Integrating deep learning neural network and M5P with conventional statistical models for landslide susceptibility modelling

Landslides are among the devastating geological hazards that cause immense damage in hilly regions. The Indian Himalayan region is plagued by numerous major landslides. Here we present results of landslide susceptibility mapping in a representative area of Jakholi region of Indian Himalaya based on a novel approach of integrating deep learning neural network (DLNN) and M5 Prime (M5P) with the conventional statistical model weight of evidence (WoE). These models were trained using 70% of inventory landslides and evaluated using 30%, considering 14 factors divided into topographical, hydrological, geological, and environmental landslide conditioning factors (LCFs). The appropriateness of factors was judged using the multicollinearity test, and the WoE model was used to evaluate the degree of association between LCFs and landslide occurrences. Precision, accuracy, Kappa coefficient, root mean square error (RMSE), and area under the receiver operating characteristic (AUC-ROC) curve were used to assess the models’ efficiency. The proposed WoE, M5P, WoE-M5P, DLNN, and WoE-DLNN models quantified 16.68%, 19.48%, 18.98%, 19.68%, and 16.67% of the area as very highly landslide-prone region, respectively. Ensemble of WoE-DLNN model with 89.96% success rate and 92.51% prediction rate in terms of AUC outperformed the WoE-M5P and individual models. It was also found that the DLNN alone performs more efficiently than WoE-M5P. Our study thus reveals that enhanced performance of WoE-DLNN ensemble could be used in other regions with similar geo-environmental settings. The results from this study can be of potential use to regional planners and governmental agencies in formulating effective landslide management plans.