画像解析 AI で河川を流れるプラスチックを遠隔監視、河川浮遊プラスチック輸送を定量化する新ソフトウェアを開発

Reliable and continuous plastic monitoring in rivers is essential for quantifying plastic flux and guiding mitigation efforts. One effective strategy for observing floating plastic transport is image-based monitoring using deep learning models. We developed river surface image monitoring software (RiSIM) to quantify floating macroplastic transport through three core processes: (1) a template matching algorithm, which identifies matching areas in a frame that resemble a template given in the previous frame; (2) deep learning models for plastic detection, classification, and object tracking; and (3) the evaluation of plastic transport rate in terms of both quantity and mass. The RiSIM-derived plastic transport rates were validated through a mark-release-recapture experiment and in-situ visual observation under both non-flood and flood conditions. The temporal variability and composition of the plastic transport rate in terms of quantity and mass were in good agreement with the ground truth data (r = 0.91 and 0.80, respectively). And also, it remained valuable for capturing the temporal variability in plastic transport rate (r = 0.87) via the comparison with in-situ visual observation, indicating that the RiSIM is valuable for assessing the increase in plastic transport rate due to a flood event. In fact, we found a significant relationship (r² = 0.36 for quantity; r² = 0.27 for mass) between daily-mean plastic transport rates and river discharge during flood events over four months. Accordingly, the RiSIM, as a near-field remote sensing technology, is a powerful tool for quantifying plastic transport and managing mis-managed plastic waste in river environments.