An interpretable deep learning framework for flood prediction in the Lower Mekong River Basin

The Lower Mekong River Basin (LMRB) is a flood-prone region experiencing increasing flood risk due to climate change and human activities. This growing challenge underscores the need for robust hydrological models capable of accurate flood prediction. Although purely deep learning approaches have demonstrated strong predictive performance, their data-driven nature does not explicitly represent the underlying physical mechanisms, which limits their interpretability.

In this study, we develop an interpretable deep learning framework based on a Long Short-Term Memory (LSTM) model to predict river discharge across multiple subbasins in the LMRB, with post-hoc interpretation provided by SHapley Additive exPlanations (SHAP). Feature contributions and dominant flood drivers are analysed using SHAP, enabling transparent interpretation of the model’s predictions. The LSTM model demonstrates high predictive performance, achieving Nash–Sutcliffe Efficiency values above 0.9 across all subbasins, although the largest flood peaks are slightly underestimated in midstream subbasins during extreme rainfall events. SHAP analysis indicates that soil-related variables are predominant contributors to discharge prediction, and their influence is partially mediated through interactions with precipitation and runoff. Furthermore, the relative importance of contributing variables changes over time: soil and vegetation-related variables dominate in earlier periods from 2013 to 2017, whereas hydrometeorological variables are more influential after 2017.

Overall, this study highlights the potential of post-hoc interpretable techniques applied to deep learning models for identifying the main contributing variables for discharge prediction and the drivers of flood events across the subbasins of the LMRB, providing valuable insights to support improved flood risk management.