A Weight‐Adaptive Ensemble Learning Method for Mapping Pine Caterpillar Infestation Risk in Northeast China Using Remote Sensing: Combining Dual Bayesian and Top‐k Strategies

ABSTRACT Pine caterpillar ( Dendrolimus ) outbreaks pose serious threats to forest ecosystems and regional economies. Long‐term analysis of infestation risk is essential for early warning and prevention. This study developed a risk assessment method for Northeast China that addresses limitations in long‐term dependency and model generalization. A multi‐factor dataset was constructed by integrating snow, soil, and other habitat variables with historical outbreak sites. A Dual‐Bayesian optimized, Top‐k adaptive weighted blending model (DBO‐Tk‐AWEL) was proposed, achieving OA 94.16%, Recall 89.66%, F1‐score 93.24%, and AUC 97.42% for 2000–2024 risk assessment. Dual‐Bayesian optimization enables automated hyperparameter and weight tuning, improving robustness and generalization in complex, high‐dimensional, small‐sample tasks. Results indicate: (1) six major high‐risk areas—Changtu–Lingyuan, Dongfeng–Fushun, Antu, Daqingshan–Qing'an, Zhalantun–Horqin, and Mohe–Oroqen. (2) spatial expansion trending eastward and northward, with increasingly concentrated high‐risk zones. (3) risk levels peaking in 2017–2020 and declining to 1.02% by 2024. (4) strong forest‐type specificity, with Korean pine–Tilia amurensis forests dominating high‐risk areas, rising red pine risk, stable Pinus tabulaeformis , low‐risk Larix–Betula mixed forests, and increasing Pinus sylvestris var. mongolica risk since 2017. (5) SHAP analysis revealing quartic polynomial relationships between key factors and risk, with temperature, solar radiation, soil temperature, snow density, and depth positively influencing risk, and precipitation, runoff, and soil humidity negatively affecting it. This study provides a robust scientific basis for targeted early warning and ecological management of pine caterpillar risks in Northeast China.