outcomes of this research, it is crucial to recognize its limitations. Data availability restricted our study to a specific timeframe and region, potentially affecting broader generalizations. Moreover, while the chosen ML algorithms offer advanced predictions, their complexity could pose challenges in real-time applications or when adapting to other regions with distinct geological characteristics. Future research could explore integrating diverse datasets for better model training and investigating simpler yet efficient algorithms. Additionally, cross-regional studies would be invaluable in assessing the universal applicability of the ML-based LSM approach.

5 Conclusion

The application of 6 ML algorithms for LSM in Inje County, South Korea, has revealed distinct efficiencies among the algorithms. RF and XG perform particularly well in accurately classifying landslide-prone areas, balancing precision and recall, and having PCC values up to about 94%. The KNN and SVM models with a PCC score of roughly 92% also performed well in correctly classifying landslide-prone areas. The two models, LR and GNB, have the lowest efficiency, corresponding to a PCC value of only about 87%. Despite this, the results demonstrate that the ML approach is superior to the existing statistical approach with logistic regression, which has an accuracy of about 75.2%.

This study provides a valuable starting point for further research into using ML techniques for LSM in South Korea and other regions. The RF and XGB models are highly recommended for LSM usage in South Korea. However, it is essential to consider the trade-offs between the performance and simplicity of implementation and interoperability. The efficiency of algorithms differs from location to region for various reasons, including geomorphology, land uses, and other factors.

The broader implications of this finding can potentially reshape disaster management policies in South Korea. A shift from traditional models to ML-based LSMs can facilitate more accurate predictions, allowing authorities to pre-emptively address and mitigate landslide risks in vulnerable zones. This study will substantially impact the understanding of landslide risk in Inje County. It will be valuable in developing effective mitigation techniques for landslides in the place.

Data availability statement

The raw data supporting the conclusion of this article will be made available by the authors, without undue reservation.

Author contributions

X-HL: Formal Analysis, Visualization, Writing–original draft. SE: Formal Analysis, Investigation, Writing–review and editing. CC: Data curation, Investigation, Visualization, Writing–original draft. DN: Methodology, Writing–review and editing. MY: Data curation, Investigation, Writing–review and editing. GL: Supervision, Writing–review and editing.

The author(s) declare financial support was received for the research, authorship, and/or publication of this article. This research was supported by the Disaster-Safety Platform Technology Development Program of the National Research Foundation of Korea (NRF), funded by the Ministry of Science and ICT. (No. 2022M3D7A1090338).

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Ado, M., Amitab, K., Maji, A. K., Jasińska, E., Gono, R., Leonowicz, Z., et al. (2022). Landslide susceptibility mapping using machine learning: A literature survey. Remote Sens. 14 (13), 3029. doi:10.3390/rs14133029

CrossRef Full Text | Google Scholar

Akinci, H., and Zeybek, M. (2021). Comparing classical statistic and machine learning models in landslide susceptibility mapping in Ardanuc (Artvin), Turkey. Nat. Hazards 108 (2), 1515–1543. doi:10.1007/s11069-021-04743-4

Atkinson, P. M., and Massari, R. (2011). Autologistic modelling of susceptibility to landsliding in the Central Apennines, Italy. Geomorphology 130 (1), 55–64. doi:10.1016/j.geomorph.2011.02.001

Ayalew, L., and Yamagishi, H. (2005). The application of GIS-based logistic regression for landslide susceptibility mapping in the Kakuda-Yahiko Mountains, Central Japan. Geomorphology 65 (1), 15–31. doi:10.1016/j.geomorph.2004.06.010

Azarafza, M., Azarafza, M., Akgün, H., Atkinson, P. M., and Derakhshani, R. (2021). Deep learning-based landslide susceptibility mapping. Sci. Rep. 11 (1), 24112. doi:10.1038/s41598-021-03585-1

PubMed Abstract | CrossRef Full Text | Google Scholar

Biswas, N. K., Stanley, T. A., Kirschbaum, D. B., Amatya, P. M., Meechaiya, C., Poortinga, A., et al. (2022). A dynamic landslide hazard monitoring framework for the Lower Mekong Region. Front. Earth Sci. 10, 1057796. doi:10.3389/feart.2022.1057796

Breiman, L. (2001). Random forests. Mach. Learn. 45 (1), 5–32. doi:10.1023/A:1010933404324

Caleca, F., Tofani, V., Segoni, S., Raspini, F., Franceschini, R., and Rosi, A. (2022). How can landslide risk maps be validated? Potential solutions with open-source databases. Front. Earth Sci. 10, 998885. doi:10.3389/feart.2022.998885

Cardinali, M., Reichenbach, P., Guzzetti, F., Ardizzone, F., Antonini, G., Galli, M., et al. (2002). A geomorphological approach to the estimation of landslide hazards and risks in Umbria, Central Italy. Nat. Hazards Earth Syst. Sci. 2 (1/2), 57–72. doi:10.5194/nhess-2-57-2002

Chae, B. G., Park, H. J., Catani, F., Simoni, A., and Berti, M. (2017). Landslide prediction, monitoring and early warning: A concise review of state-of-the-art. Geosciences J. 21 (6), 1033–1070. doi:10.1007/s12303-017-0034-4

Chen, T., and Guestrin, C. (2016). “XGBoost: A scalable tree boosting system,” in Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining (San Francisco, California, USA: Association for Computing Machinery ).

Google Scholar

Chikalamo, E. E., Mavrouli, O. C., Ettema, J., van Westen, C. J., Muntohar, A. S., and Mustofa, A. (2020). Satellite-derived rainfall thresholds for landslide early warning in Bogowonto Catchment, Central Java, Indonesia. Int. J. Appl. Earth Observation Geoinformation 89, 102093. doi:10.1016/j.jag.2020.102093

Cortes, C., and Vapnik, V. (1995). Support-vector networks. Mach. Learn. 20 (3), 273–297. doi:10.1007/BF00994018

Cover, T., and Hart, P. (1967). Nearest neighbor pattern classification. IEEE Trans. Inf. Theory 13 (1), 21–27. doi:10.1109/TIT.1967.1053964

Cramer, J. S. (2002). The origins of logistic regression . Rotterdam, Netherlands: Tinbergen Institute Working Paper No. 2002-119/4 . doi:10.2139/ssrn.360300

Dou, J., Yunus, A. P., Tien Bui, D., Merghadi, A., Sahana, M., Zhu, Z., et al. (2019). Assessment of advanced random forest and decision tree algorithms for modeling rainfall-induced landslide susceptibility in the Izu-Oshima Volcanic Island, Japan. Sci. Total Environ. 662, 332–346. doi:10.1016/j.scitotenv.2019.01.221

Eker, A. M., Dikmen, M., Cambazoğlu, S., Düzgün, Ş. H. B., and Akgün, H. (2015). Evaluation and comparison of landslide susceptibility mapping methods: A case study for the ulus district, bartın, northern Turkey. Int. J. Geogr. Inf. Sci. 29 (1), 132–158. doi:10.1080/13658816.2014.953164

Fernández, A., García, S., Galar, M., Prati, R. C., Krawczyk, B., and Herrera, F. (2018). Learning from imbalanced data sets . Springer Cham . doi:10.1007/978-3-319-98074-4

Highland, L. M., and Bobrowsky, P. (2008). The landslide handbook—a guide to understanding landslides . Reston, Virginia: U.S. Geological Survey Circular .

Hong, H., Shahabi, H., Shirzadi, A., Chen, W., Chapi, K., Ahmad, B. B., et al. (2019). Landslide susceptibility assessment at the wuning area, China: A comparison between multi-criteria decision making, bivariate statistical and machine learning methods. Nat. Hazards 96 (1), 173–212. doi:10.1007/s11069-018-3536-0

Huabin, W., Gangjun, L., Weiya, X., and Gonghui, W. (2005). GIS-Based landslide hazard assessment: an overview. Prog. Phys. Geogr. Earth Environ. 29 (4), 548–567. doi:10.1191/0309133305pp462ra

Huang, F., Cao, Z., Guo, J., Jiang, S. H., Li, S., and Guo, Z. (2020). Comparisons of heuristic, general statistical and machine learning models for landslide susceptibility prediction and mapping. CATENA 191, 104580. doi:10.1016/j.catena.2020.104580

Huang, Y., and Zhao, L. (2018). Review on landslide susceptibility mapping using support vector machines. CATENA 165, 520–529. doi:10.1016/j.catena.2018.03.003

Kadavi, P. R., Lee, C. W., and Lee, S. (2019). Landslide-susceptibility mapping in Gangwon-do, South Korea, using logistic regression and decision tree models. Environ. Earth Sci. 78 (4), 116. doi:10.1007/s12665-019-8119-1

Kim, K., Koo, K., Yoon, H., Lee, C., Won, M., Woo, C., et al. (2011). Forest disasters 2011 . Seoul: Korea Forest Research Institute .

Korea Forest Service (2021). Landslide information system. Available At: https://sansatai.forest.go.kr/gis/main.do (Accessed February 22, 2023).

Lee, B., An, H., Lee, C., Seo, K., Woo, C., Kwon, C., et al. (2021). Forest disaster white paper" . Seoul: National Institute of Forest Science , 2020.

Lee, C., Woo, C., Kim, D., Seo, J., Kwon, H., and Yun, H. (2015a). Development of landslide disaster prevention system in urban living area . South Korea: National Institute of Forest Science .

Lee, M. J., Park, I., Won, J. S., and Lee, S. (2016). Landslide hazard mapping considering rainfall probability in Inje, Korea. Geomatics, Nat. Hazards Risk 7 (1), 424–446. doi:10.1080/19475705.2014.931307

Lee, S. (2005). Application of logistic regression model and its validation for landslide susceptibility mapping using GIS and remote sensing data. Int. J. Remote Sens. 26 (7), 1477–1491. doi:10.1080/01431160412331331012

Lee, S., and Min, K. (2001). Statistical analysis of landslide susceptibility at Yongin, Korea. Environ. Geol. 40 (9), 1095–1113. doi:10.1007/s002540100310

Lee, S., Won, J. S., Jeon, S. W., Park, I., and Lee, M. J. (2015b). Spatial landslide hazard prediction using rainfall probability and a logistic regression model. Math. Geosci. 47 (5), 565–589. doi:10.1007/s11004-014-9560-z

Liu, S., Wang, L., Zhang, W., He, Y., and Pijush, S. (2023). A comprehensive review of machine learning-based methods in landslide susceptibility mapping. Geol. J. 58 (6), 2283–2301. doi:10.1002/gj.4666

Lombardo, L., Opitz, T., Ardizzone, F., Guzzetti, F., and Huser, R. (2020). Space-time landslide predictive modelling. Earth-Science Rev. 209, 103318. doi:10.1016/j.earscirev.2020.103318

Lv, L., Chen, T., Dou, J., and Plaza, A. (2022). A hybrid ensemble-based deep-learning framework for landslide susceptibility mapping. Int. J. Appl. Earth Observation Geoinformation 108, 102713. doi:10.1016/j.jag.2022.102713

Ma, Y., and He, H. (2013). Imbalanced learning: foundations, algorithms, and applications . Wiley-IEEE Press . doi:10.1002/9781118646106

Maes, J., Kervyn, M., de Hontheim, A., Dewitte, O., Jacobs, L., Mertens, K., et al. (2017). Landslide risk reduction measures: A review of practices and challenges for the tropics. Prog. Phys. Geogr. Earth Environ. 41 (2), 191–221. doi:10.1177/0309133316689344

Merghadi, A., Yunus, A. P., Dou, J., Whiteley, J., ThaiPham, B., Bui, D. T., et al. (2020). Machine learning methods for landslide susceptibility studies: A comparative overview of algorithm performance. Earth-Science Rev. 207, 103225. doi:10.1016/j.earscirev.2020.103225

Mersha, T., and Meten, M. (2020). GIS-based landslide susceptibility mapping and assessment using bivariate statistical methods in Simada area, northwestern Ethiopia. Geoenvironmental Disasters 7 (1), 20. doi:10.1186/s40677-020-00155-x

Pardeshi, S. D., Autade, S. E., and Pardeshi, S. S. (2013). Landslide hazard assessment: recent trends and techniques. SpringerPlus 2 (1), 523. doi:10.1186/2193-1801-2-523

Pham, B. T., Pradhan, B., Tien Bui, D., Prakash, I., and Dholakia, M. B. (2016). A comparative study of different machine learning methods for landslide susceptibility assessment: A case study of uttarakhand area (India). Environ. Model. Softw. 84, 240–250. doi:10.1016/j.envsoft.2016.07.005

Piciullo, L., Calvello, M., and Cepeda, J. M. (2018). Territorial early warning systems for rainfall-induced landslides. Earth-Science Rev. 179, 228–247. doi:10.1016/j.earscirev.2018.02.013

Pourghasemi, H. R., Teimoori Yansari, Z., Panagos, P., and Pradhan, B. (2018). Analysis and evaluation of landslide susceptibility: A review on articles published during 2005–2016 (periods of 2005–2012 and 2013–2016). Arabian J. Geosciences 11 (9), 193. doi:10.1007/s12517-018-3531-5

Rahman, G., Bacha, A. S., Ul Moazzam, M. F., Rahman, A. U., Mahmood, S., Almohamad, H., et al. (2022). Assessment of landslide susceptibility, exposure, vulnerability, and risk in shahpur valley, eastern hindu kush. Front. Earth Sci. 10, 953627. doi:10.3389/feart.2022.953627

Reichenbach, P., Rossi, M., Malamud, B. D., Mihir, M., and Guzzetti, F. (2018). A review of statistically-based landslide susceptibility models. Earth-Science Rev. 180, 60–91. doi:10.1016/j.earscirev.2018.03.001

Sahin, E. K., Colkesen, I., and Kavzoglu, T. (2020). A comparative assessment of canonical correlation forest, random forest, rotation forest and logistic regression methods for landslide susceptibility mapping. Geocarto Int. 35 (4), 341–363. doi:10.1080/10106049.2018.1516248

Sameen, M. I., Pradhan, B., Bui, D. T., and Alamri, A. M. (2020). Systematic sample subdividing strategy for training landslide susceptibility models. CATENA 187, 104358. doi:10.1016/j.catena.2019.104358

Shano, L., Raghuvanshi, T. K., and Meten, M. (2020). Landslide susceptibility evaluation and hazard zonation techniques – A review. Geoenvironmental Disasters 7 (1), 18. doi:10.1186/s40677-020-00152-0

Stanley, T. A., Kirschbaum, D. B., Benz, G., Emberson, R. A., Amatya, P. M., Medwedeff, W., et al. (2021). Data-driven landslide nowcasting at the global scale. Front. Earth Sci. 9, 640043. doi:10.3389/feart.2021.640043

Thirugnanam, H., Ramesh, M. V., and Rangan, V. P. (2020). Enhancing the reliability of landslide early warning systems by machine learning. Landslides 17 (9), 2231–2246. doi:10.1007/s10346-020-01453-z

Tin Kam, H. (1998). The random subspace method for constructing decision forests. IEEE Trans. Pattern Analysis Mach. Intell. 20 (8), 832–844. doi:10.1109/34.709601

Tomek, I. (1976). Two modifications of CNN. IEEE Trans. Syst. 6 (11), 769–772. doi:10.1109/TSMC.1976.4309452

Tran, T. V., Alvioli, M., and Hoang, V. H. (2022). Description of a complex, rainfall-induced landslide within a multi-stage three-dimensional model. Nat. Hazards 110 (3), 1953–1968. doi:10.1007/s11069-021-05020-0

Wang, H., Zhang, L., Yin, K., Luo, H., and Li, J. (2021). Landslide identification using machine learning. Geosci. Front. 12 (1), 351–364. doi:10.1016/j.gsf.2020.02.012

Wang, X., Du, A., Hu, F., Liu, Z., Zhang, X., Wang, L., et al. (2023). Landslide susceptibility evaluation based on active deformation and graph convolutional network algorithm. Front. Earth Sci. 11, 1132722. doi:10.3389/feart.2023.1132722

Woo, C., Kwon, H., Lee, C., and Kim, K. (2014). Landslide hazard prediction map based on logistic regression model for applying in the whole country of South Korea. J. Korean Soc. Hazard Mitig. 14 (6), 117–123. doi:10.9798/KOSHAM.2014.14.6.117

Yalcin, A. (2008). GIS-Based landslide susceptibility mapping using analytical hierarchy process and bivariate statistics in ardesen (Turkey): comparisons of results and confirmations. CATENA 72 (1), 1–12. doi:10.1016/j.catena.2007.01.003

Yalcin, A., Reis, S., Aydinoglu, A. C., and Yomralioglu, T. (2011). A GIS-based comparative study of frequency ratio, analytical hierarchy process, bivariate statistics and logistics regression methods for landslide susceptibility mapping in Trabzon, NE Turkey. CATENA 85 (3), 274–287. doi:10.1016/j.catena.2011.01.014

Yao, X., Tham, L. G., and Dai, F. C. (2008). Landslide susceptibility mapping based on support vector machine: A case study on natural slopes of Hong Kong, China. Geomorphology 101 (4), 572–582. doi:10.1016/j.geomorph.2008.02.011

Ye, C., Tang, R., Wei, R., Guo, Z., and Zhang, H. (2023). Generating accurate negative samples for landslide susceptibility mapping: A combined self-organizing-map and one-class SVM method. Front. Earth Sci. 10, 1054027. doi:10.3389/feart.2022.1054027

Yoo, N. J., Lee, C. J., Choi, Y. J., and Lee, H. S. (2009). “A study on disaster and restoration of debris flow landslides at Inje, Kangwon Province, Korea,” in 17th international conference on soil mechanics and geotechnical engineering (ICSMGE) (Alexandria, Egypt: IOS Press ).

Youssef, A. M., and Pourghasemi, H. R. (2021). Landslide susceptibility mapping using machine learning algorithms and comparison of their performance at Abha Basin, Asir Region, Saudi Arabia. Geosci. Front. 12 (2), 639–655. doi:10.1016/j.gsf.2020.05.010

Zêzere, J. L., Pereira, S., Melo, R., Oliveira, S. C., and Garcia, R. A. C. (2017). Mapping landslide susceptibility using data-driven methods. Sci. Total Environ. 589, 250–267. doi:10.1016/j.scitotenv.2017.02.188

Zhang, J., and Mani, I. (2003). “KNN approach to unbalanced data distributions: A case study involving information extraction,” in ICML'2003 workshop on learning from imbalanced data sets II (Washington, DC: ICML ).

Zhou, X., Wu, W., Qin, Y., and Fu, X. (2021). Geoinformation-based landslide susceptibility mapping in subtropical area. Sci. Rep. 11 (1), 24325. doi:10.1038/s41598-021-03743-5

Keywords: disaster management, extreme gradient boosting (XGB), feature importance, landslide, landslide probability, landslide susceptibility mapping (LSM), random forest (RF), risk map

Citation: Le X-H, Eu S, Choi C, Nguyen DH, Yeon M and Lee G (2023) Machine learning for high-resolution landslide susceptibility mapping: case study in Inje County, South Korea. Front. Earth Sci. 11:1268501. doi: 10.3389/feart.2023.1268501

Received: 28 July 2023; Accepted: 08 September 2023; Published: 20 September 2023.

Reviewed by:

Copyright © 2023 Le, Eu, Choi, Nguyen, Yeon and Lee. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

*Correspondence: Giha Lee, [email protected]

This article is part of the Research Topic

Geophysical, Climatological and Anthropogenic Hazards and Disaster: Vulnerability, Risk Assessment, and Sustainability

• Open access
• Published: 15 June 2020

GIS-based landslide susceptibility mapping and assessment using bivariate statistical methods in Simada area, northwestern Ethiopia

• Tilahun Mersha 1 , 2 &
• Matebie Meten 1  

Geoenvironmental Disasters volume  7 , Article number:  20 ( 2020 ) Cite this article

32k Accesses

90 Citations

1 Altmetric

Metrics details

Simada area is found in the South Gondar Zone of Amhara National Regional State and it is 780Km far from Addis Ababa. Physiographically, it is part of the northwestern highlands of Ethiopia. This area is part of the Guna Mountain which is characterized by weathered volcanic rocks, rugged morphology with deeply incised gorges, heavy rainfall and active surface processes. Many landslides have occurred on August 2018 after a period of heavy rainfall and they caused many damages to the local people. In this study, Frequency Ratio (FR) and Weights of Evidence (WoE) models were applied to evaluate the landslide causative factors and generate landslide susceptibility maps (LSMs). The landslide inventory map that consists of 576 active and passive landslide scarps was prepared from intensive fieldwork and Google Earth image interpretation. These landslide locations were randomly divided into 80% training and 20% validation datasets. Seven landslide causal factors including aspect, slope, curvature, lithology, land use, rainfall and distance to stream were combined with a training dataset using GIS tools to generate the LSMs of the study area. Then the area was divided into five landslide susceptibility zones of very low, low, moderate, high and very high. Later, the resulting maps have been validated by using area under the curve and landslide density index methods. The result showed that the predictive rate of FR and WoE models were 88.2% and 84.8%, respectively. This indicated that the LSM produced by FR model showed a better performance than that of WoE model. Finally, the LSMs produced by FR and WoE models can be used by decision-makers for land use planning and landslide mitigation purpose.

Introduction

Landslides are one of the recurrent natural problems that are widespread throughout the world, especially in mountainous areas which caused a significant injury and loss of human life, damage in properties and infrastructures (Parise and Jibson 2000 ; Dai et al. 2002 ; Glade et al. 2005 ; Kanungo et al. 2006 ; Pan et al. 2008 ; Girma et al. 2015 ). The term “landslide” is the movement of a mass of rock, debris or earth down a slope under the influence of gravity (Varnes 1978 ; Hutchinson 1989 ; WP/WLI - International Geotechnical Societies’ UNESCO Working Party on World Landslide Inventory 1990 ; Cruden 1991 ; Cruden and Varnes 1996 ). Landslides are caused by different triggering factors such as heavy or prolonged precipitation, earthquakes, rapid snow melting and a variety of anthropogenic activities. Landslides can involve flowing, sliding, toppling or falling movements and many landslides exhibit a combination of two or more types of movements (Crozier 1986 ; Cruden and Varnes 1996 ; Dikau et al. 1996 ).

Landslide in Ethiopia is a common phenomenon which often causes significant damage to people and property. Almost 60% of the total population in Ethiopia lives in the highland areas (Ayalew 1999 ) which is characterized by high relief, complex geology, high rainfall, rugged morphology, very deep valleys and gorges with active river incision. The rapid population growth demanded the use of areas which were not previously used for settlement, urban expansion, agricultural and other purposes thereby exposing these areas to landslide problems after rainy seasons (Temesgen et al. 2001 ; Abebe et al. 2010 ; Woldearegay 2013 ).

In recent years landslide incidences are increasing in the Ethiopian highlands due to man-made and natural causes (Meten et al. 2015b ). For instance, from 1960 to 2010 alone, Landslides have killed 388 people, injured 24 people, and damaged agricultural lands, houses and infrastructures (Ayalew 1999 ; Temesgen et al. 1999 ; Woldearegay 2008 and (Ibrahim: Landslide assessment and hazard zonation in Mersa and Wurgessa, North Wollo, Ethiopia, unpublished)). According to Abebe et al. ( 2010 ), the highlands and mountainous area of Ethiopia like the Blue Nile Gorge, the Lower Wabe-Shebele River valley, Gilgel Gibe River, Tarmaber, Kombolcha - Dessie road, Uba Dema village in Sawla, Wondogenet area and many other parts of Ethiopia are repeatedly facing problems associated with landslides. The landslides in these areas are affecting human lives, infrastructures, agricultural lands and the natural environment. As a result of this, the study of the landslide has drawn global attention to increase awareness about its socioeconomic impacts and the pressure of increasing population and urbanization on mountainous areas (Kanungo et al. 2006 ).

The current study area is found in Simada District of South Gondar Zone in the Amhara National Regional State of Northwestern Ethiopia. It is part of the northwestern Ethiopian highlands. This area is severely affected by landslide incidences in recent years. Landslide incidence in the study area occurred on August, 2018 after a heavy and prolonged rainfall that caused the death of animals, destruction of houses and wide areas of cultivated and non-cultivated lands. Therefore, this area requires a detailed investigation to evaluate the causes, types and failure mechanisms of landslides and to prepare the landslide susceptibility maps. A systematic landslide study helps to reduce the damages in infrastructures, houses and cultivated lands and loss of lives. This importance will be noticed when these landslide susceptibility maps are used by decision-makers in regional land use planning, landslide prevention and mitigation measures.

For proper and strategic land use planning, it is important to evaluate and delineate landslide prone areas using different landslide susceptibility mapping techniques. Preparing a landslide susceptibility map of a certain area is a useful tool in landslide hazard management as it shows the degree of susceptibility of an area to landslide occurrence. It is obvious that landslide susceptibility maps can be generated based on the assumption that future landslide will occur under the same condition as in the past (Pham et al. 2015 ). Interpretation of future landslide occurrence needs an understanding of conditions and processes that control landslides in the study area. Past landslides and different conditioning factors such as slope morphology, hydrogeology and geology of the area are the main parameters to assess and evaluate landslide susceptibility by integrating these conditioning factors and past landslides in a GIS environment.

GIS-based landslide susceptibility mapping techniques have been used by several researchers (Aleotti and Chowdhury 1999 ; Kanungo et al. 2009 ) which can be classified into qualitative and quantitative ones (Yalcin et al. 2011 ; Felicisimo et al. 2012 ; Peng et al. 2014 ; Wang and Li 2017 ). Qualitative techniques include geomorphological analyses and inventory methods. These are based on expert judgment and are more subjective than quantitative methods. Quantitative methods such as deterministic analyses, probabilistic approaches and statistical techniques closely rely on mathematical models which have much less personal bias but still needs experience to produce and run these models (Aleotti and Chowdhury 1999 ; Kanungo et al. 2009 ). In recent years, many landslide susceptibility maps were produced using GIS-based statistical approaches like Frequency Ratio (FR) and Weights of Evidence (WoE) models. This is because the result from these models showed good performance with high accuracy and these models are very simple to implement and can provide the contribution of each causative factor class for landslide occurrence (Lee and Pradhan 2007 ; Akgun et al. 2007 ; Dahal et al. 2008 ; Işık Yilmaz 2009 ; Pradhan, Lee and Buchroithner 2010 ; Choi et al. 2012 ; Park et al. 2012 ; Vakhshoori and Zare 2016 ; Fayez et al. 2018 ).

Several researchers have used Frequency ratio model on landslide studies (Bahrain et al. 2014 ; Meten et al. 2015a ; Haoyuan Hong et al. 2015 ; Pham et al. 2015 ; Pirasteh and Li 2017 ; Fayez et al. 2018 ; Khan et al. 2019 ) and in comparison with a few methods (Akgun et al. 2007 ; Lee and Pradhan 2007 ; Işık Yilmaz 2009 ; Choi et al. 2012 ; Park et al. 2012 ; Meten et al. 2015b ; Wang and Li 2017 ). A combination of both FR and WoE models have been applied for landslide susceptibility mapping (Regmi et al. 2013 ; Rahmati et al. 2016 ). Gholami et al. ( 2019 ) also compared the prediction capability of frequency ratio, fuzzy gamma and landslide index models. Each GIS-based statistical method requires data on past landslides, preparatory causative factors and triggering factors. To prevent or mitigate any damage from landslides, it is essential to assess the landslide prone areas. The current study aims to carryout landslide susceptibility mapping by applying FR and WoE models in order to highlight critically high and very high hazard zones. This will help to reduce and mitigate any hazard associated with future landslide occurrence.

The study area is 185.7 square kilometers which is located in Simada District of South Gondar Zone, Amhara National Regional State, Ethiopia (Fig. 1 ). The area is bounded between 38 0 11' E and 38 0 20' E longitudes and 11 0 30' N and 11 0 41' N latitudes. The typical drainage pattern of the study area is dendritic and parallel. Atkus and Kostet Rivers are the main rivers that affect the study area by eroding the banks of rivers leading to slope instability. The confluence of these rivers forms Bijena River which is the largest river in the study area. Most of the rivers in the study area flow towards the southeast direction. The physiography of the study area forms the rugged topography of Guna Mountain (Fig. 2 ) which is part of the northwestern Ethiopian highlands. The area can be classified into two main physiographic regions. These are the plateau area and the rugged terrain. The plateau areas are characterized by volcanic landscapes that represent the high flatlands of the Kefoye, Agona and Jinjero Gedel areas. These areas are water divide zones in which rivers are flowing to Abay Basin in the west and to Bashilo Basin in the south. In this area, the slopes are ranging from flat slopes on the top to steeper slopes at the plateau scarp. The rugged terrain is highly dissected by major rivers and streams which are characterized by deep narrow valleys and gorges. Slopes in these areas are steep to vertical and susceptible to erosional and landslide phenomena. The elevation of the study area ranges from 2067m to 3586 m which comprises of medium to very high relief hills. The presence of steep scarps, rugged slope faces, deep gorges and steep ridges showed that this area is prone to active surface processes and landslide incidences. Based on elevation, the climatic zones of the study area are mostly falling under the highland climatic zone. The primary wet season extends from June to September. There is great variation in the rainfall amounts with maximum rainfall occurring during the wet season which starts in June and ends in September with the heaviest rainfall occurring during the months of July and August.

Location map of the study area

Three dimensional map of the study area

In order to achieve the objectives of this research, data collection and organization, preparation of landslide inventory datasets, database construction of landslide causative factors and application of FR and WoE models were carried out to prepare the landslide susceptibility maps and validate them.

Data collection and organization

The necessary data for this study were collected from various sources. These include collecting relevant literatures from published and unpublished papers, DEM data from USGS, a regional geological map from Geological Survey of Ethiopia at a scale of 1:250000, rainfall data from National Metrological Agency of Ethiopia, a topographic map from Ethiopian Geospatial Information Agency at a scale of 1:50000 and Google Earth image from Google Earth. During field work, data collection was carried out on different rock types by describing their character, the relative degree of weathering, slope steepness, location of springs and swamps, landslide inventory mapping on both active landslide and scarp areas by measuring their length, width, accumulation zone and depth (if possible), land use and land cover, man-made activities including farming practice. After compilation of the actual field investigation, the data has been systematically processed and analyzed first in ArcGIS followed by Microsoft Excel and then finally in ArcGIS.

Preparation of landslide inventory dataset

The quality of the landslide inventories depends on the accuracy, type and certainty of the information shown in the maps. New and emerging mapping methods, based chiefly on satellite, aerial and terrestrial remote sensing technologies, can greatly facilitate the production and the update of landslide maps. Literature review has shown that the most promising approaches exploit VHR optical, monoscopic and stereoscopic satellite images, analyzed visually or through semi-automatic procedures, and VHR digital representations of surface topography captured by LiDAR sensors. A combination of satellite, aerial and terrestrial remote sensing data represents the optimal solution for landslide detection and mapping, in different physiographic, climatic and land cover conditions (Guzzetti et al. 2012 ). Ye et al. ( 2019 ) detected landslides from hyperspectral remote sensing data using a deep learning technique.

The landslide inventory dataset in the current study consist a total of 576 landslides which were identified from Google Earth image interpretation and intensive field survey. For landslide susceptibility mapping landslide polygons can be divided into training and validation datasets. The training dataset is used for constructing the predictive model while the validation dataset is used for validating the model. In this study, the specific date of landslide occurrence is not well known. Hence, the landslide polygons were randomly split into two classes with 80% for training and 20% for validation by keeping their spatial distribution into account (Fig. 3 and 4 ). In addition, the validation data sets for most of the landslide susceptibility or hazard assessments were chosen in between 20% and 30% of the total landslide inventory.

Flowchart of the research work (Note: LS = landslide)

Landslide inventory map showing the distribution of landslides

Database for landslide causative factors

To undertake landslide susceptibility analysis in the study area, a spatial database was first constructed for the causative factors within the spatial analysis tools of ArcGIS 10.4 software. The database consists of the landslide inventory datasets (training and validation) and the landslide causative factors (slope, aspect, curvature; land use, lithology, rainfall and distance from stream). These factors were subsequently evaluated by calculating their weights from the relationship between the landslide and landslide causative factors and then these results were verified. There are no strict rules or guidelines for the triggering factors to be used in different statistical approaches for landslide susceptibility mapping. Instead, the chosen factors should be operative and measurable depending on a particular area’s characteristics (Ayalew and Yamagishi 2005 ). One parameter may be an important controlling factor for landslide occurrence in a certain area but in most cases a combination of two or more landslide causative factors may be effective in addition to the triggering factor for landslide occurrence.

In this study, the triggering factor was heavy and prolonged rainfall. During the fieldwork, landslide locations were identified and marked with GPS, land use (land cover) types around the landslide scar, drainage networks and spring locations, lithological units and human activities were investigated to prepare the landslide susceptibility maps.

Generally, the selection of landslide causative factors should consider the nature of the study area and the availability of data. In this regard, a total of seven parameters were selected including slope, aspect, curvature, lithology, rainfall, land use and distance to stream. All causative factor maps were converted into raster maps with the same coordinate system (WGS 1984 UTM zone 37N) and the same pixel size (30mx30m). The rasterized training (80%) landslide map and all the causative factor maps have been overlaid and the information was extracted using the spatial analyst tool of ArcGIS to calculate the ratings or weights of all factor classes for FR and WoE models. The summation of these ratings or weights of each landslide factor will help to evaluate the spatial relationship between them and the probability of landslide occurrence in the study area.

Topographic parameters like slope, aspect, curvature and distance to stream maps were derived from Digital Elevation Model (DEM) with a cell size of 30 m by 30 m. Lithology and land use maps were prepared from intensive fieldwork and Google Earth image interpretations. The rainfall map was generated using IDW interpolation technique of the spatial analyst tool in ArcGIS from four rain gage stations near to the study area using the rainfall data from National Meteorological Agency of Ethiopia.

Frequency ratio (FR) model

Frequency Ratio model is a well-known and widely used bivariate statistical method that is used for landslide susceptibility mapping (Lee and Talib 2005 ; Akgun 2012 ; Demir et al. 2013 ; Mezughi et al. 2011 ; Yalcin et al. 2011 ; Abay and Barbieri 2012 ; Mondal and Maiti 2013 ; Paulin et al. 2014 ). In this model, processing the input data, computations and output-processes are very simple and can be easily understood. It is simple and relatively flexible to use and implement a landslide susceptibility map with accurate results (Lee and Pradhan 2007 ; Yilmaz 2009 ; Choi et al. 2012 ; Mohammady et al. 2012 ; Park et al. 2012 ). The frequency ratio model is one of the probabilistic models which are based on the observed relationship between the distribution of landslides and each landslide related factor (Lee and Talib 2005 ). To evaluate the contribution of each factor towards landslide susceptibility, the training landslide group was combined with thematic data layers separately and then the frequency ratio of each factor’s class was calculated according to the following procedures.

First, the number of pixels for landslide occurrence and non-occurrence in each factor’s class was calculated. Second, the percentage of each factor’s class having landslide to the total pixels containing landslide of the factor was calculated and the percentage of each factor class’s number of pixels to the total number of pixels in the study area was calculated. Finally, the frequency ratio of each factor class was obtained by dividing the percentage of landslide pixels to the percentage of area pixels in each factor classes (Equation 1 ).

Where; Npix ( S i , j ) = the number of pixels containing landslide within class j in factor i; Npix ( N i , j ) = the number of pixels of class j in factor i; ∑ j NPix ( S i , j ) is the number of total pixels containing landslide in the study area; ∑ j NPix ( N i , j ) is the number of total pixels in the study area.

The calculated FR value represents the degree of correlation between landslide and a certain class of the causative factor. A value of 1 is an average value for the landslide occurrence of a specific landslide causative factor class. A value more than 1 indicates a strong and positive correlation and a high probability of landslide occurrence, while a value of less than 1 indicates a negative relationship and low probability of landslide occurrence in a certain class of a landslide causative factor. The FR map of each causative factor is prepared with the help of ArcGIS by assigning the calculated FR values. Then the FR values of all the causative factor maps were overlaid and numerically added using the raster calculator of the spatial analyst tool in ArcGIS 10.4 to prepare the Landslide Susceptibility Index (LSI) map. LSI is computed by summing the FR values of all the landslide causative factor maps (Equation 2 ) and then the resulting LSI map was further reclassified in to very low, low, moderate, high and very high landslide susceptibility classes.

Where: LSI = Landslide susceptibility index, FR is the frequency ratio and n is the number of selected causative factors. The calculated values of FR for each pixel in the LSI indicate the relative susceptibility to landslide occurrence. The higher LSI pixel values have high susceptibility to landslide occurrence while the lower LSI pixel values have lower susceptibility (Akgun et al. 2007 ).

Weights of evidence (WoE) model

WoE model is a log-linear form of the Bayesian probability model for landslide susceptibility assessment that uses landslide occurrence as a training point to drive prediction outputs. It calculates both unconditional and conditional probability of landslide hazards. This method is based on the calculation of positive and negative weights to define the degree of spatial association between landslide occurrence and each explanatory variable class (Pardeshi et al. 2013 ). The positive weights (W+) indicate the occurrence of an event while the negative weight (W-) indicates the non-occurrence of an event. To evaluate W + and W - , calculating the following parameters is important.

Nmap = total number of pixels in the map

Nslide = number of pixels with landslides in the class

Nclass = number of pixels in the class

NSLclass = number of pixels with landslides in the class

The values needed for the weight of evidence formula are:

Npix1 = NSLclass

Npix2 = Nslide – NSLclass

Npix3 = Nclass – NSLclass

Npix4 = Nmap – Nslide – Nclass + NSLclass

Then the positive and negative weights are calculated as follows (Equations 3 and 4 ).

Where Npix 1 is the number of landslide pixels present on a given factor class, Npix 2 is the number of landslides pixels not present in a given factor class, Npix 3 is the number of pixels in a given factor class in which no landslide pixels are present and Npix 4 is the number of pixels in which neither landslide nor the given factor is present (Van Westen 2002 ; Dahal et al. 2008 ; Regmi et al. 2010 ). These weights are used to calculate a weight of contrast value (C) for the particular susceptibility variable (Equation 5 ).

The contrast value (C) measures the strength of a relationship between the causative factors and landslides. If the contrast value is positive, it will have a positive spatial association while the negative one will have a negative spatial association. The weighted map ( Wmap ) for each landslide causative factor can be prepared by summing the weights of contrast(C) values of each factor class. Similarly, the final landslide susceptibility index (LSI) map was prepared by summing all the weighted maps (∑ Wmap )  of each landslide causative factor through a raster calculator of map algebra in the spatial analyst tool of ArcGIS as follows (Equations 6 and 7 ).

Landslide inventory

During August, 2018, an intense rainfall in Simada area triggered many landslides that occurred mostly in rural areas. The damage was severe in the villages of Dubdubiya, Asfa Meda, Gedeba, Ditorka and at several other sites along the river courses. Particularly, in Dubdubiya and Asfa Meda villages, landslides damaged 81 dwellings, killed 14 goats, affected thousands of people, damaged hundreds of hectares of farmlands and dislocated 486 people. These problems occurred in these villages as the settlement areas are mostly located at the foot of a steep slope that is covered by weathered volcanic rocks as well as the presence of stream accumulated debris and earth flows that can suddenly burst out at the at the outlets of a mountain. Landslide inventory map of the study area (Fig. 4 ) was prepared from the combination of an intensive field survey and Google Earth image interpretations. Extensive field studies conducted from mid-November to mid-December of 2018 helped us to map known landslides using GPS and check the size and shape of these landslides in order to identify the type of movements, materials involved and to determine the state and activity of landslides (active, reactivated, dormant, etc.). This inventory data was mapped as vector-based polygon data and then converted to the raster format with a pixel size of 30m by 30m in ArcGIS 10.4.

In the present study area, a total of 576 landslides that contain 6304 pixels were identified and divided randomly into training and validation landslides by keeping their spatial distributions into account. The training landslides that accounted 80% of landslides with 5126 pixels were used for building the predictive model while the validation landslides that accounted 20% of landslides with 1178 pixels were used for validating purpose. From the total landslide polygons, 117 landslides were active landslides collected from field investigations while the remaining 459 landslide polygons were collected from time serious Google Earth image interpretations.

Landslide locations are predominantly distributed in the south-central, in the north and in the eastern parts of the study area with decreasing order of landslide density, damage on agricultural land and infrastructures. This area consists of a rugged and mountainous terrain which is characterized by steep slopes, deep gorges, high relief and fractured and weathered rocks. The common types of landslide occurrence in the study area include rock slide, rockfall, earth slide, debris slide and debris flow, rotational and translational soil slide, translational debris slide, rotational debris slide and complex types of slides. Generally, these landslides predominantly affected the rural areas in which the type of landslides and their probable causes and damages are described below.

Most prominent landslides occurred in Asfa Meda, Dubdubiya, Tej Wuha-Gedeba and Ditorka-Megersum Villages. Landslides in Asfa Meda Village occurred at the interface between thin residual soils and rhyolitic rock and most of the landslides are shallow rotational and/or translational earth slides. Most of Dubdubiya village was highly affected by stream undercutting, erosion of the slope surface, riverbank erosion and improper farming practice (Fig. 5 ). The slope materials are dominantly covered by weathered basalt and colluvial deposits. Erosional opening surfaces and tension cracks were observed during field investigation indicating that seeping water might have brought instability of the slope through internal erosion of the weathered materials. A typical example of a landslide in this village was the landslide that occurred near Arata Gabriel Church. The main causes of this landslide were stream/river undercutting, presence of spring on top of the slope and colluvial soil slope materials. The slope material in Teji Wuha and Gedeba Villages is dominantly covered with weathered tuff and thin residual soils. In this village, there is an indication of shallow groundwater since the swamp area and many springs are observed with rotational and soil creep. Creeping of soil was identified by tilting of powerlines and fences (Fig. 6 d). The common types of landslides that were observed in Ditorka and Megersum villages were rockslide (Fig. 6 a), rock fall, debris slide (Fig. 6 b) and rotational slide.

Historical landslide data in Dubdubiya Village as detected from Google Earth image. The white arrow and the red broken line in the figures indicate the direction of the slope movement and a landslide boundary respectively. a Image before the landslide event. b Image after landslide occurrence and, c Photo of the current condition

Field photo of landslide and their damage in different villages of the study area (Red and blue colored arrow indicates the direction of slope movements and River/stream flows respectively). In Ditorka and Megersum Villages- ( a ) Rockslide ( b ) Debris slide; Teji Wuha and Gedeba Villages- ( c ) Earth slide on the gravel road (damaged culvert) ( d ) Creep (tilting of powerline); in Asfa Meda Village- ( e ) and ( f ) Complex type landslides; in Dubdubiya Village- ( g ) Shallow debris flow that damaged wheat crop and ( h ) Rockslide

Landslide causative factors

The spatial distribution and density of landslides are mainly controlled by topography of an area, weather condition, geology, land use/land cover and anthropogenic factors (Khan et al. 2019 ). Consequently, evaluating the impact of these causative factors on the spatial distribution of landslides is very important in order to understand their failure mechanism and to prepare the landslide susceptibility map. In this study, seven causative factors that have been used for the preparation of landslide susceptibility maps include slope, aspect, curvature, lithology, land use/ land cover, rainfall and distance to stream. The roles played by each of these causative factors will be discussed in the following sections.

Slope is a very important parameter for landslide study as it has a direct relation with landslide occurrence. As a result, it is frequently used in preparing a landslide susceptibility map (Yalcin and Bulut 2007 ). It is well known that landslide occurs more frequently on steeper slopes due to gravity stress. The slope map (Fig. 7 a) of the study area was prepared from DEM data. It was divided into five classes such of 0 – 5 0 , 5 0 – 12 0 , 12 0 – 30 0 , 30 0 – 45 0 , and > 45 0 . For slope classes above 12 0 , the frequency ratio is increasing which indicate the higher probability of landslide occurrence in these classes (Table 1 ).

Landslide causative factor maps ( a ) Slope, ( b ) Curvature, ( c ) Distance to stream ( d ) Aspect ( e ) Land use, ( f ) Lithology and ( g ) Rainfall

Curvature map of the study area was generated from DEM data and it was classified into 3 classes of concave, convex and flat surfaces (Fig. 7 b). Following heavy rainfall, a convex or concave slope contains more water and retains this water for a longer period (Lee and Talib 2005 ). The more positive or negative values indicate the higher probability of landslide occurrence. In the flat area, the probability of landslide occurrence is very low. A positive curvature indicates that the surface was upwardly convex at that grid. A negative curvature indicates that the surface was upwardly concave at that grid and a value of zero indicates that the surface is flat.

Aspect refers to the slope orientation which is generally expressed in terms of degree from 0 0 – 360 0 . It is considered as an important factor in landslide studies as it controls slope’s exposure to sunlight, wind direction, rainfall (degree of saturation) and discontinuity conditions (Komac 2006 ). Slope aspect map (Figure 8 c) in this study area was derived from DEM data and it was divided into nine classes, namely; north (0 – 22.5, 337.5 - 360,), northeast, east, southeast, south, southwest, west and northwest (Fig. 7 d).

Landslide Susceptibility Map of the study area using: a FR and b WoE Models

Distance to stream

The proximity of the slope to the stream course is an important factor that dictates the landscape evolution of the area and an indicator of the landslide and related erosional aspects. Rivers with a number of drainage networks have a high probability of landslide occurrence as they erode the slope base and saturate the underwater section of the slope forming material (Akgun and Turk 2011 ).

Since there are many streams in the study area which flow into Kostet, Atkus and Bijena Rivers, many landslides occurred in the close vicinity of these rivers. Hence, this parameter was considered as one causal factor in landslide susceptibility analysis. Zones with parallel pattern of drainage in steep slopes are the most probable landside sites. Drainage often plays its own role in developing pore-water pressure which reduces the shear strength of slope materials. Streamlines were derived from DEM data and it was classified based on stream order.

Landslide in this area is mostly associated with 1 st , 2 nd , and 3 rd order streams. Distance from stream map was developed from Euclidean distance buffering method in the spatial analyst tool of ArcGIS 10.4. This map was classified in to five subclasses: 0 – 50, 50 – 100, 100 – 150, 150 – 200 and > 200 meter (Fig. 7 c).

Land use / land cover

Land-use change has been recognized throughout the world as one of the most important factor influencing the occurrence of rainfall-triggered landslides. Changes in land use/cover resulted from man-made activities such as deforestation, overgrazing, intensive farming and cultivation on steep slope can initiate slope instability (Glade 2003 ). Vegetation has a major contribution to resist slope movements. Vegetation having a well-spread network of root systems increases shearing resistance of the slope material. This is due to the natural anchoring of slope materials. In addition to this, it reduces the action of erosion and adds the stability of the slope. In another way, barren or sparsely vegetated slopes are usually exposed to erosion and thus it has the effect of increasing slope instability. The land use map of the study area was prepared from the Google Earth image of 2016 and the analysis was done in ArcGIS. About seven land-use types were identified including moderate forest, sparse forest, bush, grazing land, agricultural land, settlement and river (Fig. 7 e). The area is predominantly covered by agricultural land and grazing land.

Lithology is one of the most controlling parameters in slope stability since each class of materials has different shear strength and permeability characteristics (Yalcin and Bulut 2007 ). Different rock types have varied composition and structure which contribute to the strength of the slope material in a positive or negative way. The stronger rock units give more resistance to the driving forces as compared to the softer/ weaker rocks. Lithological map of the study area was prepared from existing regional geological map (with a scale of 1:250,000) as a preliminary map for further improvement of a lithologic map into a scale of 1:50,000 based on a detailed field survey. The study area contains seven lithological units namely Trachyte, Weathered tuff, Rhyolite, Weathered basalt, Residual soils, Colluvial and Alluvial Deposits (Fig. 7 f).

Rainfall is considered as an influencing factor to cause slope instability. Precipitation, particularly intense and prolonged rains are controlling factors that trigger landslides by providing water thereby increasing underground hydrostatic level and pore water pressure. When the soil undergoes such pressure changes, water within it will create negative or upward pressure, as it cannot drain quickly. When the pore water pressure is equivalent to the upper pressure, the shearing resistance of the material decrease and will lead to failure of the material. The rainfall data of the four stations that surround the study area were collected from National Metrology Agency of Ethiopia. There are various interpolation techniques in ArcGIS to interpolate rainfall over a large area based on few point data. These include Thiessen polygon, Isohyetal, average arithmetic, inverse distance weight (IDW) and Kriging. The general assumption of the IDW method of interpolation is that the value of unsampled point is the weighted average of known values within the neighborhood. Therefore, the values from a scattered set of known points can be utilized to assign rainfall values to unknown points. It can be used to compute the unknown spatial rainfall data from the known sites that are adjacent to the unknown sites (Chen and Liu 2012 ). The rainfall map of the study area was prepared using the IDW interpolation method in GIS. The rainfall data analysis showed that the maximum monthly rainfall occurs in June, July, August and September which coincides with the landslide occurrence in this area. The rainfall map of the study area was divided into five annual rainfall classes of 627 – 727, 727 – 813, 813 – 901, 901 – 994 and 994 - 1125.2 millimeters (Fig. 7 g) by the natural breaks method.

Result and discussion

Relationship between landslide and causative factors.

This study has analyzed the relationship between seven causative factors and landslide occurrence. Using the FR and WoE models, the relative frequency values and the weights of values were calculated respectively. The causative factors were classified into different classes and weights were assigned to them for both FR and WoE models as presented in Table 1 and 2 respectively. These results showed that the relative susceptibility of each class is almost similar for both models but the parameters and results are different from each other. This implies that if a factor class has lower and higher values in both models, the susceptibility will also be lower and higher respectively. In case of FR model, the spatial relationship between the causative factors and landslide is determined by FR values. The causative factor classes with FR value > 1 will have a high degree of landslide occurrence. On the other hand, for the WoE model, C describes the correlation and spatial association of the landslide with the causative factors. The positive C values indicate a positive association with more landslide occurrence and vice versa for negative C values. The weights with higher values indicate a higher degree of influence on landslide occurrence. Generally, the factor class values derived from each model showed the spatial relationship of the causative factors in their contribution to landslide occurrence. The association is more or less the same in both models.

The slope classes > 12 0 have higher contribution for landslide occurrence. The area with a slope class > 45 0 is the most landslide prone class while the area with a slope class < 5 0 is the least one. Generally, as the slope increases, the probability of landslide occurrence also increases. In case of aspect classes, the FR values of slope classes facing towards the northeast (22.5 – 67.5), east (67.5 – 112.5) and north (0 – 22.5) are greater than one indicating a higher probability of landslide occurrence. The northeast facing aspect class has got the maximum weight or rating followed by the east facing ones. The curvature range of (-3.6) - (-0.001) has a greater contribution to the slope failures. In case of lithology, three units i.e. colluvial deposit, weathered basalt and rhyolite have high probability of landslide occurrence. Colluvial deposit and weathered basalts have less strength and hence susceptible to landslides. Rhyolitic rocks in the study area formed a cliff underlying thin residual soils. As a result, most of the landslides occurred at the contact between rhyolite and thin residual soils.

The type of land use also controls the occurrence of landslide in the study area. The highest weights or ratings were observed in the land use types of grazing land, river, sparse forest and bushes indicating a high probability of landslide occurrence. The highest weighted value of grazing land is due to its exposure to erosion and weathering. In case of the relationship between landslide occurrence and the distance from stream, as the distance from stream increases, the occurrence of landslide generally decreases. Landslide occurrence is higher in the first three classes of 0 – 50m, 50 – 100m and 100 – 150m (Table 1 and 2 ). With regard to the causative factor rainfall, two classes with 813 – 901mm and 901 - 994 mm have a higher C and FR values than the other classes and are the most susceptible classes (Table 1 and 2 ). Generally, slope classes > 20 0 , land use classes of grazing land, sparse forest, river and bush; lithology of colluvial deposit, weathered basalt, alluvial deposit and rhyolite and distance to stream classes of < 150 m buffers are the most contributing factor classes among the seven landslide factor classes.

Landslide susceptibility mapping using FR and WoE models

Frequency ratio model.

Map of each causative factor is prepared with the help of ArcGIS and then the frequency ratio values were calculated. The calculated FR values for each pixel in the LSI indicate the relative susceptibility to landslide occurrence. The higher pixel values of LSI have the higher landslide susceptibility while the lower pixel values will have lower susceptibility (Akgun et al. 2007 ). The landslide susceptibility index was calculated based on the frequency ratio values that have been determined in the training process that can be added in a raster calculator of ArcGIS as follows (Equation 8 ).

Where FR sl = frequency ratio value of slope, Fr as = frequency ratio value of aspect, = FR cu = frequency ratio value of curvature, FR li = frequency ratio value of lithology, FRlu = frequency ratio value of land use, FR rf = frequency ratio value of rainfall, FR ds = frequency ratio value of distance to stream.

The LSI values for the frequency ratio model in the study area range from 2.89 to 15.09. The LSI map is reclassified to prepare the landslide susceptibility map of the study area (Fig. 8 a). There are different types of classification methods such as natural break, equal interval, manual, standard deviation and quantile. In the current study, reliable results were obtained from natural breaks method. The result of other classification methods revealed the susceptibility classes with a high degree of exaggeration where large part of the study area fall into the high susceptibility class.

Therefore, the LSI values were classified into five susceptibility classes of very low (2.89 - 5.31), low (5.31 - 6.24), moderate (6.24 - 7.23), high (7.23 - 8.39) and very high (8.39 - 15.09) using the natural breaks method of classification. The result from Table 3 showed that 8.616% (16 km 2 ), 20.474%(38km 2 ), 29.537%(54.9km 2 ), 27.898% (51.8 km 2 ) and 13.474% (25 km 2 ) areas fall into the very low, low, moderate, high and very high susceptibility classes respectively. As Fig. 8 a clearly shows, the very low and low susceptibility classes are dominantly concentrated in the northwestern and southwestern plateau part of the study area including Welela Bahir, Shomeda, Agona and Jinjero Gedel localities. Similarly, the very high and high susceptibility classes are concentrated in the south central, southeastern and eastern part of the study area particularly in Asfa Meda (Majeta), Dubdubiya (Arata Gebriel) and Ditorka-Megersum respectively and scarcely distributed in the northern part of the study area at Guna-Gedeba Village and in the western part. Moderate susceptibility classes are mostly distributed throughout the study area. The high concentrations of landslides in those high and very high susceptibility classes of the aforementioned areas were due to the presence of colluvial and alluvial deposits, stream undercutting, scattered vegetation cover, man-made activities like intensive farming, deforestation and cultivation.

Weights of evidence model

The landslide susceptibility map of the study area by WoE model was produced based on the weighted values from the seven causative factors and the training landslide (Table 2 ). The difference between W + and W - is known as the weight of contrast which is designated by C = W + - W - . This reflects the overall spatial association between the causative factors and landslides. LSI map of the study area was prepared by summing the weight of contrast values (C) of all the seven causative factors using a raster calculator in ArcGIS as follows:

Where LSI = landslide susceptibility index; C sl = weight contrast value of slope, C as = weight contrast value of aspect, C cu = weight contrast value of curvature, C li = weight contrast value of lithology, C lu = weight contrast value of land use, C rf = weight contrast value of rainfall, C ds = weight contrast value of distance to stream.

The LSI values for the WoE model in the study area range from -7.84 to 4.52. The LSI map is reclassified by the natural breaks method of classification technique in order to prepare the landslide susceptibility map of the study area (Fig. 8 a). Then, the LSI values were classified in to five susceptibility zones of very low (-7.84 - -3.72), low (-3.72 - -1.83), moderate (-1.83 - -0.28), high (-0.28 – 1.17) and very high (1.17 – 4.52). The result from Table 4 showed that 8.448%(15.7 km 2 ), 21.408%(39.8km 2 ), 33.140%(61.5km 2 ), 23.787% (44.2 km 2 ) and 13.216% (24.5 km 2 ) area fall in the very low, low, moderate, high and very high susceptibility classes respectively.

Validation of the model

Without model validation, landslide susceptibility maps will not be meaningful. As a result, validation of the predictive model is an important step for landslide susceptibility mapping (Bui et al. 2012 ). A predictive model map was constructed by overlying 80% of the landslides (training) over the causative factors. This model was validated using validation landslides (20%) that were not used for building the model. There are various types of validation techniques for landslide susceptibility maps. In the current study, the performance of the LSMs produced by FR and WoE models were evaluated using Area Under the Curve (AUC) and Landslide Density Index (LDI).

Area under the curve (AUC)

The area-under-curve (AUC) method works by creating success rate and prediction rate curves (Lee 2005 ). Landslide susceptibility maps can be validated by comparing the susceptibility maps with both the training landslide (80%) and validation landslide (20%). The success and predictive rate curves can be created for both FR and WoE models. The success rate curve is based on the comparison between the predictive model and the training landslide. The predictive rate curve is based on the comparison between the predicted map and the validation landslide. The Area Under the Curve (AUC) of the success rate represents the quality of the model to reliably classify the occurrence of existing landslides whereas the AUC of the predictive rate explains the capacity of the proposed landslide model for predicting landslide susceptibility (Pamela et al. 2018 ). AUC was calculated by reclassifying LSI into 50 classes with descending order of the values of pixels in the study area and combined with a landslide inventory. Then the rate curves were drawn through the cumulative percentage of both the training and validation landslide (y-axis) and cumulative area percentage (x-axis). The result showed that both models exhibited very good performances. However, the FR model is better with a success rate of 89.8% and a predictive rate of 88.2% than the WoE model with a success rate of 86.5% and a predictive rate of 84.8% (Fig.  9 ).

Success and Predictive rate Curves of FR and WoE models

Landslide density index (LDI)

For validation of the model, landslide pixels which have not been used for constructing the models are generally considered as the future landslide area. In this work to check the validation of the landslide susceptibility model, the testing samples that consist of 20% of the landslide pixels were overlaid over the landslide susceptibility map. The landslide density index, which is the ratio between the percentage of landslide pixels and the percentage of class pixels in each class on landslide susceptibility map, was used to validate the model (Pham et al. 2015 ). If the value of the landslide density index is increased from low to a very high susceptibility classes, then the landslide susceptibility map is considered to be valid. LDI can be calculated using the formulae in eq. 10 below and its output was presented in Table  5 . The suitability of any susceptibility map can be validated if more percentages of landslides occur in the high and very high susceptibility zones as compared to other zones (Fayez et al. 2018 ).

From Table 5 , it can be observed that the landslide density values for very high susceptibility classes are 2.743 and 2.993 with respect to WoE and FR models which are remarkably higher than the other classes. In addition to this, there is a gradual decrement in landslide density values from very high to very low susceptibility classes (Fig.  10 ). This indicates the validity of the landslide susceptibility map. Can et al. ( 2005 ) and Bai et al. ( 2010 ) stated that the landslide data should lie in either the high or very high susceptibility classes for successful validation of a LSM. Table 5 , Figs.  10 and 11 illustrate the characteristics of susceptibility classes for validation of both FR and WoE models. In general, the presence of the highest landslide percentage and density in the very high landslide susceptibility class indicates the reliability of the landslide susceptibility maps produced from FR and WoE models.

Landslide density of FR and WoE models for both training and validation landslide

Bar diagram showing the landslide percentages in different landslide susceptibility classes

Landslide posed a significant impact at Simada District of South Gondar Zone in northwestern Ethiopia on human and animal lives, agricultural lands, settlements, infrastructures and also affected the social and economic aspects of the rural community. To investigate this problem, landslide susceptibility mapping has been carried out using FR and WoE models for proper land use planning, development and management of landslide prone areas. For this, a landslide inventory map of the study area with a total of 576 landslides was divided into training and validation landslides with 80 % and 20% respectively. Seven landslide causative factors including slope, aspect, curvature, lithology, land use, rainfall and distance to stream were considered to analyze, evaluate and establish the spatial relation of these factors with landslides. From FR values and WoE contrast values, it was possible to identify which factor classes are playing a significant role for the occurrence of landslides in the study area. The FR values that are greater than 1 and the WoE contrast (C) values that are greater than 0 were found in the factor classes of slope greater than 12°; curvature classes (-3.60) - (-0.001); aspect classes facing towards N (0 – 22.5), NE (22.5 – 67.5) and E (67.5 – 112.5); distance to stream classes (< 150m); land use classes (grazing land, river, sparse forest and bushes); lithology classes (colluvial deposit, alluvial deposit, weathered basalt and rhyolite), rainfall classes (813 – 901mm and 901 – 994 mm). The LSI map of the study area was prepared based on FR values and WoE contrast values in ArcGIS 10.4 using the spatial analyst tools of raster calculator for both FR and WoE models. The LSI map in each model was reclassified into five landslide susceptibility classes of low, low, moderate, high and very high based on the natural breaks method of classification to produce the final landslide susceptibility maps. The performance of the final landslide susceptibility maps produced by FR and WoE models were validated using Landslide Density Index (LDI) and Area Under the Curve (AUC) values. The result revealed that the very low, low, moderate, high and very high values of the landslide susceptibility map are comparable with Landslide Density Index. In case of AUC, the rate curves were drawn using the cumulative percentage of the landslide in the Y-axis and cumulative percentage of map area in the X-axis. The results showed that both models exhibited very good performance. However, the FR model, which showed a success rate of 89.8% and a prediction rate of 88.2%, is better than the WoE model with a success rate of 86.5% and a prediction rate of 84.8%. This study confirmed that the bivariate statistical methods of FR and WoE models were found to be simple and effective models for landslide susceptibility mapping in the Guna mountainous chain of Simada area. The landslide susceptibility maps of the study area were prepared with a scale of 1:50,000 which can be used by civil engineers, geologists, designers and decision-makers for regional land use planning, site selection and landslide prevention and mitigation purposes.

Recommendation

The present study showed the importance of integrating various factors that are responsible for landslide occurrence in the study area. However, the quality of landslide inventory and the causative factor maps should be improved with good quality in time and space. Landslide in the study area has affected the local people who are living near to mountainous area, valleys and gorges. Their animals were died, houses and agricultural lands were destroyed and both social and economic activities were affected. Hence, besides preparing the landslide susceptibility maps of the area, suggesting the necessary preventive measures in the high and very high susceptibility classes is very essential in order to reduce the impact of future landslide hazards in the area. Hence, this study recommends planting trees & vegetation, providing proper drainage, applying gabion and check dam, relocating people and creating public awareness. In order to implement these remedial measures, further study on the geotechnical properties of soils and rocks should be conducted in this area.

Availability of data and materials

Rainfall data was collected from National Metrology Agency of Ethiopia. Topographic Map was purchased from Ethiopian Geospatial Information Agency. DEM data was freely available from http://gdex.cr.usgs.gov/gdex/ website.

Abay A, Barbieri G (2012) Landslide susceptibility and causative factors evaluation of the landslide area of Debresina, in the southwestern Afar escarpment, Ethiopia. J Earth Sci Eng 2:133–144

Google Scholar  

Abebe B, Dramis F, Fubelli G, Umar M, Asrat A (2010) Landslides in the Ethiopian highlands and the rift margins. African Earth science 56:131–138

Article   Google Scholar  

Akgun A, Dag S, Bulut F (2007) Landslide susceptibility mapping for a landslide-prone area (Findikli, NE of Turkey) by likelihood-frequency ratio and weighted linear combination models. Eng Geol V 54:1127–1143

Akgun A, Turk N (2011) Mapping erosion susceptibility by multivariate statistical method: a case study from the Ayvalik region. NW Turkey Computers & Geosciences, Turkey

Akgun AA (2012) Comparison of landslide susceptibility maps produced by logistic regression, multi-criteria decision, and likelihood ratio methods: a case study at İzmir, Turkey. Landslides 9:93–106

Aleotti P, Chowdhury R (1999) Landslide Hazard assessment: summary, review and new perspectives. Bull Eng Geol Environ 58:21–44

Ayalew L (1999) The effect of seasonal rainfall on landslides in the highlands of Ethiopia. Bull Eng Geol Environ 58:9–19

Ayalew L, Yamagishi H (2005) The application of GIS based logistic regression for landslide susceptibility mapping in the KakudaYahiko Mountains Central Japan. Geomorphology 65(1):15–31

Baharin A, Farshid S, Javad M, Barat M (2014) Using Frequency Ratio Method for Spatial Landslide Prediction. Res J App Sci Eng Technol. 7:3174–3180

Bai SB, Wang J, Lu G-N, Zhou P-G, Hou S-S, Xu S-N (2010) GIS-based logistic regression for landslide susceptibility mapping of the Zhongxian segment in the three gorges area, China. Geomorphology 115:23–31 https://doi.org/10.1016/j.geomorph.2009.09.025

Bui DT, Pradhan B, Lofman O, Revhaug I, Dick OB (2012) Spatial prediction of landslide hazards in Hoa Binh province (Vietnam): a comparative assessment of the efficacy of evidential belief functions and fuzzy logic models. Catena 96:28–40

Can, Nefeslioglu HA, Gokceoglu C, Sonmez H, Duman TY (2005) Susceptibility assessments of shallow earthflows triggered by heavy rainfall at three catchments by logistic regression analyses

Book   Google Scholar  

Chen FW, Liu CW (2012) Estimation of the spatial rainfall distribution using inverse distance weighting (IDW) in the middle of Taiwan. Paddy Water Environ 10:209–222

Choi J, Oh H-J, Lee H-J, Lee C, Lee S (2012) Combining landslide susceptibility maps obtained from frequency ratio, logistic regression, and artificial neural network models using ASTER images and GIS. Eng Geol 124:12–23

Crozier MJ (1986) Landslides: causes, consequences & environment. Croom Helm Pub, London

Cruden DM (1991) A simple definition of a landslide. Bulletin of the International Association of Engineering Geology 43(1):27–29. https://doi.org/10.1007/bf02590167

Cruden DM, Varnes DJ (1996) Landslide types and processes. National Research Council, Washington DC

Dahal RK, Hasegawa S, Nonomura A, Yamanaka M, Dhakal S, Paudyal P (2008) Predictive modelling of rainfall-induced landslide hazard in the Lesser Himalaya of Nepal based on weights-of-evidence. Geomorphology 102:496–510

Dai FC, Lee CF, Ngai YY (2002) Landslide risk assessment and management: an overview. Eng Geol 64:65–87

Demir G, Aytekin M, Akgün A, Ikizler SB, Tatar O (2013) A comparison of landslide susceptibility mapping of the eastern part of the North Anatolian Fault zone (Turkey) by likelihood-frequency ratio and analytic hierarchy process methods. Nat Hazards 65:1481–1506

Dikau R, Brunsden D, Schrott L, Ibsen M-L (eds) (1996) Landslide recognition. Identification, movements and causes. Wiley, Chichester, p 251

Fayez L, Pham BT, Solanki HA, Pazhman D, Dholakia MB, Khalid M, Prakash I (2018) Application of frequency ratio model for the development of landslide susceptibility mapping at part of Uttarakhand state, India. Int J Appl Eng Res 13:6846–6854

Felicisimo ÁM, Cuartero A, Remondo J, Quirós E (2012) Mapping landslide susceptibility with logistic regression, multiple adaptive regression splines, classification and regression trees, and maximum entropy methods: a comparative study. Landslides 10:175–189

Gholami M, Ghachkanlu EN, Khosravi K, Pirasteh S (2019) 2018 Landslide prediction capability by comparison of frequency ratio, fuzzy gamma and landslide index methods. J Earth Syst Sci 128:42 https://doi.org/10.1007/s12040-018-1047-8 Indian Academy of Sciences

Girma F, Raghuvanshi TK, Ayenew T, Hailemariam T (2015) Landslide hazard zonation in Ada Berga district, Central Ethiopia – a GIS-based statistical approach. J Geom 90:25–38

Glade T (2003) Landslide occurrence as a response to land-use change: a review of evidence from New Zealand. Catena 51(3–4):297–314

Glade T, Anderson M, Crozier MJ (eds) (2005) Landslide Hazard and Risk. https://doi.org/10.1002/9780470012659

Guzzetti F, Mondini AC, Cardinali M, Fiorucci F, Santangelo M, ChangM KT (2012) Landslide inventory maps: new tools for an old problem. Earth Sci Rev 112:42–66 Elsevier

Hong H, Xu C, Bui DT (2015) Landslide susceptibility assessment at the Xiushui area (China) using frequency ratio model. Procedia Earth and Planetary Science 15:513–517

Hutchinson JN (1989) General report: morphological and geotechnical parameters of landslides in relation to geology and hydrogeology. International Journal of Rock Mechanics and Mining Sciences & Geomechanics Abstracts 26(2):88. https://doi.org/10.1016/0148-9062(89)90310-0

Kanungo D, Arora M, Sarkar S, Gupta R (2009) Landslide susceptibility zonation (LSZ) mapping-a review. J South Asia Disaster Stud 2:81–105

Kanungo DP, Arora MK, Sarkar S, Gupta RP (2006) A comparative study of conventional, ANN black box, fuzzy and combined neural and fuzzy weighting procedures for landslide susceptibility zonation in Darjeeling Himalayas. Eng Geol 85:347–366

Khan H, Shafique M, Khan M, Bacha M, Shah S, Calligaris C (2019) Landslide susceptibility assessment using Frequency Ratio, a case study of northern Pakistan. Egyptian J Remote Sensing Space Sci 22(1):11–24. https://doi.org/10.1016/j.ejrs.2018.03.004

Komac M (2006) A landslide susceptibility model using the analytical hierarchy process method and multivariate statistics in perialpine Slovenia. Geomorphology 74(1–4):17–28

Lee S (2005) Application of logistic regression model and its validation for landslide susceptibility mapping using GIS and remote sensing data. Int J Remote Sens 26(7):1477–1491. https://doi.org/10.1080/01431160412331331012

Lee S, Pradhan B (2007) Landslide hazard mapping at Selangor, Malaysia using frequency ratio and logistic regression models. Landslides. 4:33–41

Lee S, Talib JA (2005) Probabilistic landslide susceptibility and factor effect analysis. Environ Geol 47:982–990

Article   CAS   Google Scholar  

Meten M, Bhandary NP, Yatabe R (2015a) Effect of landslide factor combinations on the prediction accuracy of landslide susceptibility maps in the Blue Nile gorge of Central Ethiopia. Geoenviron Disast 2:9 https://doi.org/10.1186/s40677-015-0016-7

Meten M, Bhandary NP, Yatabe R (2015b) GIS-based frequency ratio and logistic regression modeling for landslide susceptibility mapping of Debre Sina area in Central Ethiopia. J Mt Sci 12(6). https://doi.org/10.1007/s11629-015-3464-3

Mezughi TH, Akhir JM, Rafek AG, Abdullah I (2011) Landslide susceptibility assessment using frequency ratio model applied to an area along the E-W highway (Gerik-Jeli). Am J Environ Sci 7:43–50

Mohammady M, Pourghasemi HR, Pradhan B (2012) Landslide susceptibility mapping at Golestan Province, Iran: a comparison between frequency ratio, Dempster-Shafer, and weights-of-evidence models. J Asian Earth Sci. 61:221–236

Mondal S, Maiti R (2013) Integrating the analytical hierarchy process (AHP) and the frequency ratio (FR) model in landslide susceptibility mapping of shiv-khola watershed, Darjeeling Himalaya. Int J Disaster Risk Sci 4:200–212

Pamela, Sadisun IA, Arifianti Y (2018) Weights of evidence method for landslide susceptibility mapping in Takengon, Central Aceh, Indonesia. IOP Conference Series: Earth and Environmental Science, 118:012037. https://doi.org/10.1088/1755-1315/118/1/012037

Pan X, Nakamura H, Nozaki T, Huang X (2008) A GIS-based landslide hazard assessment by multivariate analysis landslides. J Jpn Landslide Soc 45(3):187–195

Pardeshi SD, Autade SE, Pardeshi SS (2013) Landslide hazard assessment: recent trends and techniques. SpringerPlus 2(1). https://doi.org/10.1186/2193-1801-2-523

Parise M, Jibson RW (2000) A seismic landslide susceptibility rating of geologic units based on analysis of characteristics of landslides triggered by the 17 January, 1994 Northridge, California earthquake. Eng Geol 58:251–270

Park S, Choi C, Kim B, Kim J (2012) Landslide susceptibility mapping using frequency ratio, analytic hierarchy process, logistic regression, and artificial neural network methods at the Inje area, Korea. Environ Earth Sci 68:1443–1464

Paulin L, Bursik G, Hubp M, Mejía JL, Aceves Quesada FLMP (2014) A GIS method for landslide inventory and susceptibility mapping in the Río El Estado watershed, Pico de Orizaba volcano, México. Nat Hazards 71:229–241

Peng L, Niu R, Huang B, Wu X, Zhao Y, Ye R (2014) Landslide susceptibility mapping based on rough set theory and support vector machines: a case of the three gorges area, China. Geomorphology 204:287–301 Elsevier B.V

Pham BT, Tien Bui D, Prakash I, Dholakia M (2015) Landslide susceptibility assessment at a part of Uttarakhand Himalaya, India using GIS–based statistical approach frequency ratio method. Int J Eng Res Technol 4:338–344

Pirasteh S, Li J (2017) Probabilistic frequency ratio (PFR) model for quality improvement of landslide susceptibility mapping from LiDAR-derived DEMs. Pirasteh Li Geoenviron Dis 4:19 https://doi.org/10.1186/s40677-017-0083-z

Pradhan B, Lee S, Buchroithner MF (2010) Remote Sensing and GIS-based Landslide Susceptibility Analysis and its Cross-validation in Three Test Areas Using a Frequency Ratio Model. Photogrammetrie-FernerkundungGeoinformation 2010(1):17–32. https://doi.org/10.1127/1432-8364/2010/0037

Rahmati O, Haghizadeh A, Pourghasemi HR, Noormohamadi F (2016) Gully erosion susceptibility mapping: the role of GIS-based bivariate statistical models and their comparison. Nat Hazards 82(2):1231–1258 https://doi.org/10.1007/s11069-016-2239-7

Regmi AD, Devkota KC, Yoshida K, Pradhan B, Pourghasemi HR, Kumamoto T, Akgun A (2013) Application of frequency ratio, statistical index, and weights-of-evidence models and their comparison in landslide susceptibility mapping in Central Nepal Himalaya. Arab J Geosci 7(2):725–742 https://doi.org/10.1007/s12517-012-0807-z

Regmi NR, Giardino JR, Vitek JD (2010) Modeling susceptibility to landslides using the weight of evidence approach: Western Colorado, USA. Geomorphology 115(1-2):172–187. https://doi.org/10.1016/j.geomorph.2009.10.002

Temesgen B, Mohammed U, Asfawosen A (1999) Landslide hazard on the Dabicho, Wondo genet area: the case of June 18, 1996 event. SENET: Ethiop J Sci 22(l):l27–l140

Temesgen B, Mohammed U, Korme T (2001) Natural Hazard assessment using GIS and remote sensing methods, with particular reference to the landslides in the Wondogenet area, Ethiopia. Phys Chem Earth Part C: Solar Terrestr Planet Sci (C) 26(9):665–615

Vakhshoori V, Zare M (2016) Landslide susceptibility mapping by comparing the weight of evidence, fuzzy logic, and frequency ratio methods. Geomatics Nat Hazards Risk. 7(5):1731–1752

Van Westen CJ (2002) Use of weights of evidence modeling for landslide susceptibility mapping. International Institute for Geoinformation Science and Earth Observation (ITC), Enscheda

Varnes DJ (1978) Slope movements: types and processes. In: Schuster, R.L. and Krizek, R.J. (eds.) Landslide analysis and control, National Academy of Sciences, Transportation Research Board Special Report 176, Washington, 11-33.

Wang Q, Li W (2017) A GIS-based comparative evaluation of analytical hierarchy process and frequency ratio models for landslide susceptibility mapping. Phys Geogr 38(4):318–337

Woldearegay K (2008) Characteristics of a large-scale landslide triggered by heavy rainfall in Tarmaber area, central highlands of Ethiopia. Geophys Res Abstr 10:EGU2008-A-04506 2008, SRef – ID:16077962/gra/EGU2008-A-04506

Woldearegay K (2013) Review of the occurrences and influencing factors of landslides in the highlands of Ethiopia. Momona Ethiop J Sci (MEJS) 5(1):3–31

WP/WLI - International Geotechnical Societies’ UNESCO Working Party on World Landslide Inventory (1990) A suggested method for reporting a landslide. Int Assoc Eng Geol Bull 41:5–12

Yalcin A, Bulut F (2007) Landslide susceptibility mapping using GIS and digital photogrammetric techniques: a case study from Ardesen (NE-Turkey). Nat Hazards 41:201–226

Yalcin A, Reis S, Aydinoglu AC, Yomralioglu T (2011) A GIS-based comparative study of frequency ratio, analytical hierarchy process, bivariate statistics and logistics regression methods for landslide susceptibility mapping in Trabzon, NE Turkey. CATENA 85(3):274–287 https://doi.org/10.1016/j.catena.2011.01.014

Ye C, Li Y, Cui P, Liang L, Pirasteh S, Marcato J et al (2019) Landslide Detection of Hyperspectral Remote Sensing Data Based on Deep Learning With Constrains. IEEE J Selected Topics Appl Earth Observ Remote Sensing 12(12):5047–5060. https://doi.org/10.1109/jstars.2019.2951725

Yilmaz I (2009) Landslide susceptibility mapping using frequency ratio, logistic regression, artificial neural networks and their comparison: A case study from Kat landslides (Tokat—Turkey). Comput Geosci 35(6):1125–1138

Download references

Acknowledgments

The first author would like to thank Dr. Kifle Woldearegay, Mr. Azmeraw Wubalem and Mr. Leulalem Shano for thier useful Comments and suggestions to improve this research work. He would also like to thank Mizan Tepi University for providing the scholarship opportunity to pursue his MSc study at Addis Ababa Science and Technology University.

Mizan Tepi University partially funded the first author’s MSc thesis research work.

Author information

Authors and affiliations.

Department of Geology, College of Applied Sciences, Addis Ababa Science and Technology University, Addis Ababa, Ethiopia

Tilahun Mersha & Matebie Meten

Department of Geology, College of Natural and Computational Science, Mizan Tepi University, Tepi, Ethiopia

Tilahun Mersha

You can also search for this author in PubMed   Google Scholar

Contributions

TM as a first author has mostly participated in the whole process of this research work including the field-work, data collection, database preparation and in compiling the results by himself. MM, as an advisor, commented in each phase of the research for further improvement. TM addressed the comments given from his advisor in terms of its scientific justification, methodological aspect and English correction before the final submission and finally the two authors approved this paper submission.

Corresponding author

Correspondence to Tilahun Mersha .

Ethics declarations

Competing interests.

The authors declare that we do not have any competing interests with any individual or organization.

Additional information

Publisher’s note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Mersha, T., Meten, M. GIS-based landslide susceptibility mapping and assessment using bivariate statistical methods in Simada area, northwestern Ethiopia. Geoenviron Disasters 7 , 20 (2020). https://doi.org/10.1186/s40677-020-00155-x

Download citation

Received : 09 October 2019

Accepted : 11 May 2020

Published : 15 June 2020

DOI : https://doi.org/10.1186/s40677-020-00155-x

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Frequency ratio
• Weights of evidence
• Landslide susceptibility

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• My Account Login
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Open access
• Published: 22 April 2015

Landslide susceptibility mapping using GIS-based statistical models and Remote sensing data in tropical environment

• Himan Shahabi 1 &
• Mazlan Hashim 1  

Scientific Reports volume  5 , Article number:  9899 ( 2015 ) Cite this article

54k Accesses

277 Citations

5 Altmetric

Metrics details

• Geomorphology
• Solid Earth sciences

This research presents the results of the GIS-based statistical models for generation of landslide susceptibility mapping using geographic information system (GIS) and remote-sensing data for Cameron Highlands area in Malaysia. Ten factors including slope, aspect, soil, lithology, NDVI, land cover, distance to drainage, precipitation, distance to fault and distance to road were extracted from SAR data, SPOT 5 and WorldView-1 images. The relationships between the detected landslide locations and these ten related factors were identified by using GIS-based statistical models including analytical hierarchy process (AHP), weighted linear combination (WLC) and spatial multi-criteria evaluation (SMCE) models. The landslide inventory map which has a total of 92 landslide locations was created based on numerous resources such as digital aerial photographs, AIRSAR data, WorldView-1 images and field surveys. Then, 80% of the landslide inventory was used for training the statistical models and the remaining 20% was used for validation purpose. The validation results using the Relative landslide density index (R-index) and Receiver operating characteristic (ROC) demonstrated that the SMCE model (accuracy is 96%) is better in prediction than AHP (accuracy is 91%) and WLC (accuracy is 89%) models. These landslide susceptibility maps would be useful for hazard mitigation purpose and regional planning.

Similar content being viewed by others

Landslide susceptibility zonation using the analytical hierarchy process (AHP) in the Great Xi’an Region, China

The importance of input data on landslide susceptibility mapping

Evaluating scale effects of topographic variables in landslide susceptibility models using GIS-based machine learning techniques

Introduction.

Destructive outcomes of landslides in relation to human life and the overall economic system of many nations around the globe are very severe 1 . Landslide hazard assessment and risk reduction can be accomplished by providing risk managers with easily accessible, continuous and accurate information about landslide occurrence 2 . Thus, an accurate susceptibility mapping can be key information for a large variety of users from both private and public sectors, from governmental departments and the scientific community on both local and international levels 3 .

Recently, landslide susceptibility mapping has been made possible due to the accessibility and variety of remote sensing data and thematic layers as causative factors data using GIS 4 , 5 , 6 . Most of these landslides are referred to as significant geomorphic processes which usually form an important landscaping aspect in humid tropical mountain surroundings 7 . Records have shown that in Southeast Asia, steep hill slopes, seasonally dry periods, excessive rainfall intensities and unstable soils are the main causes of frequent landslides 8 . In Cameron highlands, landslide mapping is difficult since the landslides are generally covered by dense vegetation. In addition the cloudy and rainy weather conditions are often undesirable for optical remote sensing data 9 . Consequently, it is necessary that new techniques and accurate data are used in landslide susceptibility mapping in the tropical environment.

In recent years, radar has added a new dimension to disaster management research by providing real-time and precise information. Synthetic aperture radar (SAR) is an active remote sensing system which has the capability of data collection day and night under all weather conditions 10 . In recent years, SAR data are increasingly applied to natural hazards’ researches, either by themselves or in combination with data from other remote sensing sensors 11 . SAR data such as amplitude and SAR polarimetric (POLSAR) images, interferometric DEM and other products of the interferometric DEM have proven useful. The combination of both optical and SAR data can also be utilized in geo-hazards identification and mapping to complement analysis. The SAR technique is highly popular in landslide studies 12 .

Thus, remote sensing can play a role in the production of a landslide inventory map and in the generation of thematic maps related to landslide occurrences. Several previous works have showed the potential of remote-sensing data both in the extraction of causal factors which are linked to landslide occurrences and finding of landslide areas 13 , 14 . Also, reports associated with landslide studies utilizing GIS and probabilistic models have been published 15 . The majority of the previously mentioned research has already been performed while using landslide inventory map extracted from optical satellite images and aerial photographs. Basic qualitative procedures simply use landslide index to identify areas along with related geological as well as geomorphologic features which can be susceptible to landslides 16 .

The usage of the analytical hierarchy process (AHP) approach, produced by Saaty 17 , has been utilized by many researchers world-wide 18 , 19 , 20 . Other different methods have been proposed by several investigators, including weighted linear combination (WLC) 21 , 22 . In addition, spatial multi-criteria evaluation (SMCE) method has also been applied to landslide susceptibility mapping 23 , 24 . Lately, new approaches have been employed for landslide susceptibility mapping including neuro-fuzzy 25 , 26 , 27 , decision-tree (DT) methods 28 , 29 , 30 and support vector machine (SVM) are tried and their efficiencies are evaluated 31 , 32 , 33 .

Each of the applied approaches depends on different logical explanations with the purpose of generating a landslide susceptibility mapping in an objective manner, thus decreasing the subjective evaluation of the experts. Thus, the primary distinction between earlier studies and the present research is the integration of landslide susceptibility maps extracted by AHP, WLC and SMCE methods for prepared landslide susceptibility mapping in Cameron Highlands area, Malaysia using GIS and remote sensing data. Also, the primary objective of this study is to assess the potential of AIRSAR data together with C-, L- and P-band images and WorldView-1 images for producing landslide inventory maps in the tropical forest namely in the Cameron Highlands, Malaysia.

Study area characteristics

With an area of about 660 km 2 , the Cameron Highlands is situated on an undulating plateau within the central section of the Peninsular Malaysia. The research site which is part of Cameron Highlands is bounded by Longitudes 101° 20′ 21″ E to 101° 26′ 50″ E and Latitudes 4° 24′ 37″ N to 4° 33′ 19″ N (Geographic Lat/Lon WGS 84 Projection) 34 . The research area encompasses an area of 38.4 km 2 and is situated near the northern central part of Peninsular Malaysia in Pahang state which is one of the 13 states of Malaysia ( Figure 1 ).

Location of the study area in the Cameron Highlands, Peninsular Malaysia; (a) Landsat ETM + mosaic image of Peninsular Malaysia extracted from ENVI 4.8 software created by first author (H.SH), (b) The rectangular area represents the actual study area in the Cameron Highlands created by first author (H.SH).

The geomorphology of the area is rough and has altitude ranging from 840 m to 2110 m. It has an estimated 15% of flat terrain situated adjacent and elongated to the main river. The hilly areas dominate the western and north western parts, of which Mt. Irau is the highest peak with 2110 m. Bertam and Telom Rivers are the main drainage features in this area. Its valley and tributaries are mainly flowing from north-northwest to south-southeast 35 . Megacrystic biotite granites are the most common geological structure of central mountain chain in Peninsular Malaysia 36 . Schists, phyllite, slate, as well as limestone are also significant lithologies of Cameron Highlands 37 . The highlands are usually cloud-covered during the year. The tropical forest and tea plantations, temperate vegetable and flower farms region are the major vegetable cover in the study area 38 .

The average annual rainfall of between 2,500 and 3,000 mm per year in the Cameron Highlands was calculated between March and May and from November to December. The volume of rain is another agent which impacts the fill slopes, leading to stream and gully erosion 38 . Average daytime and nighttime temperatures in the study area of 24°C and 14°C respectively, locates the Cameron Highlands in moderate climatology category. Physically, 66% of the slope gradients are more than 20°. About 8% (5,500 ha) are categorized under agriculture, 86% (60,000 ha) of the area is forested, 4% (2,750 ha) is occupied by housing and the residuum are used for recreation and other activities. The impact of present agricultural activities has led to severe soil erosion, culminating into large quantities of sediments. In addition, occurrences of landslides are another hazardous natural disaster that threatens this location which has high economic potential 39 .

In the Cameron Highlands, Malaysia, like in most tropical mountainous areas, natural hazards such as flash floods, mass movement and landslides are of great social concern and cause considerable damage to life and property. Earthflows, mudflows and landslides present a major danger in Cameron Highlands because of the nature of the topography, the climate and human activities. Generally, precipitation is a triggering factor that can cause mudflows and landslides 38 . Landslides in Cameron Highlands have also destroyed the roads and troubled economic activity such that total economic loss due to landslides was approximately US $1 billion between 1973 and 2007 39 .

Landslide inventory

Landslide inventory maps show locations and also features of landslides that have moved in the past although usually do not show the mechanism(s) that triggered them. Therefore, inventory maps provide useful information about the spatial distribution of locations of existing landslides and the potential for future landslides 40 . Landslide mapping is difficult in tropical mountainous environments as dense vegetation growth obscures landslides soon after they occur 9 .

Sometimes carrying out mitigation measures is held back due to insufficient and dependable landslide inventory map which hampers the evaluation of landslide hazard and risk. Utilizing remote-sensing data such as radar, optical satellite images and aerial photography interpretation are primary methods to obtain important, cost-effective information of landslide location 14 . The landslide information taken from remotely sensed images is especially associated with morphology, plant life and hydrologic conditions of the region 6 .

Several types of remote sensing data may be used in detecting landslide features, for instance, stereo-remote sensing products which in turn reveal the actual morphodynamical features of landslides 41 . In this study, published reports, field surveys, interpretation of digital aerial photographs (DAP) (10,000–1:50,000 scale) over a 25-year period, WorldView1 satellite imagery on the March 2011, AIRSAR data on November 2004 has been utilized for extraction of landslide inventory map.

These black and white digital images with resolution = 0.54 m pixel were taken during 1981–2006 and were acquired from the Malaysian Surveying and Mapping Department archives. WorldView-1 satellite data, which has a resolution of 0.46 meter for panchromatic band was used for detection of occurred landslides and validation of landslide inventory map obtained from AIRSAR data in the study area. The AIRSAR data were collected over the study area in November 2004, during the PacRim1 campaign. This dataset is to be compared with landslide features generated from aerial photographs and WorldView-1 satellite imagery which are also generated in UTM reference system. Multiple field investigations and ground control points (GCPs) were carried out by using global positioning system (GPS) for collection of mapping information on landslide locations (See Figure 2 ) and generating stereo models from digital aerial photography data.

Field photographs of recently occurred landslides and types of landslides that were taken from field surveys by first author (H. SH); (a) a shallow translational rockslide, ( b ) a shallow translational debris slide at the road side, ( c ) and ( d ) deep-seated rockslides. Arrow depicts the movement direction.

To identify landslides in the study area, three techniques were employed. The first technique was to compare directly by overlaying landslide vector images onto the DEMs and AIRSAR raster images. The second technique was to classify the images using ENVI 4.8 software. This is to separate landslides from the other land cover types in the surrounding area. The last technique is to separate landslides from the other land cover types in the surrounding area using segmentation followed by classification. This technique was carried out using software called “eCognition”. In eCognition software pixels are segmented into image object, so the image classification process in this software is image objects based rather than pixels based classification that was previously carried out in ENVI 4.8 software 42 .

The data used in segmentation were 1) C-, L- and P-band (wavelengths) with seven polarizations (Chh, Lhh, Lvv, Lhv, Phh, Pvv and Phv) images and 2) Slope image. In segmentation the information about the group of pixels inside the boundaries of landslide was used. The information include: 1) Spectral values of the C-, L- and P-band and 2) Average slope of the area. These spectral values may represent spectral signatures of landslides. In segmentation processes using eCognition software, all three bands with seven polarizations and a slope image were combined to identify landslides in the images. If one existing landslide can be identified it can be used in the segmented image as a sample polygon. Using classification techniques, all other polygons that have the same characteristic of pixels brightness and average slope will be highlighted as similar 42 .

Furthermore, the efficiency and quality of the SAR data and optical satellite images should be examined using a proper method. The root mean square error (RMSE) method was used for accuracy of the obtained result 43 , 44 . The each ground control points, the efficiency of the SAR data and optical satellite images is calculated based on the formula 43 :

where u is residual in the x axis; v is residual in the y axis.

Total RMSE is then derived as ( Eq. 2 ):

where n is number of GCPs; u is residual in the x axis; v is residual in the y axis.

The total RMSE was calculated for each area, SAR and optical satellite images based on nearest neighbour resampling method 45 .

Spatial database construction

Data collection is the main step in landslide susceptibility mapping whereby the relevant landslide conditioning factors are extracted to construct a spatial database. These processes are subsequently evaluated by using the relationship between the landslide and landslide causative factors and then verification of the results 25 . There are no universal guidelines regarding the selection of factors in landslide susceptibility mapping. One parameter may be an important controlling factor for landslide occurrence in a certain area but not in another one. The selection of causal factors therefore needs to take the nature of the study area and data availability into account. Collectively ten parameters of slope, aspect, soil, lithology, NDVI, land cover, distance to road, distance to drainage, precipitation and distance to fault were used to construct a spatial database using GIS, SAR data and optical satellite images processing. The database consists of vector-type spatial datasets derived from Arc GIS 9.3 software package.

In the first step a digital elevation model (DEM) of the research area was produced from the triangulated irregular network (TIN) model using AIRSAR DEM with 10-m pixel size. The slope and slope aspect parameters were obtained from the generated AIRSAR DEM with 10-m pixel size. In addition, the distance from drainage was calculated using the AIRSAR DEM. The fault lines and lithology were derived from geological map with 1:63,300-scale from Mineral and Geosciences Department of Malaysia. Also, the lineaments were derived from the structural maps and the aerial photos. The distance from the road was calculated using the 1:25,000-scale topography map. A 50-m buffer zone is chosen as a distance from road in the study area that is determined based on the occurred landslides to the closeness of the road. The soil types were acquired from a 1:25,000-scale soil map. In this research, land cover was extracted from SPOT 5 satellite image on 21 March 2010 that was calibrated by using ground control points (GCPs) obtained during field works. The supervision classification method by using ENVI 4.8 software was used to develop a statistical characterization of the reflectance for each information class of land cover map.

Furthermore, the SPOT 5 data were first classified into eight main land cover types using a supervised maximum-likelihood classification (MLC), namely grass, primary forest, rubber, cutting, secondary forest, settlements, agricultural area and water body. Field surveying was also used to justify the land cover map’s accuracy according to the order of SPOT 5 spatial resolution (~ 10 m).

Despite image pre-processing (geo-referencing and ortho-rectification, geo-rectification) on the SPOT 5 images, a 21 ground control points (GCPs) obtained during field visits were used for additional improvement in the accuracy of satellite images. Also, the SPOT 5 satellite image was used for extraction of normalized difference vegetation index (NDVI) map. The NDVI value was calculated using the formula NDVI = (IR − R)/(IR + R), where IR and Red are the near infrared (NIR) and red bands, which are from 0.7 to 1 lm and 0.6 to 0.7 lm of the electromagnetic spectrum. The NDVI value, which indicates the presence and intensity of vegetation in the study area, was classified into ten classes.

The precipitation data was prepared using the last 30 years (1981–2011) of historical rainfall data. In our methodology we used long-term precipitation for a 30 year period. An average annual rainfall contour map is mapped out from the daily rainfall data measurements. Also, the Kriging method using Arc GIS 9.3 was used for spatial interpolation on the contour maps. The flowchart for the landslide susceptibility mapping and spatial data are shown in Figure 3 . All the landslide related factors were converted to a raster grid (10 × 10-m cells) that included 1,725 rows by 5,621 columns for application of the three different GIS-based statistical approaches including analytical hierarchy process (AHP), weighted linear combination (WLC) and spatial multi-criteria evaluation (SMCE).

Different steps of preparing the map for landslide susceptibility mapping.

Landslide susceptibility mapping models

Landslide susceptibility analysis was implemented using the analytical hierarchy process (AHP), weighted linear combination (WLC) and spatial multi-criteria evaluation (SMCE) methods in a part of the Cameron Highlands, Malaysia using GIS- based statistical models and remote sensing data. Also, the maps credibility was validated using R-Index and ROC methods.

Analytical hierarchy process (AHP) in landslide susceptibility analyses

The AHP developed by Saaty 17 is a flexible tool of analyzing complicated problems focusing on site selection, urban planning and landslide susceptibility analysis 21 . AHP which is a multi-criteria decision-making and multi-objective approach allows the active participation of decision-makers namely the managers in reaching an agreement rationally 46 .

These factors are arranged in a hierarchic order and numerical values to subjective judgments based on the relative importance of each factor. Subsequently these factors are synthesized and each factor is assigned according to their importance 47 . Apart from that, reciprocal pair-wise comparison matrix is established to utilize AHP. Each layer based on a 9-point rating scale are the entries into the matrix as developed by Saaty 48 (See Table 1 ).

Generally, the specification of the values of the factors relative to each other is affiliated to the selection of the decision-maker. Nonetheless, in this research, both the determination of the decision options and comparison of the parameters were based on the comparison of landslide inventory maps 18 . The weight of each factor from the matrix weighting factor was multiplied by its weight class. The result of the susceptibility map is ascertained by factors with high local representation. These representations can be based on different parameters including natural (lithology, distance to faults, etc.), man-made (roads and other engineering structures), causal (slope, aspect, lithology, etc.) and triggering (precipitation, seismicity, etc.) 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 , 40 , 41 , 42 , 43 , 44 , 45 , 46 , 47 , 48 , 49 . The selection of the ten causal factors in this study is based on these mentioned four criteria and also data accessibility. In AHP approach, consistency ratio (CR) Eq. (3) , is utilized to show the probability that the judgments matrix was randomly created 48 .

where RI is the average of the resulting consistency index depending on the order of the matrix given by Saaty 48 and CI is the consistency index and can be expressed as Eq. (4)

where λ max is the largest or principal eigen value of the matrix and can be easily calculated from the matrix and n is the order of the matrix.

If the CR values were greater than 0.1, the AHP model was automatically rejected. The acquisitive weights were employed by using a weighted linear sum procedure. Furthermore, the acquisitive weights were employed to calculate the landslide susceptibility models 50 .

Weighted linear combination (WLC) in landslide susceptibility analysis

Weighted linear combination (WLC) is a hybrid between qualitative and quantitative methods 21 . WLC is based on the qualitative map combination approach (heuristic analysis). This technique is a popular method that is customized in many GIS and is applicable for the flexible combination of maps. Thus the tables of scores and the map weights can be adjusted based on the expert’s judgement in the domain. First, this method requires the standardization of the classes in each factor to a common numeric range. Each factor rating was based on the relative importance of each class according to field observations and existing literature, indicating the conditions as highly susceptible to slope failure 51 .

Primary-level weights and secondary-level weights are two types of parameters weights used 15 . The primary-level weights are rule-based whereby the ratings given to each class of a parameter is based on certain criteria. In this research, this criterion can be described as landslide density. In WLC method, the landslide density is a ratio between the area of landslide pixels situated inside a category of a specific factor divided by the total area of that category. The obtained result of this criterion is converted into percentage. The secondary-level (factor weights) determine the degree of exchange of one parameter versus another parameter based on opinion-based scores 21 . Both the parameters weights are combined to estimate landslide susceptibility and classify areas in relative susceptibility categories 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 39 , 40 , 41 , 42 , 43 , 44 , 45 , 46 , 47 , 48 , 49 , 50 , 51 , 52 .

Susceptibility S ( i, j ) in each pixel ( i, j ) can be expressed as the combination of the product of primary and secondary level weights Eqs. (5 and 6 ) 52 :

The weights of ‘proportional importance’ to each attribute map layer are directly affected by the decision-maker. A total score is then obtained for each alternative. This is done by multiplying the weight allocated to each attribute by the scaled value and summing the outputs of all attributes. In this method, highest overall score can be selected from the overall scores calculated for all of the alternatives 53 . The final steps for creating the landslide susceptibility map using WLC method is the combination of all weighted layers into individual maps. Then, landslide susceptibility zones were generated based on classification of the scores of these maps 21 . The WLC method can be performed using any GIS system that has overlay techniques.

Spatial multi-criteria evaluation (SMCE) in landslide susceptibility analysis

Spatial multi-criteria evaluation (SMCE) application helps and allows users to perform multi-criteria assessment in a spatial approach. In SMCE, the alternatives are locations in the form of points, lines, areas and grid cells. Therefore, criteria could occur in the form of maps 54 . Thus, SMCE is an applied science-based method that combines spatial analysis using GIS and multi-criteria evaluation (MCE) to transform spatial and non-spatial input which generates output decision 55 . Spatial multi criteria evaluation is considered as a procedure that includes the input layers, which are the spatial representation of the criteria. They are a ‘criteria tree’ that can be standardised, grouped and weighted. In addition, the input layers (resultant decision), need to be standardised from their original values to the value range of 0–1. Furthermore, the indicators have cartographic representations (natural and administrative polygons and pixel based raster maps) and different measurement scales (nominal, ordinal, interval and ratio) 23 .

The output of SMCE is one or more ‘composite index map(s), which indicates the extent to which criteria are met or not in different areas and thereby supports decision making 56 . The multi-criteria evaluation of AHP method has been used as the theoretical background of SMCE method. There are several phases in conducting the SMCE, such as problem tree analysis, standardization, weighting and map generation. Once all the criteria and related maps or attribute tables are entered in the criteria tree, the criteria have to be standardized. The values in the various input maps have different meanings and are probably showed in different units of measurement such as percentages, meters, distance in meters, land cover classes, etc. 24 . In order to standardize input maps in SMCE environment, one of the standardization methods such as numerical, Boolean and qualitative methods can be used.

The first step for standardizing map values is to convert the actual map values to a range between 0 and 1 by using a set of equations. The next step is the determination of each indicator for intermediate or overall objectives. Finally, the landslide conditioning factors are weighted by means of direct, pairwise and rank ordering comparison and the output is a composite index map 23 .

Results and Discussion

Landslide location detection using airsar data and optical satellite images.

The size and locations of landslides on the old landslide map in the study area are inaccurate, therefore not suitable for any applications in the landslide inventory map. Also, identified landslide is too small. Therefore the information about the pixel values might be insufficient for identifying landslide by segmentation and classification technique. Landslides can be identified on AIRSAR intensity image. The spectral values of the pixels representing landslides can be differentiated from the spectral values of the surrounding.

Because of many problems with the old landslide map, validation of the map was the only way to check the accuracy. The landslides that were identified on the ground as mentioned previously were also identified on aerial photograph. To validate the landslide inventory map, the landslide features that were obtained from WorldView-1 satellite images were overlaid onto the C-, L- and P-band image. Since the landslide feature was in UTM reference system, C-, L- and P-band image in the UTM reference system was used. The final landslide inventory map that was compiled in the present study is shown in Figure 4 .

Final landslide inventory map extracted from AIRSAR data and WorldView-1 satellite images; (a) types of landslide locations overlaid on AIRSAR DEM, ( b ) enlarged views of two landslides locations on WorldView-1 satellite images (panchromatic band).

Subsequently, the final landslide vector map was transformed into a grid database with a cell size of 10 × 10 m in ArcGIS 9.3. The compilation of more than 25 years of landslide inventory resulted in a total of 92 landslides identified, covering an area of 6.27 km 2 and accounting for 4.05% of the study area. The minimum, mean and the maximum landslide areas are 0.003, 0.017 and 0.123 km 2 respectively. Among the cases of landslide occurrences, 74 cases (80%) were selected during the training of landslide susceptibility mapping models and the remaining 18 cases (20%) were used for validation purpose.

The results also identified that most of the landslides detected from AIRSAR data, digital aerial photographs and WorldView-1 satellite images are shallow rotational and there are a few translational and flow types. In the present study, only the rotational landslides are counted, because the occurrences of the other types of landslides were scarce and very small. To assess the correctness and quality of the result, root mean square errors (RMSE) ( Eqs. 1 and 2 ) was performed by comparing the in-situ measurements with the resulted one. The resulted RMSE was 0.163 which is acceptable and shows the efficiency of the proposed method in recognizing the landslides with high precision.

Landslide susceptibility mappings using GIS-based statistical models

The ten factors were transmuted into a vector-type spatial database using the Arc GIS 9.3. Also, the landslide triggering factors were extracted from the database. Furthermore, in this study, the landslide susceptibility analyses were performed using the GIS-based statistical models including AHP, WLC and SMCE.

Landslide susceptibility mapping using AHP model

The rating values for each class of each triggering factor which had an influence on landslide susceptibility were also calculated by Expert choice software. The software calculates the weight of each factor and based on this weighting factors were prioritized. To get factor weights in AHP, the same method used in calculating the rating value was applied as shown in Table 2 . The software also calculates the CR. The ratio implies an acceptable level of consistency in the pair-wise comparison which is sufficient to recognize the factor weights in the landslide susceptibility model. In this research the CR value was less than 0.1 for all cases of the acquired class weights. This means a reasonable level of consistency in the pair-wise comparison.

Precipitation with AHP weight (0.142), Lithology with AHP weight (0.124) and slope with AHP weight (0.109) were found to be important parameters in occurrences of landslides in the study area, whereas distance to fault (0.074) and distance to road (0.078) received a low degree of importance. The results of the class weights, factor weights and consistency ratio of the data layers are shown in Table 2 .

As a result of the AHP analysis, we used the following equation;

where R i is the rating class of each layer and W i is the weights for each of the landslide conditioning factors.

These LSM values were then divided into four susceptibility zones including low (LS), moderate (MS), high (HS), very high (VHS) based on the natural breaks method (See Figure 5 ).

Landslide susceptibility map derived from AHP model.

Based on the result of the obtained landslide susceptibility map, 28.32% (1.77 km 2 ) of the total area show low landslide susceptibility. Moderate, high and very high susceptible zones make up 36.23% (2.27 km 2 ), 24.18% (1.70 km 2 ) and 11.27% (0.70 km 2 ) of the total area, respectively (See Figure 5 ).

Landslide susceptibility mapping using WLC model

After producing the landslide related parameter maps and the weights for their classes, this section describes the combination of the maps to estimate landslide susceptibility in the target area using the WLC approach. To be able to create a landslide susceptibility map using WLC model, the factor weights were first extracted from AHP method that are principally based on the ratings provided to each class of a factor. In order to actualise this phase, the pair-wise comparison matrix and CR of used data layers are shown in Table 3 .

In this study, the weight value of the precipitation (0.141) and slope (0.123) are the highest. On the other hand, the low WLC weights belong to distance to fault and NDVI with 0.062 and 0.073, respectively (See Table 3 ). The CR is ascertained to be 0.069 and this value expresses the appropriate amount of CR employed to acquire the comparison matrix because it is less than 0.1. Therefore the weights related to factors were multiplied by the appropriate factor maps and then all the weighted factor maps were overlaid to extract a landslide susceptibility map based on WLC model.

The allocated rates used to reclassify vector data layers and raster data layers were generated from new reclassified data. Also, the raster calculator function was used to reclassify raster layers as input parameters. By means of the raster calculation all the weighted factor maps were added to provide the final landslide susceptibility map. Based on this method, four susceptibility zones were identified such as low, moderate, high and very high (See Figure 6 ).

Landslide susceptibility map derived from WLC model.

The WLC-derived landslide susceptibility map yields very high susceptible zones of about 7.10% (0.44 km 2 ) of the total area while about 17.12% (1.07 km 2 ) is classified as being highly and moderately susceptible and 46.56% of the case study area (2.91 km 2 ) is classified as being a moderate susceptible zone. 22.34% of the study area (1.40 km 2 ) is classified as low susceptible zone. Total susceptibility value in each cell was the sum of the rasters (corresponding to primary weights) multiplied by their secondary weights, as expressed in Eq. 5 . The categorized susceptibility map was compared to the landslide inventory map to calculate the area of landslides in each class (See Figure 6 ).

Landslide susceptibility mapping using SMCE model

The first stage to generate SMCE analysis is standardization of input layers based on their original values (0–1). This standardization can be performed as a SMCE module in ILWIS (Integrated Land and Water Information System) software to acquire the composite index maps and the final landslide susceptibility map 57 .

The SMCE was built based on analyzing the weight value in bivariate statistical analysis for classes of conditioning factors (See Table 4 ). In this research, for standardization of the scale in thematic layers the fuzzy logic method was used. All comparisons are based on analytical hierarchy process (See Table 1 ). The employment of fuzzy sets to represent linguistic terms makes it possible for one to represent more effectively and continuously something which is fuzzy. In this research, in order to input parameters of SMCE in linguistic form, the following fuzzy sets were utilized:

Low = (0/1, 0.25/2, 0.75/3, 1/4)

Moderate = (0/1, 0.5/2, 1/3, 0.5/4)

High = (0/1, 1/2, 0.75/3, 0.25/4)

Very high = (0/1, 0.25/2, 0.5/3, 0.75/4)

In the foregoing fuzzy set notation, the values before the slash are the degrees of confidence and the values after slash are the members of the set 24 . The fuzzy set values of the predisposing factors of the landslides for study area are found as follows:

μ S Slope = (0.09/1, 0.73/2, 0.82/3, 0.91/4, 1/5)

μ S Aspect = (0/1, 0.87/2, 1/3, 0.75/4, 0.96/5, 0.39/6, 0.46/7, 0.65/8, 0.54/9)

μ S Soil = (1/1, 0.81/2)

μ S Lithology = (1/1, 0.87/2)

μ S NDVI = (1/1, 0.93/2, 0.72/3, 0.83/4, 0.49/5, 0.87/6, 0.42/7, 0.64/8, 0.77/9, 0.57/10)

μ S Land cover = (0.88/1, 0.56/2, 0.68/3, 1/4, 0.96/5, 0.92/6, 0.79/7, 0.62/8)

μ S Precipitation = (0.33/1, 0.81/2, 0.48/3, 0.54/4, 0.68/5, 0.76/6, 0.95/7, 0.71/8, 0.88/9, 1/10)

μ S Distance to road = (1/1, 0.95/2, 0.84/3, 0.79/4, 0.71/5)

μ S Distance to drainage = (1/1, 0.89/2, 0.94/3, 0.84/4, 0.76/5, 0.71/6, 0.63/7)

μ S Distance to fault = (1/1, 0.82/2, 0.94/3, 0.79/4, 0.71/5, 0.62/6)

In this research, grid-based analysis was performed in Arc GIS 9.3 to produce the fuzzified index maps. Criteria for landslide susceptibility mapping in this research are divided in four groups (sub-goals) such as geomorphological (slope, aspect); geological (distance to fault and lithology); environmental (soil, land cover, distance to road and NDVI); hydrological (distance to drainage and precipitation) factors. They are the input for the SMCE analysis.

The levels of the influence of sub-goals and weight value of main indicators for the study area were calculated by AHP (See Table 4 ). Based on our results in expert choice software, it can be seen that hydrological factor has the most influence on landslide occurrence (0.146). On the other hand, the geological factor which has less influence was categorized in the lowest level (0.871). Based on the results in Table 4 , in geomorphological factors, it can be seen that slope factor with weight value (0.128) is more susceptible to landslide but the aspect is less prone to landslide as it has weight value of 0.112. For geological factors, weight corresponding to lithology (0.097) is large, whereas distance to fault is lowest (0.073). In the case of environmental factors, it was observed that soil, land cover, distance to road and NDVI have a weight value of 0.123, 0.084, 0.083 and 0.087, respectively. In hydrological factors, precipitation has a higher probability of occurrence than the distance from drainage and therefore received a higher weight (0.143 vs. 0.081).

As a general result, the precipitation is highly prone to landslide occurrence and in contrast, distance to fault has the lowest impact in landslide susceptibility. For all cases of the acquired class weights (sub-goals and indicators), the consistency ratios are less than 0.1; the ratio indicates appropriate degree of consistency that was good adequate to recognize the class weights. The final landslide susceptibility map extracted from SMCE model was reclassified into four relative susceptibility zones: low, moderate, high and very high ( Figure 7 ) based on natural break classification method.

Landslide susceptibility map produced by the SMCE model.

According to the landslide susceptibility map acquired from the SMCE approach, 10.16% (0.63 km 2 ) of the entire area is found to be of low landslide susceptibility zone. Moderate and high susceptible zones showed 13.18% (0.82 km 2 ) and 29.14% (1.82 km 2 ) of the entire area, respectively. The very high landslide susceptibility zone is 47.52% (2.97 km 2 ) of the entire study area (See Figure 7 ).

Evaluation, comparison and precision of landslide susceptibility mapping methods

In this research landslide inventory map is used for methods evaluation and comparison of susceptibility mapping. In order to do that, we crossed the above mentioned map with map of susceptibility in the geographical information system and landslide index was used for evaluation of hazard classes on zoning maps ( Eq. 8 ). Landslide index is defined as follows 58 .

where Li is the index for occurrence of landslide hazard in each susceptibility zone (percent), Si is the slide area in each susceptibility zone, Ai is the area of each zone and n is the number of susceptibility classes.

In the above relation: landslide index is percentage of sliding surface ratio in any zone to the area of that zone, divided by the total sliding ratio to the total surface of the zones. In order to compare the obtained susceptibility maps, the parameter considered for precision of the predicted results is ( P ). Precision of the predicted results can be estimated by the following ( Eq. 9 ) 59 .

where Ks is the area of slide zone in upper moderate susceptibility level and S is the area of landslide in the region.

As Table 5 shows, from low up to very high level, Li amounts increase in all of the methods used. Therefore, all of the methods used for susceptibility levels have yielded acceptable results. In Table 5 , the P refers to the high efficiency of SMCE and AHP in susceptibility mapping in the studied region. From the statistical methods that were used, respectively, precision of the method ( P ) were 96% for SMCE, 91% for AHP and 89% for WLC which are all compatible with the conditions for occurrence of landslides in the region.

Validation of landslide susceptibility maps using R-index and ROC methods

The landslide susceptibility analysis was carrying out using three different approaches such as AHP, WLC and SMCE models. Furthermore, the analysis results were validated using the Relative landslide density index (R-index) and receiver operating characteristic (ROC) analysis to evaluate the correlation between the landslide susceptibility maps and landslide inventory points. Then, among the types of landslide occurrences, 74 cases (80%) had been determined throughout the training of landslide susceptibility mapping models and 18 cases (20%) had been employed for validation purpose. Validation of landslide susceptibility maps is performed with a formula defined by 60 ( Eq. (10) as follows:

where ni the number of landslides occurred in the sensitivity class i and Ni the number of pixels in the same sensitivity class i .

The results obtained from SMCE and AHP analysis methods have been more coherent with recurrent landslides occurring in sensitivity classes. In the characteristic analysis map, all susceptibility classes are correlated with the distribution of landslide occurrence. From the classes with very-high susceptibility, SMCE and AHP methods were better than WLC method. The R-index sample data sets for classes of very high hazard in AHP, WLC and SMCE maps are 19.31%, 14.75% and 58.49% respectively. From assessment of all classes, SMCE was more accurate than the other two methods. The results of validation (R-index) of AHP, WLC and SMCE models are shown in Figure 8a .

Validation of landslide susceptibility maps, (a) R-index validation of three methods for landslide susceptibility mapping, (b) Prediction accuracy assessment and success rate curve of the constructed landslide susceptibility models.

The suitability of each model was also assessed by comparing the maps with known landslide locations. The area under the receiver operating characteristic (ROC) curve, known as the AUC and ranging from 0.5 to 1.0, is widely employed to estimate the accuracy of presence or absence predictive models 20 . This curve was obtained by means of the statistical analysis software SPSS.

The success rates of the verification for the three LSM methods including AHP, WLC and SMCE are shown in Figure 8b . The results obtained from Figure 8b shows that a value of AUC for AHP was 0.8719 and the prediction accuracy was 87.19%. In the WLC and SMCE methods, the area ratios were 0.8417 and 0.9375 and the prediction accuracy for WLC and SMCE were 84.17% and 93.75%. These results obtained from R-index ( Figure 8a ) and ROC ( Figure 8b ) indicate that the SMCE model looks to be more accurate in terms of the performance of landslide susceptibility mapping and has better prediction accuracy than the other two models in the study area.

Landslide susceptibility mapping in tropical mountainous areas is usually difficult because of inadequate approachability, the vegetation and cloudy weather situations. In the present research, a landslide susceptibility evaluation was performed in a part of Cameron Highlands, Malaysia. The research involved three main phases including landslide inventory analysis, susceptibility mapping and validation. The landslide inventory map with a total of 92 landslide locations revealed the strong capability of AIRSAR data and WorldView-1 images satellite images to distinguish and detect the very small landslides of the earth surface which occurred due to the heavy precipitation with resulted accuracy of 0.163 RMSE. This RMCE accuracy is acceptable and shows the efficiency of the proposed method in recognizing the landslides with high precision.

Landslide types are very diverse and as a consequence landslide susceptibility assessment can only be achieved by specific landslide models. Hence, the determination of dominant landslide types and selection of appropriate suitable landslide models for landslide susceptibility mapping is difficult and will strongly influence the final landslide susceptibility map. In this study, the potential of GIS-based statistical models such as analytical hierarchy process (AHP), weighted linear combination (WLC) and spatial multi-criteria evaluation (SMCE) methods using GIS tools and remote sensing data have been evaluated in the study area. Ten factors as slope, aspect, soil, lithology, NDVI, land cover, precipitation, distance to fault, distance to drainage and distance to road were obtained from the spatial database. The selected causative factors for the study area were considered carefully based on the relevance, availability and scale attributes. However, some statistical analysis approaches require that the causative factors possess certain properties as conditional independence. Hence, the number of causative factors that can be used as input for landslide susceptibility modeling may be reduced depending upon the conditional circumstances. Therefore, accurate landslide susceptibility mapping on a regional scale strongly depends on the selection of landslide causative factors, which can be quite different according to various researchers.

In the AHP model, the precipitation (weight = 0.142) and lithology (weight = 0.124) parameters are positively associated with the occurrence of landslides. In the WLC model, the weight value of the precipitation (0.141) and slope (0.123) are the highest. On the other hand, the low WLC weights belong to distance to fault and NDVI with 0.062 and 0.073, respectively. Based on obtained results from SMCE model, it can be seen that hydrological group has the most influence on landslide occurrence (0.146) in this group and that precipitation has a high weight (0.143). On the other hand, the geological group which has less influence was categorized in the lowest level (0.871) and in this group, high weight belonged to lithology (0.097). As a general result extracted from SMCE model, precipitation had the highest impact on landslide occurrence and in contrast, distance to fault had the lowest impact in landslide susceptibility.

Thus, obtained results from AHP, WLC and SMCE susceptibility maps indicated that precipitation was the most important factor in occurrence of landslides in the study area. To confirm the utility of the results, three susceptibility maps were compared with 18 (20%) of 92 landslide zones using R-index and ROC methods. According to the R-index method, when field conditions and attributes are properly described by professional expertise, the SMCE method gave better results and was more accurate than the other two methods in this study. The obtained results identified high percentage for high and very high susceptibility classes in direct association with active landslide zones in the AHP (19.31%), WLC (14.75%) and SMCE (58.49%) maps. Also, the validation results by ROC method show that the area under the curve for AHP, WLC and SMCE models are 0.8719 (87.19%), 0.8417 (84.17%) and 0.9375 (93.75%). From the statistical methods that were used, precision of the method (P) was 96% for SMCE, 91% for FR and 89% for WLC. Thus, the map extracted from the SMCE model is more accurate compared to maps extracted from the other two models.

Generally, the result of validation implies that 86% of the total landslide pixels were properly categorised by the three landslide susceptibility mapping models, which indicates a significant rational carrying out with regard to comparable studies accomplished by other researchers in tropical environments using statistical models. Also, the validation results demonstrated acceptable agreement relating to susceptibility maps and the current data from landslide areas. Furthermore, current research revealed that the C-, L- and P-band images of AIRSAR data are able to provide acceptable coherence over highly vegetated areas. Also, the integration of AIRSAR data with high resolution satellite images can play important role in the production of a landslide inventory map in tropical regions. These landslide susceptibility maps can be used for optimum management by decision makers and land use planners and engineers to decrease losses caused by current and also future landslides through suitable prophylactic assessments and minimization procedures.

Nefeslioglu, H., Gokceoglu, C. & Sonmez, H. An assessment on the use of logistic regression and artificial neural networks with different sampling strategies for the preparation of landslide susceptibility maps. Eng. Geol. 97, 171–191 (2008).

Article   Google Scholar  

Dai, F., Lee, C. & Ngai, Y. Y. Landslide risk assessment and management: An overview. Eng. Geol. 64, 65–87 (2002).

Fell, R. et al. Guidelines for landslide susceptibility, hazard and risk zoning for land-use planning. Eng. Geol. 102, 99–111 (2008).

Scaioni, M. Remote sensing for landslide investigations: From research into practice. Remote. Sens. 5, 5488–5492 (2013).

Article   ADS   Google Scholar  

Gupta, R. & Joshi, B. Landslide hazard zoning using the gis approach—a case study from the ramganga catchment, himalayas. Eng. Geol. 28, 119–131 (1990).

Qiao, G. et al. Landslide investigation with remote sensing and sensor network: From susceptibility mapping and scaled-down simulation towards in situ sensor network design. Remote. Sens. 5, 4319–4346 (2013).

Thomas, M. F. Landscape sensitivity in time and space—an introduction. Catena. 42, 83–98 (2001).

Douglas, I. Hydrological investigations of forest disturbance and land cover impacts in south–east asia: A review. Philos. Trans. R. Soc. Lond.. Series B: Biological Sciences. 354, 1725–1738 (1999).

Article   CAS   Google Scholar  

Razak, K. A., Santangelo, M., Van Westen, C. J., Straatsma, M. W. & de Jong, S. M. Generating an optimal dtm from airborne laser scanning data for landslide mapping in a tropical forest environment. Geomorphology. 190, 112–125 (2013).

Pohl, C. Mapping palm oil expansion using SAR to study the impact on the CO2 cycle. I OP Conference Series: Earth and Environmental Science . 20, 12–21 (2014).

Joyce, K. E., Belliss, S. E., Samsonov, S. V., McNeill, S. J. & Glassey, P. J. A review of the status of satellite remote sensing and image processing techniques for mapping natural hazards and disasters. Prog. Phys. Geogr. 33, 1–25 (2009).

Corsini, A. et al. Space-borne and ground-based sar interferometry as tools for landslide hazard management in civil protection. Int. J. Remote. Sens. 27, 2351–2369 (2006).

Liu, J.-K., Shih, P. T. Topographic correction of wind-driven rainfall for landslide analysis in Central Taiwan with validation from aerial and satellite optical images. Remote Sens. 5, 2571–2589 (2013).

Oh, H. J., Park, N. W., Lee, S. S. & Lee, S. Extraction of landslide-related factors from ASTER imagery and its application to landslide susceptibility mapping. Int. J. Remote. Sens. 33, 3211–3231 (2012).

Akgün, A. & Bulut, F. GIS-based landslide susceptibility for Arsin-Yomra (Trabzon, north Turkey) region. Environ. Geol. 51, 1377–1387 (2007).

Kouli, M., Loupasakis, C., Soupios, P. & Vallianatos, F. Landslide hazard zonation in high risk areas of rethymno prefecture, crete island, greece. Nat. Hazards. 52, 599–621 (2010).

Saaty, T. L. The analytic hierarchy process: planning, priority setting, resources allocation (McGraw, New York., 1980).

Yalcin, A. GIS-based landslide susceptibility mapping using analytical hierarchy process and bivariate statistics in ardesen (Turkey): Comparisons of results and confirmations. Catena. 72, 1–12 (2008).

Kayastha, P., Dhital, M. & De Smedt, F. Application of the analytical hierarchy process (AHP) for landslide susceptibility mapping: A case study from the Tinau watershed, west Nepal. Comput. Geosci. 52, 398–408 (2013).

Shahabi, H., Khezri, S., Ahmad, B. B. & Hashim, M. Landslide susceptibility mapping at central Zab basin, Iran: A comparison between analytical hierarchy process, frequency ratio and logistic regression models. Catena. 115, 55–70 (2014).

Ayalew, L., Yamagishi, H. & Ugawa, N. Landslide susceptibility mapping using GIS-based weighted linear combination, the case in tsugawa area of Agano river, Niigata prefecture, Japan. Landslides. 1, 73–81 (2004).

Kouli, M., Loupasakis, C., Soupios, P., Rozos, D. & Vallianatos, F. Landslide susceptibility mapping by comparing the WLC and WOFE multi-criteria methods in the west crete Island, Greece. Environ. Earth. Sci. 72, 1–23 (2014).

Abella, E. C. & Van Westen, C. Generation of a landslide risk index map for cuba using spatial multi-criteria evaluation. Landslides. 4, 311–325 (2007).

Pourghasemi, H., Moradi, H., Aghda, S. F., Gokceoglu, C. & Pradhan, B. GIS-based landslide susceptibility mapping with probabilistic likelihood ratio and spatial multi-criteria evaluation models (North of Tehran, Iran). Arab. J. Geosci. 7, 1857–1878 (2014).

Oh, H. J. & Pradhan, B. Application of a neuro-fuzzy model to landslide-susceptibility mapping for shallow landslides in a tropical hilly area. Comput. Geosci. 37, 1264–1276 (2011).

Sezer, E. A., Pradhan, B. & Gokceoglu, C. Manifestation of an adaptive neuro-fuzzy model on landslide susceptibility mapping: Klang valley, Malaysia. Expert. Syst. Appl. 38, 8208–8219 (2011).

Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I. & Dick, O. B. Landslide susceptibility mapping at hoa binh province (Vietnam) using an adaptive neuro-fuzzy inference system and GIS. Comput. Geosci. 45, 199–211 (2012).

Pradhan, B. A. Comparative study on the predictive ability of the decision tree, support vector machine and neuro-fuzzy models in landslide susceptibility mapping using GIS. Comput. Geosci. 51, 350–365 (2013).

Saito, H., Nakayama, D. & Matsuyama, H. Comparison of landslide susceptibility based on a decision-tree model and actual landslide occurrence: The Akaishi mountains, Japan. Geomorphology. 109, 108–121 (2009).

Yeon, Y. K., Han, J. G. & Ryu, K. H. Landslide susceptibility mapping in Injae, korea, using a decision tree. Eng. Geol. 116, 274–283 (2010).

Yao, X., Tham, L. & Dai, F. Landslide susceptibility mapping based on support vector machine: A case study on natural slopes of hong kong, China. Geomorphology. 101, 572–582 (2008).

Yilmaz, I. Comparison of landslide susceptibility mapping methodologies for koyulhisar, Turkey: Conditional probability, logistic regression, artificial neural networks and support vector machine. Environ. Earth. Sci. 61, 821–836 (2010).

Peng, L. et al. Landslide susceptibility mapping based on rough set theory and support vector machines: A case of the three gorges area, China. Geomorphology. 204, 287–301 (2014).

Pradhan, B. Landslide susceptibility mapping of a catchment area using frequency ratio, fuzzy logic and multivariate logistic regression approaches. J. Indian. Soc. Remote. Sens. 38, 301–320 (2010).

Pradhan, B. & Lee, S. Regional landslide susceptibility analysis using back-propagation neural network model at cameron highland, Malaysia. Landslides. 7, 13–30 (2010).

Metcalfe, I. Tectonic evolution of the malay peninsula. J. Asian. Earth. Sci. 76, 195–213 (2013).

Pradhan, B. Remote sensing and gis-based landslide hazard analysis and cross-validation using multivariate logistic regression model on three test areas in Malaysia. Adv. Space. Res. 45, 1244–1256 (2010).

Article   CAS   ADS   Google Scholar  

Pradhan, B., Lee, S. & Buchroithner, M. F. A GIS-based back-propagation neural network model and its cross-application and validation for landslide susceptibility analyses. Computers, Comput. Environ. Urban Syst. 34, 216–235 (2010).

Matori, A. N., Basith, A. & Harahap, I. S. H. Study of regional monsoonal effects on landslide hazard zonation in cameron highlands, Malaysia. Arab. J. Geosci. 5, 1069–1084 (2012).

Shahabi, H., Ahmad, B. & Khezri, S. Evaluation and comparison of bivariate and multivariate statistical methods for landslide susceptibility mapping (case study: Zab basin). Arab. J. Geosci. 6, 3885–3907 (2013).

Roessner, S., Wetzel, H. U., Kaufmann, H. & Sarnagoev, A. Potential of satellite remote sensing and GIS for landslide hazard assessment in southern kyrgyzstan (central Asia). Nat. Hazards. 35, 395–416 (2005).

Blaschke, T. Object based image analysis for remote sensing. ISPRS J. Photogramm. Remote Sens. 65, 2–16 (2010).

Gibson, P. J., Power, C. H., Goldin, S. E. & Rudahl, K. T. Introductory remote sensing: Digital image processing and applications (Routledge, London., 2000).

Askne, J. & Santoro, M. Multitemporal repeat pass sar interferometry of boreal forests. IEEE. T. Geosci. Remote. 43, 1219–1228 (2005).

Shahabi, H., Khezri, S., Ahmad, B. B. & Musa, T. A. Application of moderate resolution imaging spectroradiometer snow cover maps in modeling snowmelt runoff process in the central Zab basin, Iran. J. Appl. Remote Sens. 8, 084699–084699 (2014).

Saaty, T. L. & Vargas, L. G. Inconsistency and rank preservation. J. Math. Psychol. 28, 205–214 (1984).

Article   MathSciNet   Google Scholar  

Sahnoun, H., Serbaji, M. M., Karray, B. & Medhioub, K. GIS and multi-criteria analysis to select potential sites of agro-industrial complex. Environ. Earth. Sci. 66, 2477–2489 (2012).

Saaty, T. L. A scaling method for priorities in hierarchical structures. J. Math. Psychol. 15, 234–281 (1977).

Feizizadeh, B., Blaschke, T., Nazmfar, H. & Rezaei Moghaddam, M. Landslide susceptibility mapping for the urmia lake basin, Iran: A multi-criteria evaluation approach using GIS. Int. J. Environ. Res. 7, 319–336 (2013).

Google Scholar  

Komac, M. A landslide susceptibility model using the analytical hierarchy process method and multivariate statistics in perialpine Slovenia. Geomorphology. 74, 17–28 (2006).

Kritikos, T. R. & Davies, T. R. GIS-based multi-criteria decision analysis for landslide susceptibility mapping at northern Evia, Greece. Z. Dtsch. Ges. Geowiss. 162, 421–434 (2011).

Hong, Y., Adler, R. & Huffman, G. Use of satellite remote sensing data in the mapping of global landslide susceptibility. Nat. Hazards. 43, 245–256 (2007).

Feizizadeh, B. & Blaschke, T. GIS-multicriteria decision analysis for landslide susceptibility mapping: Comparing three methods for the urmia lake basin, Iran. Nat. Hazards. 65, 2105–2128 (2013).

Günther, A., Van Den Eeckhaut, M., Malet, J. P., Reichenbach, P. & Hervás, J. Climate-physiographically differentiated pan-european landslide susceptibility assessment using spatial multi-criteria evaluation and transnational landslide information. Geomorphology. 224, 69–85 (2014).

Malczewski, J. GIS and multicriteria decision analysis (John Wiley & Sons, New York., 1999).

Rahman, R. & Saha, S. Remote sensing, spatial multi criteria evaluation (SMCE) and analytical hierarchy process (AHP) in optimal cropping pattern planning for a flood prone area. J. Spat. Sci. 53, 161–177 (2008).

Pellicani, R., Van Westen, C. J. & Spilotro, G. Assessing landslide exposure in areas with limited landslide information. Landslides. 11, 463–480 (2014).

Van Westen, C., Rengers, N. & Soeters, R. Use of geomorphological information in indirect landslide susceptibility assessment. Nat. Hazards. 30, 399–419 (2003).

Ayalew, L., Yamagishi, H., Marui, H. & Kanno, T. Landslides in sado island of japan: Part ii. GIS-based susceptibility mapping with comparisons of results from two methods and verifications. Eng. Geol. 81, 432–445 (2005).

Baeza, C. & Corominas, J. Assessment of shallow landslide susceptibility by means of multivariate statistical techniques. Earth Surf. Process. Landf. 26, 1251–1263 (2001).

Download references

Acknowledgements

This study was conducted as a part of post-doctoral fellowship scheme granted by Universiti Teknologi Malaysia (UTM).

Author information

Authors and affiliations.

Geoscience and Digital Earth Centre (Geo-DEC), Research Institute for Sustainability and Environment (RISE) Universiti Teknologi Malaysia, 81310 UTM, Johor Bahru, Malaysia

Himan Shahabi & Mazlan Hashim

You can also search for this author in PubMed   Google Scholar

Contributions

H.S. wrote the methods and interpretation of results. M.H. contributed to the writing and provided the fund opportunities. All the authors contributed to the interpretation and reviewed the manuscript.

Ethics declarations

Competing interests.

The authors declare no competing financial interests.

Rights and permissions

This work is licensed under a Creative Commons Attribution 4.0 International License. The images or other third party material in this article are included in the article's Creative Commons license, unless indicated otherwise in the credit line; if the material is not included under the Creative Commons license, users will need to obtain permission from the license holder in order to reproduce the material. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/

Reprints and permissions

About this article

Cite this article.

Shahabi, H., Hashim, M. Landslide susceptibility mapping using GIS-based statistical models and Remote sensing data in tropical environment. Sci Rep 5 , 9899 (2015). https://doi.org/10.1038/srep09899

Download citation

Received : 24 November 2014

Accepted : 05 March 2015

Published : 22 April 2015

DOI : https://doi.org/10.1038/srep09899

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

This article is cited by

Modeling landslides hazard along amman–jerash–irbid highway, jordan by integrating open street map (osm) and weighted linear combination (wlc) techniques.

• Abdulla Al-Rawabdeh
• Muheeb Awawdeh
• Rasheed Jaradat

Modeling Earth Systems and Environment (2024)

The influence of cartographic representation on landslide susceptibility models: empirical evidence from a Brazilian UNESCO world heritage site

• Jefferson Alves Araujo Junior
• Cesar Falcão Barella
• Alberto Fonseca

Natural Hazards (2024)

Comparative Assessment of FR and AHP Models for Landslide Susceptibility Mapping for Sikkim, India and Preparation of Suitable Mitigation Techniques

• Brototi Biswas
• Aneesah Rahaman
• Jonmenjoy Barman

Journal of the Geological Society of India (2023)

Lightweight deep learning model for automatic landslide prediction and localization

• Payal Varangaonkar

Multimedia Tools and Applications (2023)

Implementation of free and open-source semi-automatic feature engineering tool in landslide susceptibility mapping using the machine-learning algorithms RF, SVM, and XGBoost

• Emrehan Kutlug Sahin

Stochastic Environmental Research and Risk Assessment (2023)

By submitting a comment you agree to abide by our Terms and Community Guidelines . If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.