An explainable AI approach for interpreting regionally optimized deep neural networks in hydrological prediction

Abstract

The interpretation of artificial intelligence (AI) and deep learning (DL) model outcomes remains a central challenge in hydrology and rainfall-runoff modeling. This study investigates whether hyperparameter-optimized regional Long Short-Term Memory (LSTM) networks can implicitly learn hydrological processes directly from hydrometeorological data, without access to explicit catchment attributes during training. Specifically, we explore to what extent these models reinforce classical hydrological understanding or reveal new insights through explainable AI (xAI) analyses. Using hourly precipitation, temperature, and potential evapotranspiration data from 40 humid and flashy catchments in the Basque Country, Spain, we demonstrate that systematically optimized LSTMs exhibit strong generalization and scalability in regional rainfall-runoff modeling. Through a combination of correlation analysis, Random Forest (RF) modeling, Principal Component Analysis (PCA), and SHAP-based feature attribution, we quantify how catchment attributes indirectly influence LSTM performance. This multi-method approach provides a novel framework to assess the hydrological "learning maturity" of deep neural networks in regional hydrology. The results show that LSTM networks implicitly capture latent catchment characteristics that shape hydrological responses. Catchments with high runoff coefficients and higher mean annual streamflow tend to yield more accurate predictions, while catchments characterized by steep slopes, extreme flow variability, and high precipitation variability pose greater challenges due to their nonlinear hydrological behavior. SHAP analysis confirms that both catchment properties (e.g., average yearly runoff coefficient, precipitation, streamflow) and key LSTM hyperparameters (e.g., input sequence length, hidden size, dropout rate) play critical roles in predictive success, with the latter influencing the models' ability to better generalize and capture extremes. Furthermore, RF and PCA highlight essential factors influencing model accuracy, including annual precipitation, aridity index, stream density, and land cover, where broadleaf forests improve water retention and urbanization complicates runoff processes. These findings bridge the gap between the "black-box" nature of AI/DL models and hydrological interpretability, offering evidence-based guidelines for practitioners and researchers deploying LSTMs in regional hydrological contexts. Ultimately, this research underscores the potential of optimized LSTM networks to generalize hydrological processes and reinforce domain knowledge, while also advocating for the integration of catchment attributes in future model designs to enhance predictive robustness. By advancing xAI methodologies for deep learning in hydrology, this study contributes to developing more reliable and interpretable AI-driven solutions for water resource management and flood risk assessment under increasing climate variability.