5:15 PM - 7:15 PM
[MIS21-P06] Validation of image-based sensing of floating macro-plastic transport
Keywords:Macroplastics, Transport, River, Deep Learning
Quantifying macro-sized plastic debris (>25 mm) from land to the ocean is essential for revealing the global budget of plastic debris because mismanaged plastic waste found in the environment has leaked from land via rivers (Strokal et al., 2023 and Meijer et al., 2021). Especially, macroplastic transport on the water surface, which is its major pathway, is a major indicator for evaluating its export from rivers. In the present study, we discuss the applicability of quantifying floating plastic transport by applying a deep learning model to imagery obtained by a fixed camera.
We are developing a new methodology for quantifying the floating plastic transport in terms of number and mass by combining a deep learning model (Kataoka et al., 2024) and a template matching algorithm (Kataoka and Nihei, 2020). Our deep learning model can classify floating plastic transport into five major categories: drink bottles, food containers, shopping bags, other plastics, and non-plastics (Kataoka et al., 2024). This model has been retrained You Only Look Once version 8 (YOLOv8) with semantic segmentation extension, which is an instance segmentation that implements object detection and image segmentation architectures. Accordingly, object detection and image segmentation can evaluate the number and area of floating plastics from river surface images, respectively. Afterward, the floating plastic transport is evaluated via a template matching algorithm between two consecutive frames of river surface video. Finally, the mass transport can be evaluated by converting the plastic surface area to mass using the mean ratio of mass of each category to its area..
The applicability of our methodology for quantifying the transport rate in terms of both number and mass was validated through a mark-release-recapture experiment (MRRE). The MRRE was conducted from 10:00-11:30 on July 18, 2024, at the Ishite River, Ehime Prefecture, Japan. A stationary camera system was installed on a water pipe bridge over the river. The five major categories of floating plastic samples were released from upstream of the bridge, and then two surveyors collected the samples downstream. The float situation was captured by the stationary camera system. We quantified the number and mass transport rates by analyzing the river surface videos in the MRRE and then compared them with the ground truth from the MRRE. Consequently, the temporal variabilities in the number and mass transport rates quantified from the river surface videos were in good agreement with those of the ground truth (r = 0.92 and 0.62, respectively) (Fig. 1).
Ref) Strokal et al., Nat. Commun., 2023; Meijer et al., Sci. Adv., 2021; Kataoka et al., Front. Earth. Sci., 2024
We are developing a new methodology for quantifying the floating plastic transport in terms of number and mass by combining a deep learning model (Kataoka et al., 2024) and a template matching algorithm (Kataoka and Nihei, 2020). Our deep learning model can classify floating plastic transport into five major categories: drink bottles, food containers, shopping bags, other plastics, and non-plastics (Kataoka et al., 2024). This model has been retrained You Only Look Once version 8 (YOLOv8) with semantic segmentation extension, which is an instance segmentation that implements object detection and image segmentation architectures. Accordingly, object detection and image segmentation can evaluate the number and area of floating plastics from river surface images, respectively. Afterward, the floating plastic transport is evaluated via a template matching algorithm between two consecutive frames of river surface video. Finally, the mass transport can be evaluated by converting the plastic surface area to mass using the mean ratio of mass of each category to its area..
The applicability of our methodology for quantifying the transport rate in terms of both number and mass was validated through a mark-release-recapture experiment (MRRE). The MRRE was conducted from 10:00-11:30 on July 18, 2024, at the Ishite River, Ehime Prefecture, Japan. A stationary camera system was installed on a water pipe bridge over the river. The five major categories of floating plastic samples were released from upstream of the bridge, and then two surveyors collected the samples downstream. The float situation was captured by the stationary camera system. We quantified the number and mass transport rates by analyzing the river surface videos in the MRRE and then compared them with the ground truth from the MRRE. Consequently, the temporal variabilities in the number and mass transport rates quantified from the river surface videos were in good agreement with those of the ground truth (r = 0.92 and 0.62, respectively) (Fig. 1).
Ref) Strokal et al., Nat. Commun., 2023; Meijer et al., Sci. Adv., 2021; Kataoka et al., Front. Earth. Sci., 2024