Explainable machine learning for spatial risk assessment of Hyphantria cunea (Lepidoptera: Erebidae) in China

The fall webworm, Hyphantria cunea (Drury) (Lepidoptera: Erebidae), is a highly polyphagous invasive pest that poses increasing economic and ecological threats to woody plants in agricultural, forestry, and peri-urban landscapes. Spatially explicit risk assessment is crucial to support targeted monitoring and management, yet the ecological drivers of its invasion risk often vary geographically. Here, we compiled 671 occurrence records from national pest surveys, GBIF, and literature to model the invasion risk of H. cunea in China using 14 climatic, topographic, vegetation, and anthropogenic predictors. Four non-spatial ensemble machine-learning algorithms were evaluated using 5-fold spatial cross-validation. The optimal model (LightGBM) was interpreted using SHAP and GeoShapley to quantify global, model-level driver importance across the China-wide dataset and geographically varying driver effects. High-risk areas were highly concentrated in the North China Plain, Liaodong Peninsula, Shandong Peninsula, and the Beijing-Tianjin region. Elevation, mean temperature of the warmest quarter, and human population density emerged as the most influential predictors. Notably, GeoShapley revealed that the impacts of elevation and population density are not uniform but are significantly amplified within specific regional hotspots. These findings indicate that uniform, "one-size-fits-all" management strategies may be inefficient. Instead, our geographically explicit framework provides a transparent spatial decision-support tool for plant protection authorities to prioritize targeted surveillance, optimize quarantine deployment, and implement early interventions against H. cunea.