Probabilistic approach to Predict Landslide Susceptibility based on Dynamic Parameters for Uttarkashi, Uttarakhand (India)

Abstract

The changing climate and global warming affect the stability of slopes, resulting in landslides. Landslides are frequent in hilly regions all over the world. The present work compares three GIS-based machine learning techniques to predict the changes in landslide susceptibility patterns classified as low, moderate, and high from observed records. The state-of-the-art methods include Random Forest (RF), Support Vector Machine (SVM), and Multinomial Logistic Regression (MLR). The landslide inventory contains a total of 1239 locations, which are divided into three subsets for training, testing, and validation purposes. A total of seven influencing factors were selected to understand the relationship between selected factors and observed landslides. The models were compared using the Receiver Operating Characteristics (ROC) curve and other statistical measures, including accuracy, precision, recall, sensitivity, and specificity. The RF model outperformed with the highest training (RF_Accuracy=91%), testing (RF_Accuracy=88%), and validation (RF_Accuracy=86%) accuracy. The ROC values computed for the validation dataset for three models are 0.749, 0.734, and 0.874 for the MLR, SVM, and RF models respectively. The outcome of the present study could be instrumental for policy and decision-makers concerning risk planning and mitigation.