MODEL FEATURE ASSESSMENT OF CONVOLUTIONAL NEURAL NETWORK (CNN)-BASED FLOOD PREDICTIONS

Abstract

 Deep neural networks (DNN) with a deep learning approach have recently been applied to riverine flood models. We have developed the DNN-based flood model, in which a convolutional neural network (CNN) is embedded, to predict water levels during flood events with image classification techniques. Our model needed to have two major improvements for practical uses. First, lead time should be longer because our previous model provided only an-hour lead time prediction. Secondly, an effective creation of image-based inputs for CNN may have the appropriate spatial information optimized for watershed observatory networks. The former improvement provided appropriate model predictions of water level up to 3-h lead time. In addition, the simulation of a smooth flood wave was proper up to 6-h lead time. For the latter one, to reasonably predict water levels, the arrangements of stations for rainfalls and water levels based on the disctance from the predicted location to each station was better than the other arrangement such as the arrangement only of a few stations.