SHAP-Augmented Neural Networks for Landslide Susceptibility Mapping in Darjeeling-Gangtok Region

The Darjeeling-Sikkim Himalayas have been perennially affected by landslide hazards owing to their fragile terrain and complex geological and geotechnical settings. The region is of economic and strategic importance and the loss of lives and damage to property due to landslides in this area has been significant. Landslide Susceptibility Mapping (LSM) plays a critical role in evaluating the risks and provides valuable insights towards assessing the vulnerability of the region and encompassed infrastructure and settlements exposed to the hazard. Artificial Neural Networks (ANNs) are data-driven models known for their ability to universally approximate non-linear functions with complex correlations. This study aims at the comparison of the efficacy of multi-layered ANNs of different depths in the spatial analysis of classifying the region between the cities of Darjeeling and Gangtok based on their landslide susceptibility. The ANN models with one, two and five hidden layers were compared to understand the optimal complexity required for accurate and reliable landslide susceptibility analysis for the study area chosen. The SHapley Additive exPlanations (SHAP) are adopted to bring more interpretability to the model, thereby veering away from the black-box nature of the machine learning (ML) models.The optimization of the depth of ANN models revealed that 2 hidden layers were sufficient to successfully capture the complex relationships between the input parameters. Valuable insights about the possible triggers for the landslides was obtained by SHAP analysis which can be particularly helpful in further analyzing the landslide phenomenon in the study area. This study takes a significant step forward in demonstrating how advanced models (in this case, any ML models) can be made more interpretable and credible, thereby promoting informed decision-making and effective landslide risk management by taking the example ofDarjeeling-Sikkim Himalayas.

Keywords: Landslide susceptibility mapping, Multi-Layered ANNs, Explanative machine learning, SHapley Additive exPlanations (SHAP).