Emulating Predictions of Spatiotemporal Inundation Depth Using Deep Neural Network

Abstract

The high computational cost can be a bottleneck for ensemble simulations by rainfall-runoff inundation models based on physical processes. To reduce the computational cost, this study has developed a deep neural network (DNN) emulator that can rapidly predict spatiotemporal distributions of inundation depth. The emulator uses the spatiotemporal distribution of precipitation as input data and predicts the spatiotemporal distribution of inundation depth for the same period. Here, the spatiotemporal distribution of inundation depth consists of three dimensions: event, spatial pattern, and time. Therefore, we conduct dimensionality reduction by singular value decomposition as a pre-processing step prior to train the DNN. To achieve this, it is necessary to transform the data from three variables to a matrix format, and we test two different transformation methods. We found that a transformation, which reduced the size of the DNN output more, improved the prediction accuracy. The developed DNN-based emulator showed accurate inundation predictions whose accuracy is equivalent to the one that outputs only the maximum inundation depth.