A physical process and machine learning combined hydrological model for daily streamflow simulations of large watersheds with limited observation data

Physically distributed hydrological models are effective in hydrological simulations of large river basins, but the complex characteristics of hydrological features limit their application. An easy-to-use and high-efficiency hydrological model is needed for efficient water resource management in practice. Machine learning (ML) based models have the potential to provide fast mapping pathways between meteorological predictors and hydrological responses without detailed descriptions of the corresponding physical processes. However, the extensive data requirements, ignoring of spatial variability and poor performance for extreme flows limit the application of ML models. This study attempts to develop an ML-based hydrological model by combining physically based distributed hydrological model with an artificial neural networks (ANN), computer vision (CV) and a categorization approach (CA). To solve the insufficient training problem, we use a physically distributed hydrological model (GBHM) together with a stochastic rainfall generator to generate a large amount of synthetic data (GBHM-ANN). To improve the extreme flow simulation, we add the categorization approach into GBHM-ANN (GBHM-ANN-CA). To capture the spatial variability of the predictors, we also use a local binary pattern-based computer vision method to form GBHM-ANN-CA-CV model. The effectiveness of the three modeling approaches are demonstrated by synthetic case studies. We finally evaluate GBHM-ANN-CA-CV using the real data from the upper Chao Phraya Basin in Thailand. The results show that the prediction accuracy of our new data-driven model is greatly improved in data-limited watersheds. Specifically, the CV extracted spatial information can improve the robustness of the data-driven hydrological model, and the CA can greatly improve high flow simulations. The combined model yields a satisfactory accuracy for long-term daily streamflow simulations. This study demonstrates the potential of ML-based hydrological models in water resource management, especially in changing environments.