OBSERVATION OF FLOATING DEBRIS IN RIVERS USING A DEEP LEARNING MODEL AND ANALYSIS OF RUNOFF CHARACTERISTICS

Abstract

 A deep learning model was constructed to continuously detect floating debris based on river surface images captured by an existing fixed-point camera. As a result, a relatively high detection accuracy (F-measure) of 0.79 was obtained for the number of pieces of debris detected, confirming the usefulness of the model. Using the developed model, we analyzed the transport characteristics of floating debris in the Onchi River in Osaka, and found that a first flush phenomenon was observed in the early stage of flooding, and that approximately 90 % of the total number of debris in a particular flood event was transported during the rising period. An estimation formula for the number of pieces of floating debris was constructed, taking into account the amount of precipitation and the number of preceding sunny days, and the annual transport was calculated. The contribution of rainy and sunny days to the annual transport was comparable in the river. It was considered that the installation of oil fences during normal river conditions, the implementation of riverbed cleanup activities prior to flooding, and the development of techniques to collect floating debris during flooding would be effective measures to reduce the amount of floating debris.