Landslides and AI: Mapping Risks After Typhoons

After Typhoon Khanun hit mountainous areas in Taiwan, a team of researchers used artificial intelligence to detect landslides. Their approach relies on an algorithm called Random Forest, which analyzes high-resolution satellite images and 22 environmental factors. The challenge? Landslides are rare compared to other land cover types, creating data imbalance. To fix this, researchers tested several balancing techniques, including SMOTE. The Distance_SMOTE variant performed best, improving model accuracy from 74% to 85%. The resulting maps show a significant increase in landslide-affected areas after the typhoon. This method enables rapid response during emergencies, supporting land management and community recovery. The effort aligns with the UN Sustainable Development Goals, especially SDG 11, 13, and 15.

The study focuses on the application of machine learning for post-disaster land use-land cover (LULC) classification and landslide detection in Nanfeng Village, central Taiwan, following Typhoon Khanun in August 2023. High-resolution Pléiades satellite imagery was combined with 22 environmental and spectral variables, including slope, solar exposure, vegetation density, and geomorphological data. The Random Forest classifier was selected for its robustness in handling large heterogeneous datasets and its ability to estimate variable importance. However, the distribution of land cover classes was highly imbalanced, with landslides represented by a very limited number of samples. To address this challenge, several variants of the SMOTE (Synthetic Minority Oversampling Technique) were evaluated, a methodology that generates synthetic examples for minority classes. Among the tested variants, Distance_SMOTE proved most effective, improving overall model accuracy from 74% to 85% and the Kappa coefficient from 0.69 to 0.82. The F1-score for the 'landslide' class rose to 0.97, marking a significant improvement over other classes such as roads (0.85) and grasslands (0.78). The generated LULC maps allow comparison of the territory before and after the typhoon, highlighting a marked expansion of landslide-affected areas. This approach proves particularly useful for rapid assessments in emergency contexts, providing reliable data to plan restoration and risk mitigation interventions. Furthermore, the integration of advanced machine learning techniques with open satellite data represents a key step toward continuous and predictive natural hazard monitoring systems, aligned with climate adaptation strategies and territorial sustainability.