A Pilot Study of SHAP-Interpreted Machine Learning for Pixel-Level Landslide Classification from High-Resolution DEM and Satellite Imagery

Accurate delineation of current landslide extent is important for hazard assessment, sustainable watershed management, and disaster risk reduction in tectonically active mountainous regions. This study presents a pilot machine learning framework for pixel-level landslide classification in the Laonung (Laonong) Creek Watershed, southern Taiwan, using very high-resolution digital elevation model (DEM) derivatives and SPOT-6 multispectral imagery. Thirteen geomorphometric and spectral features, including slope, curvature, and six spectral indices derived from SPOT-6 bands, were extracted from 96 landslide-containing tiles within a pilot subregion of the watershed; no landslide-free tiles were included in model training or evaluation. Landslide annotations followed a geomorphic-unit delineation protocol in which optical imagery provided the primary evidence of current activity and DEM-derived hillshade supported boundary refinement. Three classifiers were evaluated using column-quartile spatially blocked four-fold cross-validation, with each fold comprising a geographically contiguous range of columns, to reduce spatial leakage: logistic regression (LR), random forest (RF), and XGBoost. All three models substantially outperformed the no-skill baseline for the resampled evaluation dataset (average precision, AP =0.250), achieving mean AP values of 0.854±0.040, 0.858±0.033, and 0.846±0.035 for LR, RF, and XGBoost, respectively. The convergence of linear and nonlinear model performance suggests that the dominant discriminatory signal is largely captured by relatively simple spectral and topographic predictors within this pilot dataset, rather than reflecting a general property of landslide classification. SHapley Additive exPlanations (SHAP) analysis across all four spatial folds identified SPOT-6 Band 3 (Red) as the dominant predictor in every fold, with NDVI a robust secondary predictor, consistent with the spectral characteristics of fresh bare-soil landslide surfaces and with the optical cues used in the annotation protocol. The results are interpreted in the context of the pilot dataset’s limited spatial extent, the resampled class distribution used for model evaluation, and unquantified label uncertainty. This study provides a transferable methodological baseline for future, larger-scale landslide classification analysis in the Laonung Creek Watershed and highlights the potential contribution of spatially explicit landslide mapping to sustainability-oriented disaster management.