Applicability of precipitation data from reanalysis as input to rainfall-runoff model using LSTM.

In recent years, global warming has progressed, and it has become important to improve the accuracy of future risk assessments. For future risk assessment, rainfall-runoff modeling is necessary to obtain flow discharge because there is no observation data in the future. Meanwhile, it is required to improve the accuracy of rainfall-runoff modeling. A rainfall-runoff model generally requires observation data for its calibration. However, observation data may be insufficient depending on the target area. Nowadays, deep learning, especially Long Short-Term Memory (LSTM) network, is frequently utilized for rainfall-runoff modeling. Then, this study employed LSTM for rainfall-runoff modeling, and also obtained precipitation data from reanalysis in order to use them as input. Reanalysis data are simulated results that are synthesized into observation data. Most reanalysis datasets cover the globe. If precipitation from a reanalysis dataset can be used for input to rainfall-runoff modeling, it may be valuable to implement a rainfall-runoff model at no or sparsely gauging basins. To investigate the usefulness of precipitation data from a reanalysis dataset for rainfall-runoff modeling, the rainfall-runoff model with LSTM was implemented at a watershed in the Hokkaido region in the northern sector of Japan, which has enough observed precipitation data. LSTM was trained with precipitation data from a reanalysis dataset and with observed precipitation data. Then, the modeling accuracy was compared. The results show that the rainfall-runoff modeling with LSTM shows high estimation accuracy not only with observed precipitation data but also with precipitation data from the reanalysis dataset although the model accuracy was worse with precipitation data from the reanalysis data than with observed precipitation data.