Performance Evaluation and Uncertainty Quantification of Deep Learning Models for Agricultural Drought Prediction

Early prediction of agricultural drought is critical for minimizing its adverse impacts. Although numerous studies have addressed drought forecasting, limited attention has been given to the uncertainty analysis of predictive models. This study employed quantile regression (QR) to quantify the uncertainty in deep learning models—specifically convolutional neural networks (CNN) and long short-term memory (LSTM) networks—for agricultural drought prediction and analyzed the capability of the developed models in predicting various categories of drought using the standardized soil moisture index (SSI). The models incorporate both local meteorological variables and global teleconnection indices and were applied to the Palakkad district of Kerala, India. The results indicate that integrating global climatic indicators enhances model performance at longer lead times and reduces predictive uncertainty. Notably, the inclusion of climatic indices resulted in an approximate 9% reduction in root mean square error (RMSE) at a 5-month lead time compared with models using only local meteorological variables. Across all lead times and grid points, CNN models consistently outperformed LSTM models. Category-wise analysis further revealed that the models are effective in predicting rare and extreme drought events. Among the tested configurations, the CNN model incorporating both meteorological and climatic indices proved to be most effective for long-lead agricultural drought prediction, based on both performance metrics and uncertainty quantification. The proposed methodology is generalizable and can be extended to other locations, contributing to more-informed drought mitigation strategies.