Runoff simulation driven by multi-source satellite data based on hydrological mechanism algorithm and deep learning network

Study region: Four sub-basins of the Songhua River basin, northeast China. Study focus: Conventional runoff models typically require in-depth knowledge of the hydrological and physical processes and are costly to construct and compute. Moreover, these models predominantly rely on ground site data, where incomplete or delayed data might introduce simulation uncertainty. Therefore, it is imperative to provide a scientifically rigorous and rational approach for simulating the runoff process, effectively addressing the limitations of existing methods. Combining a long short-term memory (LSTM) network with a modified Michel soil conservation service (MMSCS) algorithm, this study proposed the LSTM-MMSCS runoff simulation scheme. New hydrological insights for the region: The LSTM-MMSCS model was constructed by adjusting and optimizing the difference characteristics of the LSTM runoff simulation by establishing regression relationships according to the MMSCS-calculated runoff depth. LSTM-MMSCS adopted the coupling method of hydrological mechanism and deep learning to establish a simulation framework with adaptive feedback and adjustment between observed and simulated data. This scheme incorporated satellite meteorological products, solving the problem of inaccuracies caused by standard models' ineffective mining of temporal series information. LSTM-MMSCS reduced overall runoff error (RMSE was reduced from 50.07 mm to 24.47 mm) and effectively alleviated the problem of peak runoff underestimation (the relative error was reduced from 30.39% to 13.39%) compared to LSTM. Using satellite meteorological data to drive LSTM-MMSCS enabled runoff change trends visualization and aids in abnormal runoff localization.