Comparative Assessment of Machine Learning and Statistical Models for Agricultural Drought Forecasting and Driver Attribution in the North China Plain

Drought severely threatens agricultural productivity and water resources in the North China Plain (NCP). This study provides a spatiotemporal analysis and prediction of agricultural drought from 2001 to 2024 using the Vegetation Health Index (VHI) derived from satellite data. Meteorological (air temperature, precipitation) and environmental variables (soil moisture, evapotranspiration, leaf area index, enhanced vegetation index) were analysed. A multi‐method approach, including trend analysis, wavelet coherence, Granger causality, and structural equation modelling (SEM), revealed that air temperature is the dominant negative driver of vegetation health ( β = −0.444, p < 0.001), while soil moisture ( β = 0.297, p < 0.001) and vegetation greenness, particularly EVI ( β = 0.906, p < 0.001), exert significant positive influences. The northern and central NCP were identified as persistent drought hotspots. Three prediction models, Long Short‐Term Memory (LSTM), Random Forest (RF), and ARIMAX, were evaluated. LSTM showed the best overall performance, achieving the lowest RMSE (3.23) and highest R 2 (0.781), indicating its potential to capture complex temporal dynamics, while ARIMAX exhibited relatively higher forecast uncertainty. Feature importance analysis consistently identified temperature, evapotranspiration, and vegetation as important drivers across all models. This research highlights LSTM as a promising approach for agricultural drought prediction in the NCP and emphasises the roles of thermal stress and soil moisture deficit in supporting drought mitigation and sustainable water management.