A robust deep learning methodology to detect floating macro-plastic litter in rivers

Plastic pollution in rivers is a serious environmental concern. To improve the monitoring of floating macro-plastic litter in water, researchers increasingly resort to automatic detection tools based on Artificial Intelligence (AI) for Computer Vision (CV). The most advanced applications feature Deep Learning (DL) methods based on Convolutional Neural Networks (CNN) achieving state-of-the-art performances in standard CV datasets (e.g., ImageNet).

Despite promising initial results, only few studies validated the generalization ability of DL models across different locations, environmental conditions, and instrumental setups. Poor generalization results in the need for a new model for each different setting. This increases the data requirements and limits the applicability. These aspects are essential for practical implementations such as the development of a structural monitoring strategy backed by a reliable AI model.

In this work, we discuss how to develop a robust DL methodology by harnessing recent advancements in AI, such as data-centric AI and semi-supervised learning. We also show the effects of implementing these techniques on the generalization performances of a DL model by employing two different datasets of floating macro-plastic in rivers. The first is a new dataset recorded in a semi-controlled environment featuring a small drainage canal in the Netherlands; the second is a dataset available from the literature, with images from different waterways in Jakarta, Indonesia. The significant diversity among the two datasets grants a sound evaluation of model generalization performances and on the suitability of the proposed methodology for achieving increased robustness.