Expert-Interpreted Geomorphological Maps Enhanced Machine Learning for Landslide Susceptibility Mapping in Southern Taiwan

Landslide susceptibility mapping traditionally relies on topographic, hydrological, and geological factors derived from Digital Elevation Models (DEMs). However, conventional parameters may not fully capture geomorphological processes and terrain evolution histories that indicate potential future hazards. This study integrates expert-interpreted geomorphological maps into machine learning models to enhance landslide prediction in Taiwan's mountainous regions. We compared five machine learning models (Logistic Regression, Random Forest, XGBoost, CatBoost, and LightGBM) in the Laku River basin, southern Taiwan. Expert-interpreted geomorphological maps provided four critical features, debris avalanche-prone areas, rockfall zones, alluvial fans, and old landslide locations, representing historical mass movement signatures that DEM-derived parameters cannot discover. Based on testing results, XGBoost outperformed all models, and integrating geomorphological maps significantly improved performance: F1-score increased from 0.8364 to 0.8530, with recall improving by 2.9%. This enhancement was particularly evident in detecting actual landslide occurrences along landslide boundaries, critical for high-risk applications. Furthermore, SHAP analysis revealed that debris avalanche features, NDVI, and rockfall zones were the top three contributing features. Unlike Logistic Regression, which suffered from multicollinearity with geomorphological features, tree-based models effectively leveraged expert knowledge for improved decision-making. This research demonstrates that expert-interpreted geomorphological maps, encoding long-term landscape evolution, significantly enhance machine learning-based landslide susceptibility assessment through improved model interpretability and prediction accuracy.