畳み込みニューラルネットワークによるレーダ雨量を用いた河川水位予測

抄録

A river water level prediction model was proposed by applying Convolutional Neural Network using radar rainfall. Spatial distribution of hourly rainfall, accumulated rainfall and time series of water level or change of water level were used as input data. Model output is 60 minutes prediction of water level. 4 flood events out of 30, selected by peak water level, were used as validation target. NS coefficient was above 0.9 in all 4 flood events and it was confirmed that proposed model shows good accuracy. It was suggested that using both rainfall and water level results in better accuracy than using them separately. It was also suggested that setting too wide spatial range as input data causes loss of accuracy and it is better to set input data range with reference to the basin boundary.