Exploring novel super-learner-based machine learning ensembles for landslide susceptibility prediction with integrated uncertainty quantification

Western Ghats, one of the mystic and oldest mountain ranges in Southern India, historically experienced considerable number of landslides. The present study was objectified to quantify the susceptibility of landslides for Idukki district in Kerala (India), a prominent part of the Western Ghats. Fifteen geo-environmental features were taken up for the research after extensive literature survey and feature-selection technique. 1781 historical landslide locations were found through heterogeneous sources. Four advanced machine learning techniques, namely, Flexible Discriminant Analysis (FDA), Monotone Multi-Layer Perceptron (MONMLP), Model-Averaged Neural Network (MANN), Boosted Decision Tree (BDT) were selected as base-learners with the Gradient Boosting Machines as the Super-learner. The constructed models were found to perform extraordinarily with astounding AUCROC score of 0.975, 0.970, 0.965 and 0.959 with the training data, and 0.946, 0.944, 0.941 and 0.936 with the testing data for GBM-MONMLP, GBM-BDT, GBM-FDA and GBM-MANN model respectively. Taylor’s diagram was used for visual interpretation of the model performance. The models were also validated with Quality Sum (Q’s) index. The maximum Q’s index value of 3.06 was found for GBM-MONMLP followed by 2.94 (GBM-BDT), 2.71 (GBM-MANN) and 2.63 (GBM-FDA). Model calibration score substantiated excellent performance of the models. For enhanced scientific trustworthiness, uncertainty quantification was performed through coefficient-of-variation (COV) and a confidence map for each one of the susceptibility maps was constructed through matrix-intersection method. The findings of the research with fused certainty may be helpful for decision-makers of the region for doing the needful to prioritize safeguarding of human livelihood and reduce the economic depletion.