Abstract

Agricultural drought is a critical hydroclimatic hazard that severely impacts environmental and socioeconomic conditions, particularly in arid and semi-arid regions of Ningxia, China. Timely and accurate monitoring of agricultural drought is essential for understanding its dynamics and informing effective drought management strategies. This study presents a deep learning-based technique to capture the complex, nonlinear relationships between drought inducing factors using multi-source remote sensing and meteorological data from 2001-2020. Vegetation indices, precipitation, and soil-related variables were incorporated as independent predictors in a Deep Feedforward Neural Network (DFNN), with the Soil Moisture Deficit Index (SMDI) serving as the dependent variable during the growing season (April–October). For comparative analysis, two widely used machine learning algorithms Extreme Gradient Boosting (XGBoost) and Gradient Boosting Machine (GBM) were also evaluated. The results demonstrate that the DFNN model significantly outperformed both XGBoost and GBM in predicting SMDI, effectively capturing the spatial and temporal variability of agricultural drought across the growing season. In particular, the Standardized Precipitation Evapotranspiration Index (SPEI-3), which reflects short- to medium-term climatic anomalies, was identified as the most influential predictor, explaining 21.17% of the variability in agricultural drought. Moreover, the DFNN model exhibited strong stability and generalization ability during cross-validation, achieving a coefficient of determination (R²) of 0.96, an RMSE of 0.26, and an MSE of 0.07 during the training phase. Furthermore, the DFNN model provided comprehensive agricultural drought monitoring by generating consistent spatial patterns of SMDI, demonstrating its applicability for accurate drought monitoring.