Demystifying the predictive capability of advanced heterogeneous machine learning ensembles for landslide susceptibility assessment and mapping in the Eastern Himalayan Region, India

• Published in: Natural hazards (Dordrecht) Vol. 121; no. 11; pp. 13407 - 13446
• Main Authors: Dey, Sumon, Das, Swarup, Roy, Sujit Kumar
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 01.06.2025
• Subjects: Civil Engineering; Earth and Environmental Science; Earth Sciences; Environmental Management; Geophysics/Geodesy; Geotechnical Engineering & Applied Earth Sciences; Hydrogeology; Natural Hazards; Original Paper; Explainable artificial intelligence; Monotone multilayer perceptron; Bagged AdaBoost; Landslide susceptibility; Model averaged neural network; Boruta algorithm
• ISSN: 0921-030X; 1573-0840
• DOI: 10.1007/s11069-025-07325-w

Landslides are severe and frequent natural disasters that can cause significant adverse impacts on human lives and infrastructure, especially in mountainous areas around the world. Accurate susceptibility assessment and zonation are essential for disaster risk reduction and sustainable development, particularly in the Darjeeling district of West Bengal, India. This study explores the potential of unconventional machine learning techniques, such as Model-Averaged Neural Networks (MA-NNET), Bagged AdaBoost (ADABAG), and Monotone Multilayer Perceptron (MONMLP), which extend beyond standard methodology. To conduct a comprehensive investigation of landslide susceptibility, the study uses an extensive dataset of geological, topographical, hydrological, and anthropogenic factors. The landslide causal factors (LCFs) were selected through correlation analysis, multicollinearity assessment, and the Boruta algorithm. The landslide susceptibility maps constructed through the aforementioned machine-learning methods were validated using the area under the receiver operating characteristic (AUC-ROC) curve and other performance metrics such as precision, specificity, recall (sensitivity), F1-score, overall accuracy, kappa index, and balanced accuracy. According to the evaluation metrics, the ADABAG model outperforms the MA-NNET (training AUC 90.8%, testing AUC 89.8%) and MONMLP (training AUC 91.7%, testing AUC 90.08%) models, with an AUC-ROC score of 93.4% (training) and 91.3% (testing). The present study not only contributes to the advancement of landslide susceptibility assessment but also provides valuable insights into the capabilities of the employed models and their ability to predict outcomes in a geographic environment, as demonstrated by the partial dependence profile and Shapley waterfall model, a sophisticated explainable artificial intelligence technique. This research demonstrates a noteworthy capacity to substantially influence urban land use planning, disaster management, and mitigation policies in landslide-prone regions of the Darjeeling district.