Advancing Deep Learning for Regionalized Streamflow Modeling: A Comparative Study of Basin Similarity Approaches

This study evaluates the applicability of similarity-based regionalization methods across three independent large-sample hydrological datasets. The analysis covers 1318 basins distributed across three continents—North America (United States), South America (Brazil), and Oceania (Australia)—spanning diverse climatic, hydrological, and topographic conditions. A regionalization framework combining principal component analysis (PCA), hierarchical clustering, and long short-term memory (LSTM) modeling is developed. PCA is first applied to reduce dimensionality and extract multi-dimensional features from basin climate, hydrological, topographic, and spatial attributes. Hierarchical clustering is then used to group basins in this integrated attribute space, forming similarity-based basin clusters. On this basis, LSTM models are trained within each similarity group to perform regionalized prediction. Basin-wise validation experiments were conducted independently within each study region to evaluate the effectiveness of the proposed regionalization framework. Results show that similarity-based regionalization models consistently yield notable improvements in prediction performance across all three continents and exhibit improved stability and consistency across the three independent study regions. Hydrological similarity provides the best overall performance, with median and mean improvements of approximately 0.005 and 0.01, respectively, relative to other approaches. However, the most effective similarity criterion varies between regions, reflecting regional differences in the dominant influence of climate, hydrology, and topography on model performance. In some regions, specific factors show clear correlations with prediction skill. Overall, the findings demonstrate that the proposed regionalization framework can consistently improve LSTM performance across three independent study regions, highlighting its applicability under diverse hydro-climatic conditions and providing practical guidance for regional hydrological modeling.