[4Rin1] Interactive 2

[4Rin1-42] AI for Flood Forecasting: Physics-based Deep Learning and Bias Correction by Data Assimilation for Improving Prediction Accuracy

〇Osamu Okada¹, Kazuhide Sasaki¹, Takahide Ochiai¹, Taeko Hirakuni¹ (1.Foundation of River and Basin Integrated Communications)

Keywords:Disaster Mitigation and Adaptation System, Time Series Data Analysis, Physics-based Deep Learning, Data Assimilation, Bias Correction

In the flood forecasting, there is a bias caused by the input data and the model, and the forecasted data and the observed data are assimilated using the data assimilation method. For example, the Kalman filter, particle filter, and three-dimensional variational method work well for linear phenomena, non-linear and biased phenomena that change smoothly compared to the target time scale. However, suddenly changing phenomena such as flood waveforms may not work. In this study, we propose a Neural Data Assimilation (NDA) method suitable for rapidly changing data features in flood forecasting. Our results show that data assimilation using a particle filter was conventionally considered to be the best method, but the proposed method is effective for improving the accuracy of flood forecasting and extending the lead time. This research is an attempt to connect AI and data assimilation fields, and is the first step to develop new data assimilation methods in various fields in the future.