Toward improved deep learning-based regionalized streamflow modeling : Exploiting the power of basin similarity

Abstract

In deep learning (DL)-based regionalized streamflow modeling, basin similarity has demonstrated to be effective for sharing hydrological information. However, the differences in the use of hydrological information by DL due to different basin similarity strategies remain underexplored. For this, we cluster and regionalize 222 Australian basins based on hydrology, climate, landscape, and position characteristics, and compare their performance with the benchmark model. The results reveal: (1) Basin similarity strategies-based models outperform the benchmark model, demonstrating the effectiveness of basin similarity strategies; (2) Hydrology similarity yields the best model, while climate similarity is relatively stable, suggesting that the key hydrological information for improving DL performance comes from hydrology and climate characteristics; (3) The differences in DL’s utilization of hydrological information are influenced by the combined effects of basin climate, hydrology, soil, and vegetation conditions. This study provides insights into how DL-based regionalized streamflow modeling more effectively utilize basin hydrological information.