Geospatial predictive modelling of anthropogenic and natural shipwreck risks using multi-source environmental data

Integrating Underwater Cultural Heritage into Marine Spatial Planning requires understanding underlying vessel loss mechanisms. Distinguishing anthropogenic (war-related) from nature-induced vessel loss is critical for precision conservation. We present an interpretable geospatial machine learning framework to map these risk regimes in Chinese adjacent seas. Using XGBoost and SHAP with multi-source environmental data, the model achieved robust predictive performance (AUC = 0.911). Interpretability analysis revealed distinct non-linear environmental thresholds, notably depth (~56 m) and salinity (~34.3 PSU), that spatially distinguish vessel loss mechanisms. These parameters demonstrate that historical naval conflicts predominantly clustered within the deep, high-salinity continental shelf, differentiating them from nature-dominated nearshore and reef environments. We propose an Environmental Risk Zonation Map guiding differentiated survey strategies: magnetometer detection for ferrous war wrecks, and acoustic profiling for buried wooden vessels. This approach provides a data-driven basis for risk-informed heritage management.