A sophisticated model for riverine-flood predictions using convolutional LSTM and transfer learning

Abstract

This study demonstrates that our new deep neural network model could predict riverine floods with high accuracy even in a target watershed with a small amount of data, using transfer learning and convolutional LSTM (ConvLSTM) that incorporates spatial information into LSTM. To predict some larger flood events, the transfer learning creates a pretrained model in a source watershed that has a large amount of data by adjusting the sequence length (Ns), which is a parameter of spatial information in ConvLSTM, and then transfers the pretrained model to the target watershed. First, the prediction accuracy of the model was verified in the source watershed and compared with a CNN-based conventional model. The improvement of the prediction accuracy was generally observed at Ns = 2. Secondly, for the verification of the prediction accuracy in the target watershed using transfer learning, the accuracy error of the new model was conversed till 50 retraining cycles, and the new model was compared with the conventional model with and without transfer learning. As a result, the accuracy of the new model prediction was improved at Ns=2.