Research on prediction of Water Level of Low Frequency Flood on Rivers in Japan Using the Random Forest Method

In recent years, floods from heavy rainfalls have frequently occurred in many areas of Japan. In August 2016, widespread areas of Hokkaido suffered severe flood damage when four typhoons approached or passed over the region in succession. In this case, the successive typhoons came before the river water levels that had risen from the previous typhoon had fully decreased to normal levels. The soil moisture was high from the earlier rainfall, which made the river water prone to rise even when only small amounts of rain fell. Because of such weather and soil conditions, sufficiently accurate forecasting of river water levels was difficult for the August 2016 case. Embankment failures and inundation damage occurred along the Ishikari River which is in Western Hokkaido and along the Tokachi River which is in Eastern Hokkaido. In this study, we took the August 2016 case and forecast the water level with a 6-hour lead time (LT) and with a 12-hour LT for four locations: one at the lower reaches and the other at the middle reaches of the Ishikari River and the Tokachi River. Water level forecasting using the random forest (RF) method, which is a machine-learning method for determining the relative contribution of each explanatory variable, was used. As the result of water level forecasting using the RF method, a prediction model with a Nash-Sutcliffe coefficient of 0.70 or higher for all observation stations was developed. We also proposed a related-factor correlation method, which possesses a failsafe function in water level prediction when used by river management offices. In this method, the water level is predicted by creating a regression equation using the explanatory variables that are high in the order of contributing indexes, which are obtained by using the RF method. Water level forecasts given by the related-factor correlation method yielded nearly equal results to those given by the RF method. Water level forecasting using the related-factor correlation method is thought to be a practical forecasting method for river management offices, because the method enables timely and accurate forecasting at the time of disasters. Water level forecasts with an extended LT of 12 hours had considerable accuracy, and these can be used for avoiding evacuation during dark hours.