An explainable deep learning approach for investigating the mechanism of agricultural drought in the Poyang Lake Basin

This study proposes an explainable deep learning approach to investigate the driving mechanisms of agricultural drought. As a prerequisite, accurate long short-term memory (LSTM) models were developed to predict volumetric soil moisture across four layers (0-7 cm, 7-28 cm, 28-100 cm, and 100-289 cm) in the Poyang Lake Basin. Agricultural drought events were subsequently identified using the standardized soil moisture index (SSMI). The Expectation gradient (EG) and additive decomposition (AD) were then applied to analyze how the LSTM model responds to meteorological inputs and to examine its internal workings. EG analysis shows that in the shallow soil layer (0-7 cm), during the 0-15 days preceding drought onset, the EG score for temperature-related inputs was 56.93, higher than the 45.24 for precipitation. This suggests that temperature-related inputs are the dominant drivers of shallow-layer drought formation. In contrast, in the deep soil layer (100-289 cm), the cumulative EG score for precipitation over the 45-90 days before drought onset reached 102.57, significantly exceeding the 58.93 for temperature-related inputs. This highlights the lagged and accumulative nature of deep-layer droughts, primarily driven by long-term precipitation deficits. Seasonal analysis reveals that winter droughts are primarily driven by long-term precipitation deficits, whereas short-term high temperatures and sharp declines in precipitation trigger summer droughts. AD analysis further reveals that shallow soil layers typically respond to drought approximately 7-10 days before its onset, while deep soil layers can begin capturing and retaining precipitation signals as early as 30-45 days prior to the drought event. This study improves the explainability of deep learning models, advances the understanding of agricultural drought mechanisms, and provides methodological support for drought monitoring and early warning.