Abstract: The relationship between rainfall and runoff is of great significance for the allocation of water resources and the protection of water and land resources in agricultural areas, but it is difficult to deal with the rainfall-runoff process under different land use types in small watersheds. The long short-term memory model (LSTM) and the Xinʼanjiang model combined with ensemble Kalman filter (EnKF) technology were used to explore the simulation effectiveness of data-driven machine learning (ML) model on rainfall-runoff process under different land use patterns, and the simulation effectiveness was compared with that of SWAT hydrological model. The estimation effectiveness of EnKF on hydrological parameters ensembles in the Xinʼanjiang model and the patterns of filter-estimated parameters were studied, and the runoff processes for different agricultural land use types based on the calibrated parameters were simulated. The results showed that the runoff value was easier to learn in the case of high runoff with a slightly small slope and the low runoff process with a large slope. The simulation accuracy and stability of the SWAT model were not as good as those of the LSTM model, but SWAT model could reflect the local soil hydrological conditions to a certain extent, which was convenient for genetic analysis. The EnKF technology had the functions of parameter update and parameter estimation, which could optimize the runoff simulation effectiveness of the Xinʼanjiang model.

Key words: rainfall-runoff simulation, data driven, data assimilation, LSTM, EnKF, Xinʼanjiang model, land use pattern, optimize forecasting