Machine Learning‐based landslide susceptibility analysis in volcanic islands: A comparison with traditional approaches

Landslide susceptibility mapping is essential for risk mitigation and territorial planning across diverse geologic settings. This study applies a machine learning approach (Random Forest regression implemented in ArcGIS Pro) to model landslide susceptibility across the Canary Islands, Spain. A geostatistical dataset was compiled using key predictive variables including slope, elevation, vegetation (NDVI), extreme rainfall (99th percentile), and geological units (from the official Geotechnical Map by SITCAN). The model was trained using the BDMOVES landside inventory from Spanish Geological Survey (IGME-CSIC), which contains 1144 documented events across the archipelago, categorized by their destructive potential, calculated using event velocity (extremely slow to extremely fast) as proxy. An additional 2288 randomly located absence points (locations where no events have been recorded) were included, resulting in a dataset of 3432 geolocated training points. Validation of the model yielded a coefficient of determination (R²) of 0.90, indicating excellent predictive performance. The results were then compared to a traditional (heuristic multicriteria) susceptibility model applied to the island of La Palma, demonstrating that the Random Forest model significantly improves the spatial precision and statistical robustness of susceptibility assessments. Further ground truthing in the islands of Tenerife and El Hierro reinforced the initial results. This study represents a substantial methodological advance in landside susceptibility modeling for volcanic islands, and underscores the value of integrating official datasets with machine learning techniques for enhanced geohazard analysis.