A data-driven approach for flood prediction using grid-based meteorological data

Establishing a physically-based hydrological model for flood prediction in ungauged or data-limited catchments has always been a difficult problem. In this study, a data-driven approach based on the NASA Global Land Data Assimilation System (GLDAS) data is proposed for flood prediction with the assistance of the Gamma Test. A runoff generation model together with a routing model based on the most advanced deep learning model, i.e. Long Short-Term Memory (LSTM) network is established. By calculating the noise of the input data, Gamma Test can effectively avoid the overfitting phenomenon with the LSTM network. Taking the a small-scale mountainous catchment from northern China as an example, Gamma Test in this study is used to help select the optimal combination of the GLDAS inputs for the runoff generation model, and meanwhile the input grids involved in the routing model. The established models are then verified based on the Nash-Sutcliffe Efficiency coefficient (NSE), and it is found that the surface runoff generation model shows a good performance, with an average NSE value of 0.8708, while the baseflow-groundwater runoff generation model also results in an acceptable performance with an average NSE value of 0.6320. After the involvement of the observed rainfall data, the model performance (NSE) of the routing model has increased from 0.5738 to 0.7144. Finally, the runoff generation models and the routing model are integrated, and the grid-based GLDAS meteorological data are used directly to simulate the streamflow at the catchment outlet. The integrated model performs well with an NSE value of 0.7909, which indicates the feasibility of this data-driven approach for flood prediction using the gird-based meteorological data. The methodology adopted in this study provides a reference for flood prediction using data-driven models in ungauged or data-limited catchments.