A deep learning model for predicting river flood depth and extent

Abstract

This paper presents an innovative deep learning (DL) framework to (a) automatically identify river geometry and flood extent, and (b) predict river flooding depth. To do that, U-Net, an advanced convolutional neural network (CNN), was modified and given the designation of U-NetRiver. With the modification, the model received an input composite image with two bands of ground elevation and flooding discharge, and the output was water depth. The model was trained and validated based on the outputs from iRIC (a two-dimensional hydraulic model) for a segment of the Green River in the state of Utah. The results showed that the U-NetRiver could identify the river shape and wetted areas for flooded regions automatically. The maximum difference of predicted river depth obtained from U-NetRiver and the one obtained from the hydraulic model was 2.7 m. This result suggests a 29% improvement in prediction of the maximum flood depth in the river.