Ensemble deep learning framework for landslide susceptibility mapping and road vulnerability index development in the Chittagong Hill Tracts, Bangladesh

Landslides pose a persistent threat to life, livelihoods, and critical infrastructure in the Chittagong Hill Tracts (CHT) of Bangladesh, a region marked by steep slopes, intense monsoonal rainfall, and increasing anthropogenic pressure. This study presents a comprehensive landslide susceptibility assessment for the monsoon-dominated Chittagong Hill Tracts (CHT), addressing data scarcity challenges. An ensemble model integrating Convolutional Neural Networks, Deep Neural Networks, and Long Short-Term Memory networks was developed using ten standardized 30-meter resolution geo-environmental variables. A balanced dataset of 5,082 samples was generated through spatially stratified random sampling. Spatial-block 10-fold cross-validation demonstrated strong model performance with an F1-score of 0.8923 and an AUC of 0.9558, outperforming individual models. Key drivers identified include annual rainfall and proximity to fault lines, highlighting the role of hydro-tectonic factors. Additionally, a Road Vulnerability Index based on six physical parameters revealed that 13% of the regional road network is highly vulnerable to landslides. These findings offer high-resolution, interpretable susceptibility and infrastructure risk maps to support disaster risk reduction strategies, contributing to Sustainable Development Goals 11 and 13. This framework advances understanding of landslide hazards and infrastructure vulnerability in data-limited mountainous regions.
Ensemble deep-learning framework for landslide susceptibility mapping and Road Vulnerability Index (RVI) development in the Chittagong Hill Tracts, Bangladesh (Figure).