Advanced deep learning strategies for detection and quantification of macroplastics in rivers along the Peruvian coast

Abstract

Rivers are the primary contributors to plastic waste pollution entering the oceans, largely due to inadequate solid waste management, especially in the Global South. Macroplastics become difficult to remove from water bodies, and eventually fragment into smaller polymers, affecting wildlife and human health. However, methods for estimating these flows still face significant limitations. This study develops a methodological framework that incorporates artificial intelligence, particularly Deep Learning, to detect and classify eight classes of mixed inorganic municipal solid waste (MSW), with a focus on macroplastics present in rivers. This approach considers the spatial and temporal dynamics of the watercourse under study by using YOLOv11, a convolutional neural network model, by training and validating images captured by drones. A section of the river Rímac (Lima, Peru) was examined for one year. Results suggest that the YOLOv11 model is suitable for the rapid counting of certain macroplastic classes, such as tires, and black and colored bags. The model showed very high accuracy for tires (mAP = 0.94) in the testing stage, whereas for plastic bags values were above 0.74. Lower precision was identified for other categories, such as furniture and PET bottles due to debris size, abundance or chromatic contrast. Temporal changes in abundance were analyzed, with relevant changes observable between dry and wet seasons. This research validates the potential for establishing fieldwork projects covering larger areas to capture images of MSW mixes in rivers along the Peruvian coast, enabling future development of an automatic monitoring system.