Advancing marine microplastic monitoring through deep learning-based image segmentation

Abstract Microplastic pollution poses significant threats to marine ecosystems and human health, requiring efficient and standardized monitoring within a One Health framework. This study presents the development and evaluation of an artificial intelligence (AI)-driven image segmentation model for detecting microplastics in images collected along the coastline of Sousse, Tunisia, a region particularly vulnerable to high pollution. The dataset includes 1080 images of plastics categorized into classes: high-density polyethylene (HDPE), low density polyethylene (LDPE), polyamide (PA), polyethylene terephthalate (PET), polypropylene (PP), and polystyrene (PS), as well as an "other" category comprising non-plastic and miscellaneous beach-collected items, such as cigarette butts and dried Posidonia oceanica . To ensure reliability with AI analysis, all fragments were first identified, photographed under controlled lighting to capture morphological variability. Using advanced polygon segmentation, the model enables pixel-level annotation and outperforms traditional bounding box methods, especially for irregularly shaped particles. Background subtraction and adaptive thresholds enhance accuracy, reducing false positives. Model performance is evaluated using the Intersection-over-Union (IoU) metric, measuring overlap between predicted segmentations and manually annotated ground truth data. Additionally, it discusses the ecological relevance of identified microplastic types in the regional context, highlighting potential impacts on coastal biodiversity and environmental health. Findings contribute to the growing field of marine pollution monitoring by demonstrating the practical utility and challenges of AI-powered segmentation, supporting future efforts toward scalable, ecologically informed, and standardized microplastic detection methods.