Beyond Accuracy: Trustworthy LSTM-Based Hydrological Modelling Assessed with XAI and Hydrological Signatures — A Case Study in Rio Grande do Sul, Brazil

Hydrological modelling is essential for water resource management, decision making, extreme events forecasting, and for advancing an integrated understanding of the water cycle. In this context, two main approaches dominate: physics-based (or process-based) models, which simulate hydrological processes such as streamflow using fundamental physics equations, and data-driven models, which use statistical or machine learning techniques to map inputs to outputs. Although Artificial Intelligence (AI) techniques have shown promising results in predictive accuracy, particularly in data-rich basins, their inherently black-box nature raises concerns about whether their internal representations align with real hydrological processes. This is especially critical when models are applied to extreme events, non-stationary conditions, or scenarios beyond the training distribution, where high performance metrics alone may not guarantee reliable or physically meaningful predictions. In this study, we evaluated the performance of a Long Short-Term Memory (LSTM) model for drought modelling modeling and assessed how effectively it could represent real-world hydrological behavior in the Rio Grande do Sul watersheds available in the Catchment Attributes and Meteorology for Large-sample Studies (CAMELS-BR) dataset. The focus on these basins is particularly relevant given the region's hydrological importance, susceptibility to extreme events (e.g., droughts and floods), and distinct characteristics compared to temperate regions, where most legacy models were developed. The model was trained using data from 55 different basins across the state. This multi-basin approach allows the LSTM to learn universal hydrological patterns while maintaining the ability to predict low flow conditions in individual watersheds. The model inputs combined dynamic hydrological variables (e.g., precipitation and evapotranspiration) with static catchment attributes  (e.g., aridity, soil properties, and topography). Accumulated rainfall features were constructed over 3-30 day windows to capture watershed memory effects as a proxy to soil moisture dynamics. In addition, Explainable AI (XAI) techniques together with hydrological signatures (e.g. runoff ratio, baseflow index and elasticity) were applied to assess the physical soundness of the LSTM model in the region. Following this, the internal structure of the LSTM - particularly the cell states - were analyzed and compared with hydrological behavior (e.g., soil water accumulation, groundwater dynamics, rainfall inputs) in both situations where XAI and hydrological signatures highlighted, or did not highlight, physical consistency. The LSTM’s effectiveness in Brazilian watersheds highlighted its potential as a complementary tool for low flow and drought modelling, offering a valuable alternative for water resources management. XAI analyses and hydrological signatures highlighted the physical soundness of the multi-basin model, but also indicated that improvements were needed, as the internal structure did not consistently track physical hydrological behavior in some cases, hindering the extrapolation of the LSTM model to assess drought conditions in different meteorological settings (e.g., climate change scenarios).