Hydrological image building using curve number and prediction and evaluation of runoff through convolution neural network

Abstract

This study developed a runoff model using a convolution neural network (CNN), which had previously only been used for classification problems, to get away from artificial neural networks (ANNs) that have been extensively used for the development of runoff models, and to secure diversity and demonstrate the suitability of the model. For this model's input data, photographs typically used in the CNN model could not be used; due to the nature of the study, hydrological images reflecting effects such as watershed conditions and rainfall were required, which posed further difficulties. To address this, the method of a generating hydrological image using the curve number (CN) published by the Soil Conservation Service (SCS) was suggested in this study, and the hydrological images using CN were found to be sufficient as input data for the CNN model. Furthermore, this study was able to present a new application for the CN, which had been used only for estimating runoff. The model was trained and generalized stably overall, and R2, which indicates the relationship between the actual and predicted values, was relatively high at 0.82. The Pearson correlation coefficient, Nash-Sutcliffe efficiency (NSE), and root mean square error (RMSE), were 0.87, 0.60, and 16.20 m3/s, respectively, demonstrating a good overall model prediction performance.