Deep Learning based Mass-Flux Estimation of floating (plastic) litter in waterways
F. Boccacci (TU Delft - Civil Engineering & Geosciences)
R. Taormina – Mentor (TU Delft - Civil Engineering & Geosciences)
M. Hrachowitz – Graduation committee member (TU Delft - Civil Engineering & Geosciences)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Plastic debris in aquatic environments has become an increasing concern for wildlife and human health. Rivers and canals serve as major sources of plastic transport to the oceans, yet global estimates of plastic flux remain highly uncertain, partly due to non-uniform and sparse field measurements. This poses a significant challenge for effective mitigation strategies, as the contribution of individual streams can vary by multiple orders of magnitude.
This thesis develops and tests deep-learning models that use images from fixed cameras to detect and classify floating litter based on material types. Following these detections, a conversion into a mass estimate is made based on the size of the litter object. A combination of real-world datasets and syn- thetically generated images was used to train a YOLOv11 architecture for a single-class task and two multi-class setups. The performance and generalization capability was evaluated on in-domain test sets and out-of-domain locations, while mass estimates were validated against physical sampled litter at two locations.
The results indicate robust performance for single-class models, while the introduction of material classes caused a performance decrease of 12% in mAP@50-95, due to the visual ambiguity of heterogeneous plastic types. The integration of synthetic data improved generalization to unseen locations through higher recall, albeit at the cost of higher false positive rates. Field validation showed that mass estimates based on traditional methods, such as human visual counting, can cause an uncertainty of up to an order of magnitude. Object detection models tended to underestimate the total mass due to heavy outliers in the ground-truth and a detection bias towards plastic categories. However, the estimate at the urban site was within 20% of the recovered plastic load. Overall, results indicate that deep learning can provide conservative and reviewable estimates of plastic mass flux from camera data. Detailed material classification is feasible for visually distinct categories, but remains data-limited for more heterogeneous materials.