Landslide susceptibility modelling using hybrid bivariate statistical-based machine-learning method in a highland segment of Southern Western Ghats, India

The unique physiographic conditions, abundant rainfall, and anthropogenic disturbances in the mountainous Western Ghats (WG) region often cause landslides, which leads to loss of life and properties. The present study aims to generate a landslide susceptibility map for a mountainous WG region using novel hybrid methods. For this, a landslide inventory database is prepared with 82 landslide locations and twelve landslide contributing factors, viz., lithology, geomorphology, slope angle, soil texture, relative relief, distance from the streams, distance from the roads, distance from the lineaments, topographic wetness index (TWI), rainfall, land use/land cover, and slope curvature were used for modelling. Initially, four bivariate approaches, viz., information value (InfoV), evidential belief function (EBF), certainty factor (CF), and index of entropy methods (IoE), were applied to model landslide susceptibility, and the accuracy was estimated. Thereafter, these models were integrated with random forest (RF) to build hybrid models. The computed hybrid models were evaluated using ROC-AUC and confusion matrix-based statistical measures. Hybrid models show higher accuracy with ROC-AUC values ranging from 0.894 to 0.928 in the training phase and 0.906 to 0.925 in the testing phase. The RF-InfoV model was proven to be comparably efficient for landslide susceptibility analysis.