Catchment runoff simulation by coupling data assimilation and machine learning methods

Accurate catchment runoff simulation under the changing environment has a great significance in the flood disaster prevention and regional water resources management.The machine learning (ML) approach has been widely and successfully applied in runoff modelling during recent years, which, however, has not yet fully considered the potential impact of changes in hydrological intermediate state variables.This study proposed a coupled ML-based model for runoff simulating by integrating the ensemble Kalman filter (EnKF), the principal component analysis (PCA) and the long short-term memory (LSTM), which denoted as EnKF-PCA-LSTM.The specific steps include: ① The dynamic update of hydrological intermediate state variables via the EnKF method; ② The integration of updated state variables into the input set for predictor selection by the PCA method; ③ Runoff simulation through the combination of chosen predictors with the LSTM model.Taking the Ganjiang River basin as a case study, we provided a comprehensive assessment on the runoff simulation performance of the EnKF-PCA-LSTM, and performed comparisons against that of the original LSTM model and the physical hydrological model HYMOD.Results show that the EnKF-PCA-LSTM outperforms both the LSTM and HYMOD models, as reflected by the higher Nash-Sutcliffe efficiency coefficients, the Kling-Gupta efficiency coefficient and the Nash-Sutcliffe efficiency for the log-transformed runoff (0.954, 0.971 and 0.972, respectively).This finding suggests that considering the hydrological intermediate state could effectively improve the accuracy and stability of ML models in terms of runoff simulation, which undoubtedly provides valuable insight into the catchment runoff modeling.