Predicting the Landslide Susceptibility in Eastern Sikkim Himalayan Region, India Using Boosted Regression Tree and REPTree Machine Learning Techniques

Abstract

AbstractIn the mountainous parts of the world, landslides are considered as the most dangerous to people and property. The number and the amount of damage caused by landslides have been steadily growing globally. As a result, slope instability management is critical in the hilly region’s ecological and socioeconomic dynamics. The Rorachu river basin of Sikkim Himalaya has been selected for the present study which is Sikkim’s most landslide-prone area. The key intention of the research is to use computer-based machine learning techniques to create landslide susceptibility maps (LSMs) and compare the models’ efficiency. Nineteen variables, including stimulating and environmental factors, were chosen to better explain the current spatial relationship with the landslide. Two popular machine learning techniques, i.e., Reduced Error Pruning Tree (REPTree) and Boosted Regression Tree (BRTree) have been incorporated to prepare LSMs. Randomly selected landslide and non-landslide points were used to build two different databases: training data and evaluation data. During the collection of both training and testing sites, a 70:30 ratio was retained. Tolerance (TOL) and Variance Inflation Factor (VIF) was used to estimate multicollinearity, and Information Gain Ratio (IGR) was taken to evaluate the importance of the variables. The findings show that multicollinearity is minimum in the landslide causing factors, and rainfall is perhaps the most crucial component in landslide occurrence. Receiver Operating Characteristics curve (ROC) coupled with statistical techniques like Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) were exercised to portray the models’ accuracy for both training and testing datasets. The result enlightens that higher landslide susceptibility classes contain most of the slide area. In both the training and testing datasets. The BRTree model has the highest Area Under Curve (AUC) value. The BRTree model’s AUC values for training and testing datasets were found to be 0.896 and 0.9, respectively. The result of RMSE and MAE also hails the superior representation of the REPTree model. The RMSE and MAE value of the REPTree model was noticed 0.120, 0.189 (for training dataset), and 0.129, 0.192 (for validation dataset), while the value was obtained 0.150, 0.235 (for training dataset) and 0.139, 0.21 (for validation dataset) in BRTree model respectively. As a result, both models performed admirably, but the BRTree model performed better than the REPTree model.KeywordsReduced Error Pruning Tree (REPTree)Boosted Regression Tree (BRTree)Landslide susceptibilityROC curveRorachu river basin