Advanced AI techniques for landslide susceptibility mapping and spatial prediction: A case study in Medellín, Colombia

Landslides, a global phenomenon, significantly impact economies and societies, especially in densely populated areas. Effective mitigation requires awareness of landslide risks, yet temporal links between occurrences are often neglected, challenging model performance due to non-stationary triggering and predisposing factors. This study presents a novel landslide susceptibility model that incorporates spatial and temporal dependencies, including landslide recurrence. We applied AI models — Naive Bayes, Linear Discriminant Analysis, Quadratic Discriminant Analysis, Decision Trees, Random Forest, and Support Vector Machine (SVM) — to a case study in Medellín, a mountainous city in northwest Colombia. Using heuristic methods, we evaluated geological and geomorphological characteristics to identify high-risk areas. Integrating temporal data from four consecutive periods allowed us to enhance estimation robustness by incorporating random effects. Our findings identify slope, stream distance, geology, geomorphology, and mean annual precipitation as key factors influencing landslide susceptibility in Medellín. The SVM model demonstrated superior performance with an accuracy of 85%, closely aligning with previous studies. This research underscores the importance of temporal dynamics in landslide susceptibility assessments, improving prediction accuracy and supporting more effective risk management. • This study introduces a landslide model integrating spatial and temporal dependencies. • Heuristic methods assess Medellín’s geology and geomorphology using risk conditions. • AI models like SVM and Random Forest assessed landslide risk in Medellín, Colombia. • Slope, stream distance, geology, geomorphology, and rainfall drive landslide susceptibility. • Temporal data over four periods improve model accuracy by integrating random effects.