Spatial Prediction of Landslide Susceptibility in Kodagu Using Random Forest Models: Insights from the 2018 Landslide Inventory

ABSTRACT Landslides are among the most destructive natural hazards, causing substantial socio-economic losses and fatalities, particularly in the Kodagu region of the Southwestern Ghats of India. Identifying landslide-prone zones is therefore crucial for risk mitigation and sustainable land-use planning. This study presents a machine learning based approach for landslide susceptibility mapping (LSM) using a newly generated landslide inventory corresponding to the 2018 landslide events in Kodagu District. This article presents a machine learning model for mapping landslide susceptibility using a newly generated landslide inventory of the 2018 landslide events in Kodagu District to evaluate the effectiveness of Random Forest (RF) techniques. Optimisation techniques, namely grid search and random search, were employed to enhance the effectiveness of RF models in predicting landslide-prone areas, thereby ensuring a more accurate and reliable landslide susceptibility map. Accordingly, six conditioning factors namely elevation, slope, aspect, Stream Power Index (SPI), curvature, and Topographic Wetness Index (TWI) were incorporated in the analysis. The Results indicate that although grid search achieves predictive performance comparable to random search, it requires substantially higher computational time. The RF classifier delineated 2502 km2 and 1579 km2 as low- and high-susceptibility zones, respectively, whereas the RF regressor identified 876 km2, 595 km2, and 574 km2 as very high, high, and moderate susceptibility zones, respectively. The Variable importance analysis revealed slope, elevation, curvature, and SPI as the dominant factors influencing landslide occurrence. The RF regressor achieved higher prediction accuracy (95%) than the RF classifier (85%), owing to its flexibility allows it to detect gradual environmental changes and include borderline regions often excluded by the classifier. Therefore, the Landslide Susceptibility Map (LSM) generated by the RF regression model is more reliable and has more discrimination power compared to the RF classifier model. Based on the newly generated landslide inventory map of the 2018 landslide in Kodagu, this preliminary study aims to assist new researchers in identifying the most suitable choice between the RF regressor and classifier for prediction modelling.