INVESTIGATION ON ACCURATE PREDICTION OF BASIN-WIDE RIVER WATER TEMPERATURE USING DEEP LEARNING

Abstract

 In this study, we developed a basin-wide river temperature prediction model for five first-class rivers in the Chugoku district, and investigated the change in model accuracy when rainfall was also used as an input layer, in order to realize highly accurate basin-wide river temperature prediction. Specifically, the model was compared under two conditions: one in which only temperature was input, and the other in which rainfall was input in addition to temperature. These were conducted under the condition that water temperatures at all locations in the watershed were predicted simultaneously using meteorological data at all locations in the watershed as input. As a result, the input of air temperature alone overestimated the water temperature by more than 5°C during the outflow period from June to August. Especially at the main river sites with large catchment areas, the input of rainfall also reduced the temperature to a maximum of 2°C and the mean absolute error rate (MAPE) was less than 10% at most of the target sites. These results show that using not only air temperature, but also rainfall in the input layer enables highly accurate river water temperature prediction on a basin-wide scale.