Integration of Explainable Artificial Intelligence to Identify Significant Landslide Causal Factors for Extreme Gradient Boosting based Landslide Susceptibility Mapping with Im

01/10/2022
by   Muhammad Sakib Khan Inan, et al.
13

Landslides have been a regular occurrence and an alarming threat to human life and property in the era of anthropogenic global warming. An early prediction of landslide susceptibility using a data-driven approach is a demand of time. In this study, we explored the eloquent features that best describe landslide susceptibility with state-of-the-art machine learning methods. In our study, we employed state-of-the-art machine learning algorithms including XgBoost, LR, KNN, SVM, Adaboost for landslide susceptibility prediction. To find the best hyperparameters of each individual classifier for optimized performance, we have incorporated the Grid Search method, with 10 Fold Cross-Validation. In this context, the optimized version of XgBoost outperformed all other classifiers with a Cross-validation Weighted F1 score of 94.62 by incorporating TreeSHAP and identified eloquent features such as SLOPE, ELEVATION, TWI that complement the performance of the XGBoost classifier mostly and features such as LANDUSE, NDVI, SPI which has less effect on models performance. According to the TreeSHAP explanation of features, we selected the 9 most significant landslide causal factors out of 15. Evidently, an optimized version of XgBoost along with feature reduction by 40 other classifiers in terms of popular evaluation metrics with a Cross-Validation Weighted F1 score of 95.01 97

READ FULL TEXT VIEW PDF

Authors

page 3

05/18/2018

Optimizing for Generalization in Machine Learning with Cross-Validation Gradients

Cross-validation is the workhorse of modern applied statistics and machi...
03/13/2019

Predicting class-imbalanced business risk using resampling, regularization, and model ensembling algorithms

We aim at developing and improving the imbalanced business risk modeling...
10/18/2021

Gradient boosting with extreme-value theory for wildfire prediction

This paper details the approach of the team Kohrrelation in the 2021 Ext...
10/03/2021

Classification of Viral Pneumonia X-ray Images with the Aucmedi Framework

In this work we use the AUCMEDI-Framework to train a deep neural network...
09/12/2020

Interpretable Machine Learning Approaches to Prediction of Chronic Homelessness

We introduce a machine learning approach to predict chronic homelessness...
02/21/2019

Stacking with Neural network for Cryptocurrency investment

Predicting the direction of assets have been an active area of study and...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Landslides are subject to a great concern for Geological and Environmental researchers from all over the world due to it’s irreparable and execrable impact on environment, society and economy. Landslide is a natural calamity which is characterized by the movement of a mass rock, debris or earth down a slope. Several environmental factors like heavy rainfall, geographical factor like location, land near volcanoes etc contribute to the occurrence of landslides. Especially, in some areas, Rainfall, slope, Land Use and Land Cover Change, and Elevation etc are one of the most influencing factors for Landslide (Mind’je et al., 2020). Hilly and Coastal areas all over the globe are mostly vulnerable to frequent and most devastating landslides. More than thousand people are killed by landslides every year around the globe, including an average of 25 - 50 deaths all alone in United States. (Survey, 2021). According to the study of Sultana (2020), considering the time period of 2000-2018, the yearly average number of landslides in Bangladesh is 19, with a 4% rate of increase per year, which ultimately results in 38 fatalities and 54 injuries on average. Landslides often damage road networks in hilly areas causing great direct or indirect consequential economic losses due to hindrance in communication with the hilly parts of a country (Winter et al., 2016). Landslides are great threat to socio-economic conditions of a country (Perera et al., 2018). The impact of a landslide may be extended to destruction of important infrastructure, cultivable land and natural resources. It may lead to blockage of rivers and intensify the risk of floods (FAO, 2021). The study Landslide Susceptibility is sensitive and arduous due to the presence of uncorrelated or non-linearly correlated environmental factors responsible for Landslides (Huang et al., 2020a). In this context, an extensive statistical data driven analysis could help us to detect useful hidden and confounding patterns for identification of landslides to support effective measures to prevent this disaster.

In recent years, the global issue of Landslide, an alarming threat to mankind, has drawn attention of the Artificial Intelligence researchers. Considering the exigency of an early and automated prediction of Landslides, along with Geologists and Environmental scientists, AI (Artificial Intelligence) researchers from all over the world has devoted themselves in the extensive study of Landslide Susceptibility Mapping using Artificial Intelligence based methods. As a result, over the years a decent number of state of art studies have been conducted for Landslide Susceptibility Mapping with Machine Learning, Deep Learning and Artificial Neural Networks. The immense potential of Machine Learning algorithms can be utilized to automate and improve the efficiency of the analysis and prediction of Landslide Susceptibility

(Sahin et al., 2020).

Machine Learning methods including Naïve Bayes (NB), Multilayer Perceptron (MLP), Kernel Logistic Regression (KLR), and J48-bagging was employed for Landslide Susceptibility Mapping considering the area of Youfang district, China in the study of

Hong et al. (2019)

and it is observed from the study that MLP(Multilayer Perceptron) outperformed all other classifiers and proved to be an efficient tool for landslide study of this area.

Sahana et al. (2020)

proposed a hybrid neural network classifier integrating Multilayer Perceptron (MLP) and Bagging for an efficient mapping susceptibility of rainfall induced landslides to support identification of vulnerable areas for disaster prevention and management. Several other studies have also found that Multilayer Perceptron(MLP) or a hybrid combination of Multilayer Perceptron(MLP) and Particle Swarm Optimization is an compelling classifier in the study of Landslides

(Pham et al., 2017; Li et al., 2019)

. Deep Learning algorithms excelled the study of Landslide Susceptibility prediction and analysis due to their deep architectures supporting robust in-built feature extraction strategies and great capacity to tackle confounding and sensitive factors.

Huang et al. (2020a)

found in their study that a Fully Connected Sparse Autoencoder based model outperformed the traditional models for extracting optimal non-linear features from environmental factors. Moreover, a spatially explicit deep neural network was proposed by

(Dao et al., 2020), where for feature selection Relief-F method was integrated to quantify the utility of the conditioning factors. Deep Neural Networks found to be comparatively outperforming over the conventional machine learning classifiers in the domain of Landslide Susceptibility in several other state of art studies (Bui et al., 2020; Zhu et al., 2020).

Though Deep Learning based methods can efficiently predict Landslides, the architectures of these algorithms are complex in structure and computationally expensive. Even for training purpose, to build an effective model, deep neural networks need a large amount of training data.

In this context, to find an efficient but less computationally expensive framework for Landslide study, researchers have studied the potential of Ensemble based, Probabilistic and Hybrid Machine Learning classifier for Landslide Susceptibility Prediction. Bagging based Reduced Error Pruning Trees (BREPT), a novel hybrid machine learning classifier, was proposed by Pham et al. (2019) with notable performance in Landslide Susceptibility. Similarly, another tree based hybrid classifier was proposed by Thai Pham et al. (2019) with impressive AUC scores. Also, Classification and Regression Tree (CART) algorithm outperformed Mulitlayer Perceptron based classifiers in some recent studies(Huang et al., 2020b). Some state art studies outlined the extreme potential of ensemble based or gradient boosting based machine learning classifiers through comparative analysis of ensemble machine learning methods and gradient boosting algorithm with several other methods (Fang et al., 2021; Sahin, 2020)

. Tree based enemble algorithms showed prominent performance; Random Forest leading with impressive results in the study of

Merghadi et al. (2020). Hybrid Random Forest based models with GeoDetector and RFE for factor optimization were developed by (Zhou et al., 2021)

. Extreme Boosting algorithms tends to outperform in comparison with linear classifier by robust feature extraction and abilty to deal with outliers

(Rabby et al., 2021)

. However, Support Vector Machine and Logistic Regression exhibited significant performance in Landslide Susceptibility prediction and also outperformed ANN(Artificial Neural Network) in some studies

(Chen et al., 2018; Kalantar et al., 2018)

. A novel machine learning method with the integration of unsupervised machine learning method K- Means Clustering and supervised Decision Tree based classifier was proposed by

Guo et al. (2021).

Despite a large number of outstanding researches have been conducted on Landslide Susceptibility using Machine Learning and Deep Learning methods, no study have solved the issue of explainability of these state of art Artificial Intelligence algorithms. As the structure of these state of art algorithms are very complex from a mathematical point of view, it often engenders a BlackBox problem which may sometimes lead to the inefficiency of these framework in the future. And also, no previous studies have dedicatedly investigated the possibility of successful prediction of Landslides using less number of geological and topological factors with the help of machine learning based automated systems. In this concern, in this study, we have employed Explainable Artificial Intelligence framework for predicting Susceptibility of Landslides. The key contributions of this study are as follows;

  • Identification of most important geological, topological and hydrological factors that best corroborate the performance of automated machine learning model for the assessment and early prediction of Landslide Susceptible Area.

  • Extensive performance analysis of machine learning classifiers to identify most suitable Machine Learning algorithm for landslide susceptibility mapping.

  • Proposal of an optimized integrated ensemble based machine learning solution which needs less number of landslide causal features to efficiently classify Landslides by reducing costs and time consumed in the primary stage Landslide related studies.

2 Methods

In this section, an illustration of the methods that implied in our study, have been delineated in a detailed manner. The research framework of our study is graphically represented at Figure 1.

Figure 1: Our Research Framework

2.1 Normality Test

Normality test is a statistical analysis-based method that is used to determine whether or not the distribution of data follows a Gaussian Distribution. In data mining, it is often important to comprehend the distribution of data that helps to discern whether to use parametric or non parametric statistical methods for exploratory data analysis. If the data follows a Gaussian distribution, then parametric statistical methods are incorporated, else non-parametric statistical methods are incorporated. In this context, to understand the data distribution of our feature variables involved in Landslide Susceptibility prediction we employed a popular and reliable normality evaluation method

(Yap and Sim, 2011)

, Sharpio-Wilks, that generates p-values based on test statistics (W-statistic). The W-statistic or test statistic is computed as

(Royston, 1992);

(1)

Where denotes the test statistic value, denotes the ith order statistic and denotes the mean of the sample. And also, , the coefficient can be stated as;

(2)

Here

is a vector norm that can be donated by;

(3)

The following Table 1 represents the W-Statistic values computed through the Sharpio-Wilks test for each individual feature variable of our dataset.

Feature Name W-Statistic Feature Name W-Statistic
PROFILE 0.752 SLOPE 0.771
PLAN 0.636 ASPECT 0.864
CHANGE 0.685 TWI 0.868
LANDUSE 0.665 SPI 0.749
ELEVATION 0.735 DRAINAGE 0.809
NDVI 0.857 ROAD 0.642
RAINFALL 0.818 GEOLOGY 0.847
FAULTLINES 0.684
Table 1: Sharpio-Wilk Test Statistic of Feature Variables

P-values illustrate that from a given data sample how likely the data was drawn from a Gaussian distribution based on a certain threshold. Evidently, a p-value less than or equal to the threshold value of 0.05 delineates that that data sample does not follow a normal or Gaussian distribution and a p-value greater than 0.05 delineates that data was drawn from a normal distribution. In this study, based on the normality test analysis, we found that the data samples from any individual independent variable were not drawn from a normal distribution. However, numerous characteristics of the dataset are continuous in nature when it comes to the area of Geotechnical Engineering, which is poorly reflected by the data distribution. According to this research, a linear machine learning model would not be the best match to capture complicated multi-co-linearity difficulties when applied to a different dataset’s global point of view. Considering the analysis, we successfully applied the geotechnical engineering domain to identify the best potential machine learning solution throughout our study.

2.2 Chi-Square Test

In data mining and inference based research, it is important to understand whether or not a certain feature contributes to the final outcome. The statistical feature significance test helps us to identify and select features that strongly corroborate the final prediction. In Geo-technical Science, identifying key factors that best predict a Landslide Susceptibility is an indispensable part. In our study, we decided to employ a non-parametric statistical significance-based test, the Chi-Square test, to identify the importance of our feature variables in Landslide Susceptibility prediction based on the normality test results. The Chi-Square test is a non-parametric statistical significance analysis method that is suitable for analyzing the significance of independent variables. To state mathematically (Singhal and Rana, 2015; McHugh, 2013),

(4)

Here,

denotes "Degree of Freedom",

denotes "observed value" and denotes "expected value", and

is the ”ith” position is in the contingency table. The value of

is computed through,

(5)

Here, denotes "Number of Data Instances and denotes "Number of Features". Afterward, we determine the statistically significant P-values for the independent variable against the dependent variable using the computed Chi-Square and degree of freedom values. The p-values computed through the Chi-square test are shown in Table 2.

According to the p-value threshold of 0.05, we can infer that all of our feature variables show statistical evidence to espouse the prediction of Landslide Susceptibility.

Feature Name P-Value Feature Name P-Value
PROFILE 3.19E-94 SLOPE 9.73E-291
PLAN 1.13E-123 ASPECT 4.04E-17
CHANGE 1.24E-57 TWI 2.89E-308
LANDUSE 6.56E-158 SPI 8.93E-11
ELEVATION 7.11E-162 DRAINAGE 7.77E-17
NDVI 9.96E-98 ROAD 0
RAINFALL 6.00E-82 GEOLOGY 4.32E-88
FAULTLINES 1.27E-48
Table 2: p-value Scores From Chi-Square Test

Here, a p-value of less than or equal to 0.05 indicates that the particular categorical feature variable significantly contributes to the classification of Landslide Susceptibility with strong statistical evidence.

2.3 Machine Learning Classifiers

In our study, we have employed 5 machine learning algorithms for experimental purpose and optimized their performance considering popular evaluation metrics.

2.3.1 Support Vector Machine

Support Vector Machine (SVM) is a machine learning classifier which works in a supervised manner. It delineates a boundary to disparate the data point through analyzing the datapoints from training set based on differences in data distribution across individual classes. This boundary is called decision boundary which helps to classify between classes like Landslide Susceptible or Non-Landslide Susceptible. The decision boundary of SVM is a hyperplane in an N-dimensional space which evidently classifies the data points of different classes. Data points that are closer to the hyperplane and evinces a significant impact on the hyperplane’s direction and orientation are called support vectors. The performance of the SVM is optimized through support vectors. The weights associated with these support vectors corroborates the classifier to shape to orientations of the hyperplane. SVM is an empirically popular classifier that has exhibited noteworthy performance in the domain of Landslide Susceptibility Mapping

(Huang and Zhao, 2018).

SVM can be further divided into to major category, Linear SVM and Non-Linear SVM. The category of SVM is decided based on the kernel function it uses to draw the hyperplane. In Non-Linear SVM, the radial basis function is computed as, for Kernel, K,

(6)

Where and are data points and is a free parameter that controls the degree of generalization. For better optimization, the generalization parameter and kernel function is tweaked to find the combination that best supports the classification of Landslide Susceptibility.

2.3.2 Logistic Regression

Logistic Regression is a linear machine learning classifier, well suitable for handling categorical and numerical feature variables and capable of robustly predicting binary outcomes. Logistic Regression implies a log odds ratio which is an iterative maximum likelihood method to predict whether a given set of features belongs to a certain class. Logistic Regression follows the logistic function or sigmoid function to draw a decision boundary between the two classes, Landslide Susceptible area and Non-Landslide Susceptible area. For an uni-variate logistic regression, To state mathematically,

(7)

Where, is target variable, indicates bias and, indicates the coefficient or weights of feature variable . Logistic Regression is widely used in the domain of Landslide Susceptibility Mapping due to it’s simple structure and effective performance for binary classification problems (Lombardo and Mai, 2018).

2.3.3 K Nearest Neighbours

K Nearest Neighbours (KNN) is a non-parametric machine learning algorithm. It is also called a lazy learning algorithm, much suitable for smaller datasets (Guo et al., 2003). To classify data instances into certain category, KNN employs a neighbourhood similarity analyzing method incorporating distance metrics like Manhattan Distance, Euclidean Distance etc, to successfully create discriminating supervised clusters of individual class labels like Landslide Susceptible or Non-Landslide Susceptible. For final classification of Landslide Susceptibility, it would follow a majority voting criteria, considering the class labels of a certain number of nearest neighbours based on the results of distance metrics analysis. The generalized formula of computing distances for KNN is;

(8)

Here, and are datapoints of according feature variable and target variable, and controls which distance metrics to be used. For Manhattan Distance is set to 1 and for Euclidean Distance is set to 2.

2.3.4 AdaBoost

Adaptive Boosting (AdaBoost) is a tree based machine learning classifiers, which aggregates multiple weak learners (decision trees) in an ensemble manner to build a robust classifier (Freund and Schapire, 1997). A single decision tree or weak learner in Adaboost is called a stump which has maximum depth of 1 with one root and two leaves. The algorithm assigns more weight on difficult to classify instances and less on those which are well classified. Thus, it creates a forest of stumps to efficiently classify a Landslide Susceptible and Non-Landslide Susceptible area. The amount of say for each stump is determined based on the classification error which is basically the sum of weights for incorrectly classified samples (Schapire, 2013). The number stumps to use for classification has an significant impact on the classification performance. It is an extremely fast classifier compared to other tree based classifier which makes is a strong candidate for landslide susceptibility prediction.

2.3.5 Extreme Gradient Boosting(XgBoost)

Extreme Gradient Boosting (XgBoost) algorithm is a tree based ensemble learning classifier that solves the overfitting issue previously present in decision tree-based classifiers with its improved gradient boosting strategy with built-in regularization and impressive gains in speed (Chen and Guestrin, 2016). The improved regularization strategy makes this algorithm so robust that it has evinced notable performance for solving problems that include a large amount of unstructured data (text, images, etc) or dataset containing outliers and features that are sensitive in nature (Khan Inan et al., 2021). In comparison to the previous general Gradient Boosting (GBM), a number of new design features such as robustness in handling missing values, approximate and sparsity aware split-finding algorithm, parallel computing, cache-aware access, block compression, and sharding have made XgBoost an effective choice for complex and sensitive classification problems.

The gradient descent method is employed to build every decision trees uniquely, followed by beginning with a certain threshold and modifying the weights in an iterative manner by minimizing residuals in every single iteration. So, the trees built after every iteration remains unique as the error or mistakes done by the previous tree is minimized or regularized in the next tree to build.

Mathematically, residuals are calculated to tackle the problem of unique trees. Residuals are errors between observed and predicted values. Each tree starts with a single leaf and all of the residuals go to the leaf.

(9)

Here,

denotes previously predicted probability for ith leaf. Similarity Score,

, is calculated for each new leaf.

(10)

Here, in equation of (10), is the regularization parameter that controls pruning of trees, and and denotes residuals and cover accordingly.

The Gain of the trees are computed through calculating the similarity scores of left, right and root nodes. It tells us to where to split the data.

(11)

Here, stands for Gain, , , denote accordingly, similarity score of Left, Right and Root Node. Considering the immense potential of Extreme Gradient Boosting algorithm we adopted the algorithm for our study of Landslide Susceptibility Prediction.

2.4 Exhaustive Grid Search

Grid Search is an exhaustive searching method popularly used for hyperparameter tuning of Machine Learning algorithms. This method incorporates grid based parameter search by which it computes every possible combination of parameters from a given set of values. It helps to optimize the performance along with reducing overfitting issue of Machine Learning algorithms to build an efficient classifier based on certain data. In our study, to optimize the machine learning algorithms for landslide susceptibility prediction, we have incorporated the Grid Search method for finding best hyperparameters for every individual classifiers validating by 10-Fold Stratified Cross Validation of Weighted F1 Scores. The grid of hyperparameters which has been optimized with Grid Search depicted in the Table 3.

Classifier HP Definition Parameters Grid
XgBoost max_depth Maximum Depth of a Tree [2,3,5,6,8]
n_estimators Number of trees
[500, 1500,
3000, 5000]
learning_rate Learning Rate
[0.01, 0.1,
0.05, 0.3, 0.5]
gamma Regularization Parameter
[0, 0.1, 0.5,
1, 2]
subample Percentage of Training Rows
[0.5, 0.7,
0.8, 0.9, 1]
KNN n_neighbors Number of Neighbours
[3, 5, 7, 9,
11, 13]
p
Exponent of
Minkowski distance
[1, 2]
LR C Regularization Parameter
[0.001, 0.01, 0.1,
1, 10, 100]
solver Algorithm to use
[’newton-cg’, ’lbfgs’,
’liblinear’, ’sag’, ’saga’]
penalty Regularization Algorithm [’l1’, ’l2’, ’elasticnet’]
SVM C Regularization Parameter [0.01, 0.1, 1, 10, 100]
kernel Kernel Trick [’linear’, ’rbf’]
Adaboost n_estimators Number of trees
[0.001, 0.01, 0.1,
0.15, 0.2, 0.3, 0.5, 1]
learning_rate Learning Rate
[10, 50, 100,
500, 1000, 1500, 3000]
Table 3: Range of Hyperparemeters (HP) and Definition

2.5 TreeSHAP Explanation

TreeSHAP is an extension of SHAP (SHapley Additive exPlanations) method which elucidates the prediction or output of Machine Learning algorithms by computing Shapley values for a given data instance that delineates what is the sum of contributions from it’s individual feature variables (Lundberg and Lee, 2017; Lundberg et al., 2018)

. It is a game theoretic approach instigated as a fast and model-specific alternative to KernelSHAP for decision tree based algorithms. Shapley values, a coalitional game theory technique, illustrates how to allocate the prediction of individual instances among the characteristics in a fair manner. Shapley values are computed as

(Molnar, 2021);

(12)

Here, denotes prediction function of Machine Learning classifier and stands for total number of features. represents any subset of features that doesn’t include the feature and is the size of that subset.

In our study, we have incorporated TreeSHAP method for analyzing the predictions of Extreme Gradient Boosting algorithm to solve the blackbox issue of machine learning algorithms in a motivation to build Explainable Artificial Intelligence based solution along with identifying which of the features corroborates mostly an effective classification of Landslide Susceptibility.

3 Experimental Results

3.1 Dataset Description

In our study, we have experimented our research methods on a benchmark dataset of Landslide Susceptibility Mapping considering three Upazilas of Rangamati Hill District, Bangladesh which was prepared by Rabby et al. (2021).This district is extremely susceptible to landslides due to its geographical location and several natural factors. The natural climate factors like heavy rainfall and geographical factor-like the highest average slope gradients have made this district prone to landslides and a convincing candidate in terms of the study area for landslide research. The following dataset contains 196 data instances of each of the two target classes, Landslide Susceptible and Non-Landslide Susceptible. 15 important Geological, Topological and Hydrological factors for Landslide Susceptibility Mapping are presented in this dataset. In our study, we have included these 15 landslide causal factors as feature variable for predicting Landslide Susceptibility. The following feature variables are; ’PROFILE’ (Profile Curvature), ’PLAN’ (Plan Curvature), ’CHANGE’,’LANDUSE’, ’ELEVATION’, ’SLOPE’, ’ASPECT’, ’TWI’, ’SPI’, ’DRAINAGE’ (Distance from Drainage Network), ’NDVI’ (Normalized Vegetation Index), ’RAINFALL’, ’FAULTLINES’ (Distance to Fault lines), ’ROAD’ (Distance to Road Network), ’GEOLOGY’. Table 4 delineates the methods utilized to compute these landslide causal factors by incorporating remote sensing technologies (Abedin et al., 2020). A brief description of the features which have been analyzed and utilized in this study are given below:

Landslide Casual Factor Methods Used to Determine the Factor
PROFILE CURVATURE ArcGIS Curvature Function
PLAN CURVATURE ArcGIS Curvature Function
LAND COVER CHANGE Anderson scheme Level-I method (Satelite Images)
LANDUSE CHANGE Anderson scheme Level-I method (Satelite Images)
ELEVATION
Advanced Spaceborne Thermal Emisssion and
Reflection Radiometer (ASTER) Global Digital
Elevation Map (GDEM)
NDVI LandSat 8 level 2 imagery
RAINFALL

Kriging Interpolation

FAULTLINES
Euclidean
Distance tool in ArcGIS
SLOPE Slope tool in ArcGIS
ASPECT Aspect tool in ArcGIS
TWI ASTER GDEM
SPI ASTER GDEM
DRAINAGE
Euclidean
Distance tool in ArcGIS
ROAD
Euclidean
Distance tool in ArcGIS
GEOLOGY Geological Survey
Table 4: Methods To Compute Landslide Causal Factors

3.1.1 PROFILE (Profile Curvature)

PROFILE (Profile Curvature) is an important factor used in the study of Landslide Susceptibility Mapping. The curvature in the downslope direction along a line produced by the intersection of an imaginary vertical plane with the ground surface is known as Profile Curvature. The driving and resistive strains within a landslide in the direction of motion are affected by profile curvature. (Carson and Kirkby, 1972; Meten et al., 2015). The data count distribution of the feature PROFILE is depicted in Figure 2.

Figure 2: Count Distribution of Data of Profile Curvature

3.1.2 PLAN (Plan Curvature)

The curvature of topographic contours or the curvature of a line generated by the intersection of an imaginary horizontal plane with the ground surface is referred to as Plan Curvature (PLAN). The convergence or divergence of landslide material and water in the direction of landslide motion is controlled by plan curvature (Ohlmacher, 2007; Meten et al., 2015). The data count distribution of the feature PLAN is depicted in Figure 3.

Figure 3: Count Distribution of Data of Plan Curvature (PLAN)

3.1.3 Change

The loss of natural areas, notably forests, to urban or suburban development, or the loss of agricultural regions to urban or exurban development is referred to as land cover change (CHANGE) (Sealey et al., 2018). A number of studies have found that land cover change (CHANGE) is a notable and strongly influencing factor towards determination of the susceptibility of Landslides for a particular area (Promper et al., 2014; Restrepo and Alvarez, 2006). The data count distribution of the feature CHANGE is depicted in Figure 4.

Figure 4: Count Distribution of Data of Change

3.1.4 Landuse

Land use is usually described as a sequence of human-performed activities on land with the goal of obtaining goods and advantages from the usage of land resources (Reichenbach et al., 2014). In the hilly or mountainous areas, Land use change (LANDUSE) can increase or decrease the possibility of landslides with potential influence (Chen et al., 2019). The data count distribution of the feature LANDUSE is depicted in Figure 5.

Figure 5: Count Distribution of Data of Landuse

3.1.5 Elevation

Landslide vulnerability is frequently assessed using elevation. The altitude of terrain is referred to as elevation. According to Dou et al. (2015), the ground at various heights will have varying levels of sensitivity which is a key factor to identify the probability of possible Landslide events. The data count distribution of the feature ELEVATION is depicted in Figure 6.

Figure 6: Count Distribution of Data of Elevation

3.1.6 Slope

The angle measured between a horizontal plane and a particular location on the ground surface is known as the slope angle (SLOPE) (Whitworth et al., 2011). SLOPE is one of most influential factors that can lead to causing serious landslides. SLOPE also has a notable correlation with other geological and topological factors which made it an early alarm to assess the susceptibility of landslides in the hilly parts of the world. In general, the likelihood of a landslide rises as the slope rises (Meten et al., 2015).The data count distribution of the feature SLOP is depicted in Figure 7.

Figure 7: Count Distribution of Data of Slope Angle (SLOPE)

3.1.7 Aspect

The aspect at a location on the ground surface, according to some researchers, is the direction that the tangent plane passing through that point faces and is represented in degrees. The aspect, in its most basic form, is a data type that indicates the geographical direction in which the slopes grow (Tanoli et al., 2017). The data count distribution of the feature ASPECT is depicted in Figure 8.

Figure 8: Count Distribution of Data of ASPECT

3.1.8 Twi

The Topographic Wetness Index (TWI), also known as the compound topographic indicator, is a wetness index that measures steady-state conditions. It is widely used to quantify the influence of topography on hydrological processes which has great influence in the occurance of Landslides (Mattivi et al., 2019). The data count distribution of the feature TWI is depicted in Figure 9.

Figure 9: Count Distribution of Data of TWI

3.1.9 Spi

The erosive force of a stream or water flow is measured by the SPI (Stream Power Index). The slope and contributing area are used to calculate SPI. SPI approximates the locations on the landscape where gullies are more likely to form (Abedin et al., 2020). The data count distribution of the feature SPI is depicted in Figure 10.

Figure 10: Count Distribution of Data of SPI

3.1.10 Drainage

DRAINAGE refers to distance to drainage network. It is usually noticed that area near to drainage network are more prone to landslides which makes it a very crucial feature for the study of landslides (Abedin et al., 2020). The data count distribution of the feature DRAINAGE is depicted in Figure 11.

Figure 11: Count Distribution of Data of DRAINAGE

3.1.11 Ndvi

The Normalized Difference Vegetation Index (NDVI) is used to calculate the density of green on a given plot of land. It measures vegetation by comparing the amount of near-infrared light reflected by vegetation to the amount of red light absorbed by vegetation. It has been identified as a good indicator of landslide susceptibility according to geotechnical researchers (Dahigamuwa et al., 2016). The data count distribution of the feature NDVI is depicted in Figure 12.

Figure 12: Count Distribution of Data of NDVI

3.1.12 Rainfall

The amount of rainfall in a particular hilly area is a great indicator of landslide susceptibility. Excessive rainfall is often considered as a potential trigger for sudden and destructive landslides in the hilly regions (Abedin et al., 2020). Heavy rainfall can induce soil saturation, and debris flow can occur on certain slopes triggering the possibilty of rainfall induced landslides (Chen et al., 2017). The data count distribution of the feature RAINFALL is depicted in Figure 13.

Figure 13: Count Distribution of Data of RAINFALL

3.1.13 Faultlines

Fault lines (FAULTLINES) are geological variables in a Lanslides research that suggest tectonic breaks and reduce rock strength. In general, areas closer to the Faultline are more prone to landslides than areas further away (Regmi et al., 2014; Abedin et al., 2020). The data count distribution of the feature FAULTLINES is depicted in Figure 14.

Figure 14: Count Distribution of Data of FAULTLINES

3.1.14 Road

ROAD refers to the distance from the road of land that is a crucial measure to assess the landslide susceptibility of an area. Roads assist to concentrate drainage, while road cuttings harm the slope structure. Landslides near roadways might occur if the required precautions are not taken (Chen et al., 2017). The data count distribution of the feature ROAD is depicted in Figure 15.

Figure 15: Count Distribution of Data of ROAD

3.1.15 Geology

Geology is concerned with the permeability and strength of a region’s rocks and soil, and hence with landslides (Ayalew and Yamagishi, 2005). Understanding the geology of land area is always considered to be a crucial factor for effective study of Landslides. The data count distribution of the feature GEOLOGY is depicted in Figure 16.

Figure 16: Count Distribution of Data of GEOLOGY

In this study, the above mentioned 15 features has been taken into consideration for building a robust explainable machine learning model that best corroborates the prediction of Landslide Susceptibility adopting the geotechnical engineering domain. The study’s features have also been analyzed in various state-of-the-art investigations, with a high level of correlation in terms of landslide susceptibility mapping. Before employing the state of the art machine learning algorithms, we have spitted our dataset into a training and testing ratio of 67:33 with stratified sampling strategy and randomization seed of 15 using Scikit-Learn library.

3.2 Evaluation Metrics

In our study, to evaluate the performance of our machine learning models, we considered the analysis of below mentioned popularly used evaluation metrics.

  • True Positive (TP): The case when the certain area is Landslide Susceptible and the model also classified as Landslide Susceptible.

  • False Positive (FP): The case when the certain area is not Landslide Susceptible but the model classified as Landslide Susceptible.

  • True Negative (TN): The case when the certain area is Not Landslide Susceptible and the model also classified as Not Landslide Susceptible.

  • False Negative (FN): The case when the certain area is Landslide Susceptible but the model classified as Not Landslide Susceptible.

  • Accuracy: It defines correctly classified areas with susceptible to Landslides. It can be computed as;

    (13)
  • Recall: It is defined as the ratio of the number of positive samples that have been correctly predicted as Landslide Susceptible corresponding to all Landslide Susceptible samples in the data. It can computed as;

    (14)
  • Precision: It is defined as the ratio of the number of positive samples that have been correctly predicted as Landslide Susceptible corresponding to all samples predicted as Landslide Susceptible. It can computed as;

    (15)
  • F1-Score: It is delineated as the term that balances between recall and precision. It can be defined as;

    (16)

3.3 Evaluation Stage 1: Optimization of Algorithms

In this section, we have optimized the performance of our Machine Learning classifiers using Grid Search method to find the best combination of Hyper-parameters for individual classifiers with 10-Fold Stratified Cross Validation to reduce the overfitting issue. We have employed the Scikit-Learn implementation in Python for SVM, KNN, Adaboost and Logistic Regression. And the Python Library of XgBoost for implementing the Extreme Gradient Boosting Algorithm. Grid Searching was performed by incorporating GridSearchCV method from popular Scikit-Learn library. In this stage of evaluation, all of the 15 feature variables have been employed for training and testing the individual classifiers. The best combinations of hyperparameters of the machine learning classifiers obtained through Grid Searching, are illustrated on the Table   5.

Classifier Hyperparameters
XgBoost
max_depth=3
n_estimators=3000
learning_rate=0.1
gamma=0
subample=0.7
KNN
n_neighbors =7
p = 1
Logistic Regression
C=0.01
solver=’l2’
penalty=’newton-cg’
SVM
C=10
kernel=rbf
Adaboost
n_esitmators=1000
learning_rate=1
Table 5: Best Combination of Hyperparameters

A graphical log-scale comparison of 10-Fold Stratified Cross-validation Scores with the obtained best hyper-parameters combination of the machine learning classifiers employed in our study is represented in the Figure  17.

Figure 17: Cross Validation Scores With All Features

The Confusion Matrix results of the machine learning classifiers on the test dataset is presented on the Table

6.

SVM KNN LR AdaBoost XgBoost
TP 57 60 58 57 59
FP 8 8 3 3 5
TN 57 57 62 60 60
FN 8 5 7 8 6
Table 6: Confusion Matrix Before Feature Reduction

Precision, Recall, F1-Score and Accuracy scores of the machine learning classifiers is presented on the Table  7.

Metrics SVM KNN LR AdaB XgBoost
Accuracy 87.69 90 92.31 90 91.54
Precision (N) 87.69 91.94 89.86 88.24 90.91
Recall (N) 87.69 87.69 95.38 92.31 92.31
F1 Score (N) 87.69 89.76 92.54 90.23 91.6
Precision (P) 87.69 88.24 95.08 91.94 92.19
Recall (P) 87.69 92.31 89.23 87.69 90.77
F1 Score (P) 87.69 90.23 92.06 89.76 91.47
Table 7: Performance Evaluation with All Features

In this context, based on the comparison of the performances of the individual machine learning classifiers, it is cleared that the optimized version of XgBoost(Extreme Gradient Boosting) is outperforming all other Machine Learning classifiers in evaluation criteria based on 10 Fold Stratified Cross Validation Mean F1 Weighted Scores. However, the findings on the test set demonstrate that the Logistic Regression model outperforms all other classifiers.But, as the logistic regression model does not outperform in terms of 10 Fold Cross Validation Scores, it may be deduced that the test dataset scores do not completely reflect the actual efficiency due to lack of the bulk of data samples. Apart from that, based on F1 Scores on test data samples, XgBoost is still showing the second best performance. Moreover, tree based gradient boosted methods have better explainability than other machine learning classifiers. Thus, in this evaluation stage, Optimized Extreme Gradient Boosting is identified as the most efficient algorithm to classify Landslide Susceptible area using the 15 feature variables from the utilized dataset.

3.4 Evaluation Stage 2: TreeSHAP Analysis

As optimized XgBoost is outperforming all other machine learning classifers besides it has robust explainability, we further decided to proceed with the optimzed XgBoost model for further analysis. In this stage of evaluation, we have integrated the TreeSHAP method with the optimized XgBoost (Extreme Gradient Boosting) Classifier(our best performing classifier from previous evaluation stage) to understand the prediction criteria of XgBoost and the level of contribution of individual features that strongly corroborate in Landslide Susceptibility prediction. Explainable Artificial Intelligence is an essential demand of present time as the complex and deep architectures of these machine learning models make it hard to interpret by engendering the BlackBox problem which is defined as not being able to detect where the model is actually looking or how much the individual features are contributing for a certain prediction. Considering the exigency and sensitivity involved in the study of Landslide Susceptibility, we interpreted the performance of XgBoost by analyzing the SHAP Feature Importance and SHAP Summary Plot.

3.4.1 SHAP Feature Importance

The SHAP Feature Importance plot illustrates the mean absolute shapley values for individual features. Figure  18 presents a graphical illustration of the mean absolute shapley values of all the features used for training the optimized Extreme Gradient Boosting (XgBoost) classifier before predicting Landslide Susceptibility. It is a bar plot where features are sorted in descending order based mean absolute SHAP values. Features with large absolute values like, SLOPE, ELEVATION, ROAD, TWI have significant impact on the prediction to support optimized XgBoost model for detecting Landslide Susceptible areas efficiently. On the other hand, FAULTLINES, PLAN, SPI, NDVI, LANDUSE have low absolute mean values indicating a very low influence in the model’s performance.

Figure 18: SHAP Feature Importance

3.4.2 SHAP Summary Plot

The SHAP Summary Plot illustrated at Figure  19 combines feature importance with feature effects for the optimized XgBoost model of Landslide Susceptibility prediction.

Figure 19: SHAP Summary Plot

In the following summary plot, each point is a Shapley value for a feature and a data instance. Here, in the graph, the vertical axis delineates features sorted according to the level of influence in the model’s prediction and horizontal axis delineates Shapley values indicating the positive or negative correlation of the feature of an instance with the target variable. To understand the the values of features graphically, a color based comparison is drawn. The overlapping dots are jittered in the vertical axis direction which give us an idea of the Shapley value distribution per feature variable. The features are ranked in order of significance. The following summary plot strongly illustrates; SLOPE, ROAD and TWI as the most significant features with wide spread distribution along the horizontal axis supporting efficient prediction for Landslide Susceptibility. In contrast, instances of PLAN, NDVI and LANDUSE and some other features are not well distributed over the horizontal axis and posses a Shapley value of near to 0, indicating a very less or almost no influence on the model’s performance for Landslide Susceptibility Prediction.

The study of features’ contributions to the assessment of Landslide Susceptible region derived from TreeSHAP analysis can be validated by considering evidence from state-of-the-art studies by geo-science researchers. Firstly, Çellek (2020) has emphasized the importance of measuring Slope Angle (SLOPE) for classfying Landslides with strong evidences. Consequently, the optimized XgBoost model proposed in our study is also utilizing the SLOPE feature as top most significant feature for prediction of Landslides. Furthermore, ELEVATION has a great correlation with SLOPE in the study of Landslides and helps researches to understand the vulnerability of an area towards Landslides (Rabby et al., 2021). This fact has also been reflected in our investigation of Landslides through TreeSHAP analysis. Additionally, TWI and ROAD are also top contributing factors for assessing the susceptibility of Landslides which have been reflected in our model building process and studies by several geotechnical researchers (Chen et al., 2017; Abedin et al., 2020; Rabby et al., 2021). Thus, the following analysis strengthen the hypothesis of this study, and supports the proper adoption of Geoscience and Geotechincal Engineering domain at every stage of our proposed research framework.

3.5 Evaluation Stage 3: Improved Landslide Feature Selection

In machine learning-based solutions, feature reduction is extremely important. An efficient model that can forecast Landslides with high accuracy using only a few characteristics is a gift in the domain of geotechnical engineering for Landslide Susceptibility Mapping. From the evaluation scores of previous stages, the optimized version of XgBoost seems to be the best fit for portending landslide susceptibility. So, in this stage of experimental setup, we have elected XgBoost as our proposed model for integration in landslide susceptibility mapping and performed feature reduction for XgBoost using the analysis of SHAP values. For this, we have retrained XgBoost with Grid Searching and eliminated 6 features including ASPECT, FAULTLINES, SPI, PLAN, NDVI and LANDUSE which has comparatively a very low influence in the prediction of XgBoost model’s performance according to TreeSHAP analysis. The overfall research methodology of proposed integrated optimized XgBoost model which can efficiently predict landslide susceptibility by using less number landslide causal factors is depicted in Figure 20.

Figure 20: The Complete Research Workflow for Building the Proposed Integrated XgBoost Model for Landslide Susceptibility Mapping using Less Number Landslide Causal Factors

Comparison of 10 Fold Stratified Cross Validation Mean Weighted F1 Scores before and after reduction of features illustrated in Figure 21.

Figure 21: Comparison of Cross Validation Scores of XgBoost with and without feature reduction

It is clearly visible that after feature reduction the Cross Validation score of optimized XgBoost has been increased from 94.62% to 95.01%.

Parameter Name Value
gamma 0
learning_rate 0.1
max_depth 3
n_estimators 1500
subsample 1
Table 8: Final Set of Hyperparameters of proposed optimized XgBoost model with Reduced Features

Table 8 delineates the final hyperparameter settings of the XgBoost model with reduced features.

Learning Curve helps us to determine whether or not increasing the amount of data is going to ameliorate the performance of model. The learning rate curve of XgBoost classifier from Figure  22 clearly depicts that there are room for improvement in the performance with more data in case of the optimized XgBoost model with reduced features. The point where the two lines in the graphs will converge, after that particular point the room for improvement with more data will be narrowed down.

Figure 22: Learning Curve of XgBoost
Figure 23: ROC Curve of XgBoost

4 Discussion

In this study, after a rigorous analysis of machine learning classifiers for predicting Landslides in an automated manner, it is found that, optimized version of XgBoost classifier is a promising method which successfully adopts geotechnical engineering domain based on the outperforming results of Table 7 and Figure 17. Considering the massive potential of XgBoost model, this study further extended the research to explain the prediction criterion of XgBoost model by integrating TreeSHAP analysis in a motivation of solving the blackbox problem with Explainable Artificial Intelligence. According to the TreeSHAP analysis, 6 features were eliminated from the feature set and the XgBoost model was retrained along with exhaustive hyperparameter tuning. Here, after feature elimination, the 10 Fold Cross Validation Weighted F1 Score of XgBoost model has increased to 95.01%. It corroborates that feature reduction has been beneficial for the XgBoost model. Moreover, the optimized XgBoost model can generate outperforming results with even less features.

Reduction of features is a crucial contribution presented in this study that open doors for geoscience researchers to conduct effective studies on Landslides utilizing less features. The features that have been eliminated in the final model building process in this study can reduce the cost of the industry level works on the domain of Landslide Susceptibility prediction. For example, to compute the features like LANDUSE and NDVI, high-tech remote and satellite based technologies are required which increases the cost of the investigation and demands more time and technological manpower. In the primary stage, if the susceptibility of Landslide can be predicted without the need of these features which have high cost for computation purpose it is going to be great help for geotechincal researchers for both industry and academia. The optimized version of Xgboost with proposed architecture is outperforming other classifiers and previous settings of model even after eliminating the above mentioned 6 features. Due to the elimination of these 6 features, it would be possible to save more cost and time along with better utilization of the existing technological resources involved in Landslide study.

Comparatively, Optimized Extreme Gradient Boosting has outperformed all other machine learning classifiers even with reduction of 40% of the features with 10 Fold Stratified Cross-Validation Weighted F1 Scores of 95.01% and ROC-AUC score of 97%. It has also helped us to identify eloquent features that strongly corroborates in the prediction of Landslide Susceptibility. In this context, we prefer to adopt the optimized version of Extreme Gradient Boosting (XgBoost) with the identified eloquent features for Landslide Susceptibility Prediction using data-driven analysis.

Figure 24: AUC Score Comparison with Other State of Art Studies

From Figure 24, in comparison with the recent state of the art studies for Landslide Susceptibility Prediction using Artificial Intelligence based methods, the optimized XgBoost model with Reduced Features that outlined in our study, has outperformed the best model from the study of Rabby et al. (2021) by almost 6% , Mandal et al. (2021) by almost 4%, Sahin et al. (2020) by almost 8% and Huang et al. (2020b) by almost 10% in terms of AUC scores.

5 Conclusion

Landslides are another name of nightmare for the people living in the hilly areas from all over the world which causes a huge number of causalities every year exacerbating socio-economic condition along with unwanted death tolls. An early prediction of Landslide Susceptibility can be a great blessing for mankind. Realizing this exigency, we have performed extensive analysis to predict Landslide Susceptible area with state of the art Artificial Intelligence based methods. In our study, we have successfully identified the eloquent features which are most useful for an automatic and early prediction Landslide Susceptibility with Machine Learning algorithms, through the integration of Explainable and Interpretable Artificial Intelligence, a milestone achieved in the study of Landslides and Geo-Technical science. Besides, due to the optimization of models and even with reduction of features our best performing model has outperformed the recent similar state of art studies and the previous study with similar dataset also. Moreover, this study also highlights that machine learning based feature selection is more suitable than the statistical analysis based feature selection in domain of Landslide study. The proposed model of this study, is a cost effective solution for the assessment and early prediction of Landslides at the primary investigation with less number of geological and topological features. Reduction of geological, topological and hydrological features would allow geotechincal researchers to conduct more economic and efficient research in less time. In future, we would like to collect more empirical data of Landslides and integrate state of art deep architectures including Generative Adversarial Networks (GANs) for more efficient and early prediction of Landslides to address the problem of Landslide Susceptibility Mapping for saving mankind from sudden catastrophes.

Acknowledgment

We want to thank Dr. Fahim Irfan Alam, Ph.D. (Associate Professor, University of Chittagong), expert in Machine Learning for Remote Sensing Applications and Hypersectral Imaging, for valuable discussions and intellectual reviews to make the research outperforming.

References

  • Mind’je et al. (2020) Richard Mind’je, Lanhai Li, Jean Baptiste Nsengiyumva, Christophe Mupenzi, Enan Muhire Nyesheja, Patient Mindje Kayumba, Aboubakar Gasirabo, and Egide Hakorimana. Landslide susceptibility and influencing factors analysis in rwanda. Environment, Development and Sustainability, 22(8):7985–8012, Dec 2020. ISSN 1573-2975. doi: 10.1007/s10668-019-00557-4. URL https://doi.org/10.1007/s10668-019-00557-4.
  • Survey (2021) U.S. Geological Survey. How many deaths result from landslides each year?, 2021. URL https://www.usgs.gov/faqs/how-many-deaths-result-landslides-each-year?qt-news_science_products=0#qt-news_science_products.
  • Sultana (2020) Neegar Sultana. Analysis of landslide-induced fatalities and injuries in bangladesh: 2000-2018. Cogent Social Sciences, 6(1):1737402, 2020. doi: 10.1080/23311886.2020.1737402. URL https://doi.org/10.1080/23311886.2020.1737402.
  • Winter et al. (2016) Mike G. Winter, Barbara Shearer, Derek Palmer, David Peeling, Clare Harmer, and Jonathan Sharpe. The economic impact of landslides and floods on the road network. Procedia Engineer, 143:1425–1434, 2016. ISSN 1877-7058. doi: https://doi.org/10.1016/j.proeng.2016.06.168. URL https://www.sciencedirect.com/science/article/pii/S1877705816306154. Advances in Transportation Geotechnics III.
  • Perera et al. (2018) E. N. C. Perera, D. T. Jayawardana, P. Jayasinghe, R. M. S. Bandara, and N. Alahakoon. Direct impacts of landslides on socio-economic systems: a case study from aranayake, sri lanka. Geoenvironmental Disasters, 5(1):11, Aug 2018. doi: 10.1186/s40677-018-0104-6. URL https://doi.org/10.1186/s40677-018-0104-6.
  • FAO (2021) United Nations FAO. Landslides : FAO in Emergencies, 2021. URL http://www.fao.org/emergencies/emergency-types/landslides/en/.
  • Huang et al. (2020a) Faming Huang, Jing Zhang, Chuangbing Zhou, Yuhao Wang, Jinsong Huang, and Li Zhu. A deep learning algorithm using a fully connected sparse autoencoder neural network for landslide susceptibility prediction. Landslides, 17(1):217–229, Jan 2020a. ISSN 1612-5118. doi: 10.1007/s10346-019-01274-9. URL https://doi.org/10.1007/s10346-019-01274-9.
  • Sahin et al. (2020) Emrehan Kutlug Sahin, Ismail Colkesen, Suheda Semih Acmali, Aykut Akgun, and Arif Cagdas Aydinoglu. Developing comprehensive geocomputation tools for landslide susceptibility mapping: Lsm tool pack. Comput. Geosci., 144:104592, 2020. ISSN 0098-3004. doi: https://doi.org/10.1016/j.cageo.2020.104592. URL https://www.sciencedirect.com/science/article/pii/S009830042030577X.
  • Hong et al. (2019) Haoyuan Hong, Junzhi Liu, and A-Xing Zhu. Landslide susceptibility evaluating using artificial intelligence method in the youfang district (china). Environ Earth Sci, 78(15):488, Aug 2019. ISSN 1866-6299. doi: 10.1007/s12665-019-8415-9. URL https://doi.org/10.1007/s12665-019-8415-9.
  • Sahana et al. (2020) Mehebub Sahana, Binh Thai Pham, Manas Shukla, Romulus Costache, Do Xuan Thu, Rabin Chakrabortty, Neelima Satyam, Huu Duy Nguyen, Tran Van Phong, Hiep Van Le, Subodh Chandra Pal, G. Areendran, Kashif Imdad, and Indra Prakash. Rainfall induced landslide susceptibility mapping using novel hybrid soft computing methods based on multi-layer perceptron neural network classifier. Geocarto International, 0(0):1–25, 2020. doi: 10.1080/10106049.2020.1837262. URL https://doi.org/10.1080/10106049.2020.1837262.
  • Pham et al. (2017) Binh Thai Pham, Dieu Tien Bui, Hamid Reza Pourghasemi, Prakash Indra, and M. B. Dholakia. Landslide susceptibility assesssment in the uttarakhand area (india) using gis: a comparison study of prediction capability of naïve bayes, multilayer perceptron neural networks, and functional trees methods. Theor. Appl. Climatol., 128(1):255–273, Apr 2017. ISSN 1434-4483. doi: 10.1007/s00704-015-1702-9. URL https://doi.org/10.1007/s00704-015-1702-9.
  • Li et al. (2019) Deying Li, Faming Huang, Liangxuan Yan, Zhongshan Cao, Jiawu Chen, and Zhou Ye. Landslide susceptibility prediction using particle-swarm-optimized multilayer perceptron: Comparisons with multilayer-perceptron-only, bp neural network, and information value models. Applied Sciences, 9(18), 2019. ISSN 2076-3417. doi: 10.3390/app9183664. URL https://www.mdpi.com/2076-3417/9/18/3664.
  • Dao et al. (2020) Dong Van Dao, Abolfazl Jaafari, Mahmoud Bayat, Davood Mafi-Gholami, Chongchong Qi, Hossein Moayedi, Tran Van Phong, Hai-Bang Ly, Tien-Thinh Le, Phan Trong Trinh, Chinh Luu, Nguyen Kim Quoc, Bui Nhi Thanh, and Binh Thai Pham. A spatially explicit deep learning neural network model for the prediction of landslide susceptibility. CATENA, 188:104451, 2020. ISSN 0341-8162. doi: https://doi.org/10.1016/j.catena.2019.104451. URL https://www.sciencedirect.com/science/article/pii/S0341816219305934.
  • Bui et al. (2020) Dieu Tien Bui, Paraskevas Tsangaratos, Viet-Tien Nguyen, Ngo Van Liem, and Phan Trong Trinh. Comparing the prediction performance of a deep learning neural network model with conventional machine learning models in landslide susceptibility assessment. CATENA, 188:104426, 2020. ISSN 0341-8162. doi: https://doi.org/10.1016/j.catena.2019.104426. URL https://www.sciencedirect.com/science/article/pii/S0341816219305685.
  • Zhu et al. (2020) Li Zhu, Lianghao Huang, Linyu Fan, Jinsong Huang, Faming Huang, Jiawu Chen, Zihe Zhang, and Yuhao Wang.

    Landslide susceptibility prediction modeling based on remote sensing and a novel deep learning algorithm of a cascade-parallel recurrent neural network.

    Sens., 20(6), 2020. ISSN 1424-8220. doi: 10.3390/s20061576. URL https://www.mdpi.com/1424-8220/20/6/1576.
  • Pham et al. (2019) Binh Thai Pham, Indra Prakash, Sushant K. Singh, Ataollah Shirzadi, Himan Shahabi, Thi-Thu-Trang Tran, and Dieu Tien Bui. Landslide susceptibility modeling using reduced error pruning trees and different ensemble techniques: Hybrid machine learning approaches. CATENA, 175:203–218, 2019. ISSN 0341-8162. doi: https://doi.org/10.1016/j.catena.2018.12.018. URL https://www.sciencedirect.com/science/article/pii/S0341816218305538.
  • Thai Pham et al. (2019) Binh Thai Pham, Ataollah Shirzadi, Himan Shahabi, Ebrahim Omidvar, Sushant K. Singh, Mehebub Sahana, Dawood Talebpour Asl, Baharin Bin Ahmad, Nguyen Kim Quoc, and Saro Lee. Landslide susceptibility assessment by novel hybrid machine learning algorithms. Sustainability, 11(16), 2019. ISSN 2071-1050. doi: 10.3390/su11164386. URL https://www.mdpi.com/2071-1050/11/16/4386.
  • Huang et al. (2020b) Faming Huang, Zhongshan Cao, Jianfei Guo, Shui-Hua Jiang, Shu Li, and Zizheng Guo.

    Comparisons of heuristic, general statistical and machine learning models for landslide susceptibility prediction and mapping.

    CATENA, 191:104580, 2020b. ISSN 0341-8162. doi: https://doi.org/10.1016/j.catena.2020.104580. URL https://www.sciencedirect.com/science/article/pii/S0341816220301302.
  • Fang et al. (2021) Zhice Fang, Yi Wang, Ling Peng, and Haoyuan Hong. A comparative study of heterogeneous ensemble-learning techniques for landslide susceptibility mapping. Int J Geogr Inf Sci, 35(2):321–347, 2021. doi: 10.1080/13658816.2020.1808897. URL https://doi.org/10.1080/13658816.2020.1808897.
  • Sahin (2020) Emrehan Kutlug Sahin. Comparative analysis of gradient boosting algorithms for landslide susceptibility mapping. Geocarto International, 0(0):1–25, 2020. doi: 10.1080/10106049.2020.1831623. URL https://doi.org/10.1080/10106049.2020.1831623.
  • Merghadi et al. (2020) Abdelaziz Merghadi, Ali P. Yunus, Jie Dou, Jim Whiteley, Binh ThaiPham, Dieu Tien Bui, Ram Avtar, and Boumezbeur Abderrahmane. Machine learning methods for landslide susceptibility studies: A comparative overview of algorithm performance. Earth Sci. Rev., 207:103225, 2020. ISSN 0012-8252. doi: https://doi.org/10.1016/j.earscirev.2020.103225. URL https://www.sciencedirect.com/science/article/pii/S0012825220302713.
  • Zhou et al. (2021) Xinzhi Zhou, Haijia Wen, Yalan Zhang, Jiahui Xu, and Wengang Zhang. Landslide susceptibility mapping using hybrid random forest with geodetector and rfe for factor optimization. Geosci. Front., 12(5):101211, 2021. ISSN 1674-9871. doi: https://doi.org/10.1016/j.gsf.2021.101211. URL https://www.sciencedirect.com/science/article/pii/S167498712100075X.
  • Rabby et al. (2021) Yasin Wahid Rabby, Md Belal Hossain, and Joynal Abedin. Landslide susceptibility mapping in three upazilas of rangamati hill district bangladesh: application and comparison of gis-based machine learning methods. Geocarto International, 0(0):1–27, 2021. doi: 10.1080/10106049.2020.1864026. URL https://doi.org/10.1080/10106049.2020.1864026.
  • Chen et al. (2018) Wei Chen, Hamid Reza Pourghasemi, and Seyed Amir Naghibi. A comparative study of landslide susceptibility maps produced using support vector machine with different kernel functions and entropy data mining models in china. Bull. Eng. Geol. Env., 77(2):647–664, May 2018. ISSN 1435-9537. doi: 10.1007/s10064-017-1010-y. URL https://doi.org/10.1007/s10064-017-1010-y.
  • Kalantar et al. (2018) Bahareh Kalantar, Biswajeet Pradhan, Seyed Amir Naghibi, Alireza Motevalli, and Shattri Mansor. Assessment of the effects of training data selection on the landslide susceptibility mapping: a comparison between support vector machine (svm), logistic regression (lr) and artificial neural networks (ann). Geomatics, Natural Hazards and Risk, 9(1):49–69, 2018. doi: 10.1080/19475705.2017.1407368. URL https://doi.org/10.1080/19475705.2017.1407368.
  • Guo et al. (2021) Zizheng Guo, Yu Shi, Faming Huang, Xuanmei Fan, and Jinsong Huang. Landslide susceptibility zonation method based on c5.0 decision tree and k-means cluster algorithms to improve the efficiency of risk management. Geosci. Front., page 101249, 2021. ISSN 1674-9871. doi: https://doi.org/10.1016/j.gsf.2021.101249. URL https://www.sciencedirect.com/science/article/pii/S1674987121001134.
  • Yap and Sim (2011) B. W. Yap and C. H. Sim. Comparisons of various types of normality tests. J. Stat. Comput. Simul., 81(12):2141–2155, 2011. doi: 10.1080/00949655.2010.520163. URL https://doi.org/10.1080/00949655.2010.520163.
  • Royston (1992) Patrick Royston. Approximating the shapiro-wilk w-test for non-normality. Stat. Comput., 2(3):117–119, Sep 1992. ISSN 1573-1375. doi: 10.1007/BF01891203. URL https://doi.org/10.1007/BF01891203.
  • Singhal and Rana (2015) Richa Singhal and Rakesh Rana. Chi-square test and its application in hypothesis testing. Journal of the Practice of Cardiovascular Sciences, 1, 01 2015. doi: 10.4103/2395-5414.157577.
  • McHugh (2013) Mary L McHugh. The chi-square test of independence. Biochemia medica, 23(2):143–149, 2013.
  • Huang and Zhao (2018) Yu Huang and Lu Zhao. Review on landslide susceptibility mapping using support vector machines. CATENA, 165:520–529, 2018. ISSN 0341-8162. doi: https://doi.org/10.1016/j.catena.2018.03.003. URL https://www.sciencedirect.com/science/article/pii/S0341816218300791.
  • Lombardo and Mai (2018) Luigi Lombardo and P. Martin Mai. Presenting logistic regression-based landslide susceptibility results. Eng. Geol., 244:14–24, 2018. ISSN 0013-7952. doi: https://doi.org/10.1016/j.enggeo.2018.07.019. URL https://www.sciencedirect.com/science/article/pii/S0013795218301212.
  • Guo et al. (2003) Gongde Guo, Hui Wang, David Bell, Yaxin Bi, and Kieran Greer. KNN Model-Based Approach in Classification. In Robert Meersman, Zahir Tari, and Douglas C. Schmidt, editors, On The Move to Meaningful Internet Systems 2003: CoopIS, DOA, and ODBASE, Lecture Notes in Computer Science, pages 986–996, Berlin, Heidelberg, 2003. Springer. ISBN 9783540399643. doi: 10.1007/978-3-540-39964-3_62.
  • Freund and Schapire (1997) Yoav Freund and Robert E Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. J. Comput. System Sci., 55(1):119–139, 1997. ISSN 0022-0000. doi: https://doi.org/10.1006/jcss.1997.1504. URL https://www.sciencedirect.com/science/article/pii/S002200009791504X.
  • Schapire (2013) Robert E. Schapire. Explaining AdaBoost. In Bernhard Schölkopf, Zhiyuan Luo, and Vladimir Vovk, editors, Empirical Inference: Festschrift in Honor of Vladimir N. Vapnik, pages 37–52. Springer, Berlin, Heidelberg, 2013. ISBN 9783642411366. doi: 10.1007/978-3-642-41136-6_5. URL https://doi.org/10.1007/978-3-642-41136-6_5.
  • Chen and Guestrin (2016) Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining, pages 785–794, 2016.
  • Khan Inan et al. (2021) Muhammad Sakib Khan Inan, Rubaiath E Ulfath, Fahim Irfan Alam, Fateha Khanam Bappee, and Rizwan Hasan. Improved sampling and feature selection to support extreme gradient boosting for pcos diagnosis. In 2021 IEEE 11th Annual Computing and Communication Workshop and Conference (CCWC), pages 1046–1050, 2021. doi: 10.1109/CCWC51732.2021.9375994.
  • Lundberg and Lee (2017) Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Adv Neur In, 30:4765–4774, 2017.
  • Lundberg et al. (2018) Scott M Lundberg, Gabriel G Erion, and Su-In Lee. Consistent individualized feature attribution for tree ensembles. arXiv preprint arXiv:1802.03888, 2018.
  • Molnar (2021) Christoph Molnar. Interpretable machine learning, Jun 2021. URL https://christophm.github.io/interpretable-ml-book/shapley.html#shapley.
  • Abedin et al. (2020) Joynal Abedin, Yasin Wahid Rabby, Ikramul Hasan, and Humaira Akter. An investigation of the characteristics, causes, and consequences of june 13, 2017, landslides in rangamati district bangladesh. Geoenvironmental Disasters, 7(1):23, Aug 2020. ISSN 2197-8670. doi: 10.1186/s40677-020-00161-z. URL https://doi.org/10.1186/s40677-020-00161-z.
  • Carson and Kirkby (1972) Michael A Carson and Michael J Kirkby. Hillslope form and process. 1972.
  • Meten et al. (2015) Matebie Meten, Netra PrakashBhandary, and Ryuichi Yatabe. Effect of landslide factor combinations on the prediction accuracy of landslide susceptibility maps in the blue nile gorge of central ethiopia. Geoenvironmental Disasters, 2(1):9, Mar 2015. ISSN 2197-8670. doi: 10.1186/s40677-015-0016-7. URL https://doi.org/10.1186/s40677-015-0016-7.
  • Ohlmacher (2007) Gregory C Ohlmacher. Plan curvature and landslide probability in regions dominated by earth flows and earth slides. Eng. Geol., 91(2-4):117–134, 2007.
  • Sealey et al. (2018) Kathleen Sullivan Sealey, P.-M. Binder, and R. King Burch. Financial credit drives urban land-use change in the united states. Anthropocene, 21:42–51, 2018. ISSN 2213-3054. doi: https://doi.org/10.1016/j.ancene.2018.01.002. URL https://www.sciencedirect.com/science/article/pii/S2213305417300164.
  • Promper et al. (2014) C. Promper, A. Puissant, J.-P. Malet, and T. Glade. Analysis of land cover changes in the past and the future as contribution to landslide risk scenarios. Appl. Geogr., 53:11–19, 2014. ISSN 0143-6228. doi: https://doi.org/10.1016/j.apgeog.2014.05.020. URL https://www.sciencedirect.com/science/article/pii/S0143622814001155.
  • Restrepo and Alvarez (2006) Carla Restrepo and Nora Alvarez. Landslides and their contribution to land-cover change in the mountains of mexico and central america 1. Biotropica, 38(4):446–457, 2006.
  • Reichenbach et al. (2014) P. Reichenbach, C. Busca, A. C. Mondini, and M. Rossi. The influence of land use change on landslide susceptibility zonation: The briga catchment test site (messina, italy). Environ. Manage., 54(6):1372–1384, Dec 2014. ISSN 1432-1009. doi: 10.1007/s00267-014-0357-0. URL https://doi.org/10.1007/s00267-014-0357-0.
  • Chen et al. (2019) L. Chen, Z. Guo, K. Yin, D. P. Shrestha, and S. Jin. The influence of land use and land cover change on landslide susceptibility: a case study in zhushan town, xuan’en county (hubei, china). Nat Hazard Earth Sys, 19(10):2207–2228, 2019. doi: 10.5194/nhess-19-2207-2019. URL https://nhess.copernicus.org/articles/19/2207/2019/.
  • Dou et al. (2015) Jie Dou, Dieu Tien Bui, Ali P. Yunus, Kun Jia, Xuan Song, Inge Revhaug, Huan Xia, and Zhongfan Zhu. Optimization of causative factors for landslide susceptibility evaluation using remote sensing and gis data in parts of niigata, japan. PloS one, 10(7):e0133262, 2015.
  • Whitworth et al. (2011) Malcolm Whitworth, Ian Anderson, and Graham Hunter. Chapter seventeen - geomorphological assessment of complex landslide systems using field reconnaissance and terrestrial laser scanning. In Mike J. Smith, Paolo Paron, and James S. Griffiths, editors, Geomorphological Mapping, volume 15 of Developments in Earth Surface Processes, pages 459–474. Elsevier, 2011. doi: https://doi.org/10.1016/B978-0-444-53446-0.00017-3. URL https://www.sciencedirect.com/science/article/pii/B9780444534460000173.
  • Tanoli et al. (2017) Javed Iqbal Tanoli, Chen Ningsheng, Amar Deep Regmi, and Li Jun. Spatial distribution analysis and susceptibility mapping of landslides triggered before and after mw7. 8 gorkha earthquake along upper bhote koshi, nepal. Arabian J. Geosci., 10(13):1–24, 2017.
  • Mattivi et al. (2019) Pietro Mattivi, Francesca Franci, Alessandro Lambertini, and Gabriele Bitelli.

    Twi computation: a comparison of different open source giss.

    Open Geospatial Data, Software and Standards, 4(1):6, Jul 2019. ISSN 2363-7501. doi: 10.1186/s40965-019-0066-y. URL https://doi.org/10.1186/s40965-019-0066-y.
  • Dahigamuwa et al. (2016) Thilanki Dahigamuwa, Qiuyan Yu, and Manjriker Gunaratne. Feasibility study of land cover classification based on normalized difference vegetation index for landslide risk assessment. Geosciences, 6(4), 2016. ISSN 2076-3263. doi: 10.3390/geosciences6040045. URL https://www.mdpi.com/2076-3263/6/4/45.
  • Chen et al. (2017) Chi-Wen Chen, Hongey Chen, Lun-Wei Wei, Guan-Wei Lin, Tomoyuki Iida, and Ryuji Yamada. Evaluating the susceptibility of landslide landforms in japan using slope stability analysis: a case study of the 2016 kumamoto earthquake. Landslides, 14(5):1793–1801, 2017.
  • Regmi et al. (2014) Netra R. Regmi, John R. Giardino, and John D. Vitek. Characteristics of landslides in western colorado, usa. Landslides, 11(4):589–603, Aug 2014. ISSN 1612-5118. doi: 10.1007/s10346-013-0412-6. URL https://doi.org/10.1007/s10346-013-0412-6.
  • Ayalew and Yamagishi (2005) Lulseged Ayalew and Hiromitsu Yamagishi. The application of gis-based logistic regression for landslide susceptibility mapping in the kakuda-yahiko mountains, central japan. Geomorphology, 65(1-2):15–31, 2005.
  • Çellek (2020) Seda Çellek. Effect of the slope angle and its classification on landslide. Nat. Hazards Earth Syst. Sci. Discuss., pages 1–23, 2020.
  • Mandal et al. (2021) Kanu Mandal, Sunil Saha, and Sujit Mandal. Applying deep learning and benchmark machine learning algorithms for landslide susceptibility modelling in rorachu river basin of sikkim himalaya, india. Geosci. Front., 12(5):101203, 2021. ISSN 1674-9871. doi: https://doi.org/10.1016/j.gsf.2021.101203. URL https://www.sciencedirect.com/science/article/pii/S1674987121000670.