Deep Neural Network-Based Landslide Susceptibility Mapping in Himachal Pradesh: A Comparative Study with Conventional Models

Landslides pose a persistent threat in the Himalayan terrain of Himachal Pradesh, India, driven by complex interactions among geomorphology, climate, and anthropogenic activity. This study presents a Deep Neural Network (DNN)-based framework for landslide susceptibility mapping (LSM) using high-resolution geospatial data. Key geomorphological and environmental factors were extracted and analyzed in a GIS environment, and historical landslide data were used to prepare training (75%) and validation (25%) sets. The DNN model’s performance was compared with the widely used machine learning technique Random Forest (RF) and the statistical method Logistic Regression (LR). The DNN achieved superior results with 88.69% accuracy, an AUC of 0.84, and an MSE of 0.0926, outperforming both RF and LR. The resulting susceptibility map identifies critical high-risk zones, providing actionable insights for disaster risk reduction and land-use planning. These findings demonstrate the potential of deep learning to enhance LSM accuracy and scalability in data-rich, hazard-prone regions.