Interpretable machine learning on large samples for supporting runoff estimation in ungauged basins

The distribution of flowmeter data and basin characteristic information exhibits substantial disparities, with most flow observations being recorded at a limited number of well-monitored locations. The perennial challenge of achieving reliable and robust hydrological modeling in ungauged catchments through regionalization has persisted. The increasing availability of large-scale hydrological datasets, coupled with recent advancements in machine learning techniques, offers new opportunities to explore patterns of association between basin attributes and hydrological parameters to enhance streamflow predictions. A novel parameter cross-regional transfer approach based on interpretable machine learning (XGBoost) is proposed to accurately predict runoff processes in ungauged regions by leveraging well-trained models across numerous basins within climate zones. We validate the effectiveness of this framework across 5,764 basins in a large sample dataset (Caravan), employing Nash-Sutcliffe Efficiency (NSE), RMSE and bias to assess performance. And a comparison is made with deep transfer learning based on LSTM and Transformer. Results indicate that the proposed method achieves NSE values exceeding 0.2 for 75 % of the ungauged basins, demonstrating superior performance and more stable accuracy compared to pure deep learning models, owing to its incorporation of physical constraints. Furthermore, the response of parameters to basin attributes within different climatic zones in the large-sample context is elucidated through SHAP values, enriching the understanding of hydrological features through data-driven inverse inference. These findings underscore the capability of interpretable machine learning to leverage hydro-physical regularities extracted from abundant basin features, thereby enhancing the accuracy of runoff predictions in ungauged regions.