Improving Basin‐Wide Flood Estimation From a Global Hydrological Model Through Spatiotemporal‐Pattern‐Based Machine Learning

ABSTRACT When estimating future flood events using a global hydrological model (GHM), the large uncertainties associated with general circulation models (GCMs) and bias in the GHM model pose significant challenges. In the meantime, most future flood estimations are conducted only at specific gauge stations due to limited data availability and are unable to support basin‐wide water resources planning and management. To address these issues, we propose a spatiotemporal‐pattern‐based machine learning method, DSGPR‐EOF, which is a combination of Dual‐stage Sparse Gaussian Process Regression (DSGPR) and Empirical Orthogonal Function (EOF) analysis. DSGPR‐EOF is developed to improve the accuracy of basin‐wide flood estimations, including flood peak discharge, flood peak time, and flood volume. We apply the proposed method to the Brahmaputra River Basin (BRB), known for its topographical and climatic diversity, to evaluate the effectiveness and efficiency of the method. DSGPR‐EOF is shown to lead to higher accuracy in flood peak discharge estimation than the widely used multi‐GCMs ensemble mean method and several mainstream machine learning methods, including Support Vector Regression (SVR), Artificial Neural Network (ANN), and Long Short‐Term Memory (LSTM). Comprehensive comparisons reveal that DSGPR‐EOF achieves the lowest relative error of peak discharge (3.36%) among all compared methods, with particularly notable advantages in capturing higher‐order temporal patterns of flood dynamics. The errors in the estimated 10 and 100‐year flood peak discharges of DSGPR‐EOF are reduced by 68.6% and 54.5%, respectively, compared to SGPR‐EOF method. The accuracy of flood peak and volume estimated by DSGPR‐EOF method is highly consistent spatially. Furthermore, when model‐derived reference discharge data are substituted by observed data, flood peak estimation is shown to further improve over the entire basin even though the substitution is made only at locations of gauging stations. These findings underscore the practical significance of the DSGPR‐EOF method for basin‐wide flood estimation.