Integrated drought monitoring and analysis: A novel framework based on multi-source remote sensing data and ensemble machine learning

Against the backdrop of climate change, drought risks are escalating in critical agricultural regions, highlighting the need for effective monitoring tools. Existing in-situ and remote sensing-based drought monitoring methods suffer from low accuracy, poor spatial representativeness, and insufficient explanatory power. To address these gaps, we propose a novel framework integrating multi-source remote sensing data and an ensemble machine learning (ML) model. This approach was validated using the Beijing-Tianjin-Hebei-Shandong-Henan region in China as a case study. The results of this study indicate that the Bayesian-weighted ensemble model effectively captures the nonlinear relationships between drought and its driving factors across multiple time scales (1, 3, 6, and 12 months), thereby enhancing prediction accuracy. For the Standardized Precipitation Evapotranspiration Index (SPEI), the model achieves R2 values ranging from 0.71 to 0.74 across the four time scales. Additionally, it attains over 78% accuracy in classifying different drought severity classes, with a 98% accuracy rate for extreme drought detection. Correlation analysis identifies precipitation anomalies (Pa, R = 0.31) and potential evapotranspiration (PET) as key correlates of short-term drought (SPEI-1). SHAP (SHapley Additive exPlanations) further quantifies their contribution at 21% each, confirming them as dominant drivers. For long-term drought, correlation analysis shows soil moisture is critical (R > 0.27, P