| Abstract Submission No. | ABS-3.1-0063 |
| Title of Abstract | Soil Profiles Meet Explainable AI: A Catena-Based Approach to Improving Landslide Prediction |
| Authors | Achu A L*, Girish Gopinath |
| Organisation | Kerala University of Fisheries and Ocean Studies (KUFOS) |
| Address | Kerala University of Fisheries and Ocean Studies (KUFOS) Kochi, Kerala, India Pincode: 682 506 E-mail: achu.geomatics@gmail.com |
| Country | India |
| Presentation | Oral |
| Abstract | Landslides are influenced by multiple geo-environmental factors, yet most machine-learning based susceptibility models primarily depend on DEM-derived topographic variables, leaving the contribution of depth-wise soil properties underexplored. This study incorporates soil geotechnical and hydrological properties at multiple depths into landslide susceptibility modelling, employing Random Forest (RF) combined with SHAP-based Explainable AI (XAI) to enhance model transparency and interpretability. The research was conducted in the Muthirapuzha River Basin (MRB), Southern Western Ghats, India, using both grid-based and slope unit (SU)-based mapping units. Findings show that depth-wise soil parameters substantially improve model accuracy and mitigate overestimation compared to models using only surface or topographic variables. Key predictors included Field Capacity (FCY), Chemical Index of Alteration (CIA), Liquid Limit (LLT), and Unsaturated Hydraulic Conductivity, along with topographic factors such as slope angle and Topographic Wetness Index (TWI). SU-based models outperformed grid-based ones in AUROC and offered deeper insights into landslide depth and volume. SHAP values and waterfall plots effectively explained model outputs and provided site-specific insights, demonstrating their utility for landslide risk assessment. The consistent ranking of predictor importance across mapping units further supports the robustness of the selected variables. Overall, this study highlights the significance of incorporating depth-wise soil parameters in landslide susceptibility modelling and advocates for SHAP-based XAI methods to strengthen model transparency and applicability, particularly in tropical mountainous regions. |
| Keywords | Landslides, Explainable AI, Soil parameters, Western Ghats, Kerala, India |