Uncovering the impact of catchment attributes in regionalized models through explainable artificial intelligence

Study region 531 ungauged basins in the contiguous United States (CONUS). Study focus: We propose XNLSTM, an interpretive framework combining regionalized Long Short-Term Memory (LSTM) modeling, 10-fold basin cross-validation, and explainable artificial intelligence (XAI) methods, including Shapley Additive Explanations (SHAP), the Morris method, and permutation feature importance (PFI), to examine how catchment attributes affect runoff prediction in ungauged basins. The framework jointly uses multiple XAI methods to reveal interpretational variability and identify catchment attributes with relatively stable importance across methods. It then trains models under different catchment attribute configurations derived from the XAI-based interpretation results and interprets fold-specific models to assess the stability of catchment attribute importance across basin splits and configurations. New hydrological insights for the region: SHAP, Morris, and PFI produce different rankings for the same LSTM model, but p_mean, aridity, and other catchment attributes remain consistently important. Across the three XAI rankings, climatic and topographic attributes are consistently more important than soil, geological, and vegetation attributes for overall runoff prediction. Using more catchment attributes, and using attributes with higher XAI-derived importance, improves cross-basin generalization. However, lower ranked attribute combinations show slight tendencies to reproduce low-flow and high-flow metrics closer to observations. Importance rankings are relatively stable for all-attribute models but change more clearly when models use attribute subsets. Attributes such as elev_mean, pet_mean, and soil_conductivity remain important across basin splits and configurations.