A Fully Automated YOLO-SAM Framework for High-Precision Microplastic Analysis in Environmental Sample

Abstract Microplastics have emerged as pervasive contaminants that threaten ecosystems and human health, yet their accurate identification in environmental samples remains a persistent challenge due to their diverse morphologies, translucency, and low optical contrast. Current deep learning approaches often rely on controlled laboratory conditions or manual intervention, which restricts scalability in real-world monitoring. Here, we propose a fully automated segmentation framework that integrates a YOLOv11 detector with a domain-adapted Segment Anything Model (SAM). The detector provides high-recall bounding box prompts, while the fine-tuned SAM performs pixel-level segmentation optimized for microplastic features.Using a dataset of 5,653 annotated microscopy images encompassing five morphological categories, the framework achieved a mean Intersection over Union (mIoU) of 0.777 in fully automated mode, surpassing DeepLabV3+ by more than 22%. When provided with manual prompts, the fine-tuned SAM reached an mIoU of 0.900, establishing a new benchmark in microplastic segmentation. Importantly, the system demonstrated robustness to bounding box perturbations and consistent performance across challenging particle types, including fibrous and semi-transparent structures.This end-to-end YOLO-SAM pipeline eliminates the need for manual annotation, offering a scalable, accurate, and robust solution for high-throughput microplastic analysis under environmentally relevant conditions. By bridging the gap between laboratory development and field deployment, this approach enables more efficient monitoring of microplastic pollution and supports the advancement of ecological risk assessment and regulatory strategies.