Comparative Analysis of YOLO Architectures for Real-Time Debris Segmentation and Area Estimation

Accurate and real-time monitoring of debris is critical for effective environmental mitigation. While current models are proficient in 2D object detection, a key challenge remains in the real-time quantification of debris surface area. To address this, this paper presents a comparative analysis of five You Only Look Once (YOLO) architectures (YOLOv5s-seg, YOLOv6s-seg, YOLOv8s-seg, YOLOv11s-seg, and YOLOv12s-seg) to identify the most suitable model for estimating area of debris. Our evaluation reveals that YOLOv11s-seg provides the highest accuracy, achieving a leading mAP@50 of 0.21091 on the segmentation task and translating this to a superior area estimation accuracy of 97.03%. However, a distinct trade-off between accuracy and efficiency was identified, with YOLOv8sseg emerging as the most computationally efficient model, demonstrating the fastest inference speed on a CPU. These findings provide a clear recommendation: YOLOv11s-seg is the optimal choice for applications prioritizing maximum precision, while YOLOv8s-seg presents a compelling alternative for real-time systems where processing speed is the primary constraint. The proposed model offers a significant contribution to sustainable agriculture, particularly for water-based farming like seaweed, by enabling automated marine debris counting and establishing a foundation for future volume estimation.