Development of high-precision scum detection model for early warning of river scum

Abstract

In urban rivers where combined sewer overflow inflow occurs, scum forms after rainfall, causing prob- lems such as bad smells, deterioration of the landscape, and adverse effects on the ecosystem. As a coun- termeasure, early warning through monitoring of scum is desired. In this study, we examined the monitoring and image recognition requirements with the aim of detecting scum formation at an early stage, and then used data from a river monitoring camera on the Hirano River in Osaka Prefecture to experimentally clarify the image recognition model and loss function that match the characteristics of scum images. Based on the experimental results, we added images with low accuracy in detecting scum to the training data and con- ducted final training, and developed a high-precision scum detection model that satisfied the requirements by achieving Scum IoU95%.