Explainable Deep Learning for Assessing Terrestrial Hydroclimate Impacts on Coastal River Discharge

Abstract

Coastal river discharge plays a critical role in regulating coastal water quality, ecosystem health, and community resilience. However, understanding how terrestrial hydroclimate influences past and future discharge remains limited due to coarse and biased climate inputs, structural limitations of hydrologic models, and the black-box nature of many deep learning approaches. Here we present an explainable deep learning framework (X-CNN-LSTM) that integrates convolutional neural networks (CNNs) and long short-term memory (LSTM) networks to capture spatial and temporal controls on discharge, incorporates trend-preserving, super-resolution deep learning to downscale climate simulations, and enables saliency-based interpretability. The X-CNN-LSTM achieves Kling–Gupta Efficiencies (KGE) of 0.825, 0.786, and 0.714 for the Alabama, Tombigbee, and ACF Rivers, respectively, outperforming standalone deep learning models, process-based models, and flood reanalysis data. Saliency-based analysis identifies runoff and soil moisture as dominant predictors and highlights key regions driving discharge variability. Additionally, the use of trend-preserving, multivariate super-resolution deep learning to downscale climate inputs significantly improves discharge simulation accuracy during the historical period. To assess future climate impacts, we drive X-CNN-LSTM with super-resolution deep learning downscaled CMIP6 climate projections under SSP2-4.5 and SSP5-8.5 scenarios. Results show consistent decreases in low flow extremes and increase in interannual variability, signaling heightened hydrologic risks. These trends are robust across multiple downscaled CMIP6 climate and Earth system models. Overall, this explainable deep learning framework offers a notable advancement in high-resolution climate impact assessment of coastal river discharge.