Applying deep learning methods for the bridge-based monitoring of floating macroplastics on rivers

The increasing plastic pollution in water bodies poses a serious threat to the environment. Rivers are a major pathway in the transport of macroplastic litter from source areas (e.g. urban areas) into the environment (e.g. beaches, shores, lakes and oceans). For this reason, quantifying and monitoring macroplastic loads in rivers is a key step in assessing pollution levels and developing effective preventive measures. In-situ monitoring is time-consuming and tedious, and it only covers a limited timeframe. The potential of Deep Learning (DL) methods to automatically detect objects in water bodies has recently been demonstrated. In this study we propose a framework to automatically monitor macroplastic loads in rivers using DL. The approach was evaluated on the River Rhine using various items (e.g. plastic bottles, caps, bags, polystyrene, tree branches with and without leaves) collected at the test site. An RGB camera was installed on Niederwerth Bridge (Koblenz, Germany) which captured images of the river at 1-second intervals. One boat was used to introduce the objects upstream and another boat to collect them again downstream. Our dataset consists of about 800 images with objects manually labelled in three groups: plastic bottles, plastic litter and vegetation. We employed the well-tested YOLOv5 network, pre-trained on the MS COCO dataset. Despite the limited amount of training data, the validation showed promising results with a mean average precision ([email protected]) of about 94%. The model can be improved by including more diverse training data from different rivers, in different environmental conditions (e.g. illumination, water turbidity, waves) and using a wider variety of objects.