Decoding Landslide Susceptibility in Wayanad District of Kerala, India, Using Machine Learning Approach

Abstract

Landslide Susceptibility Zonation (LSZ) is essential for comprehending and predicting landslide events, especially in areas prone to natural hazards. This study assesses and contrasts the efficacy of two machine learning (ML) algorithm, Random Forest (RF) and Support Vector Machine (SVM), in producing high resolution LSZ maps for the Wayanad area in Kerala, India. The region is significantly susceptible to landslides, as evident by a disastrous occurrence on July 30, 2024, which led to more than 300 deaths and impacted almost 5,000 individuals. LSZ map was created using twelve landslide conditioning factors (LCFs) at a spatial resolution of 12.5 × 12.5 m. The evaluation of multicollinearity confirmed the independence of the explanatory factors. The model training utilized a balanced dataset consisting of 314 landslide and 314 non-landslide sites. The RF model revealed high susceptibility zones including 23.8% of the study region, while the SVM model recognized 19.5%. These zones are primarily located along the southwestern, western, and northwestern boundaries of Wayanad. The predictive capacities of the models, assessed using Receiver Operating Characteristic (ROC) analysis, demonstrated accuracies of 95.8% for RF and 93.5% for SVM, reflecting the strong performance of both techniques. The findings highlight the efficacy of ML algorithm, particularly RF, in LSZ, offering critical insights for hazard mitigation and land-use planning in comparable geologically vulnerable areas. © King Abdulaziz University and Springer Nature Switzerland AG 2025.