A deep learning-based composite agricultural drought index for monitoring and impact assessment in Central Asia

Drought is a major natural hazard that seriously threatens agricultural production and food security. However, most existing drought indices rely on a single variable or scale, limiting their ability to capture the complexity of agricultural drought. In this study, we propose a Composite Agricultural Drought Index (CAEDI), developed using an unsupervised Convolutional Autoencoder (CAE) to integrate multiple drought-related indicators, including precipitation, land surface temperature, vegetation condition, and soil moisture. Guided by soil moisture anomalies as a weak physical prior, the model extracts drought-relevant features through nonlinear multivariate fusion. We evaluated CAEDI using established drought indices, in-situ observations, and crop yield data across Central Asia. CAEDI exhibits strong correlations with SPEI-2 and SPEI-3 (R > 0.80, p < 0.01) and outperforms individual indices in identifying agricultural drought. It also demonstrates high spatial and temporal consistency with conventional drought indices, strong associations with yield variability, and effectively distinguishes between wet and dry zones, showing good agreement with meteorological observations. Furthermore, we developed an Agricultural Drought Impact Index (ADI) by integrating CAEDI with Extreme Gradient Boosting (XGBoost) and SHapley Additive Explanations (SHAP). ADI effectively captures yield losses and highlights drought-sensitive phenological stages, particularly during heading and grain filling. Overall, CAEDI provides an objective, scalable, and physically grounded approach for agricultural drought monitoring and impact assessment, especially in data-scarce regions. Its integration of multi-source indicators through deep learning shows promises for enhancing early warning systems and supporting climate-resilient agricultural management.