Abstract
INTRODUCTION: Accurate leaf area quantification is vital for early phenotyping in small-seeded crops such as broccoli (Brassica oleracea var. italica), where dense, overlapping, and irregular foliage makes traditional measurement methods inefficient. METHODS: This study presents YOLOv11-AreaNet, a lightweight instance segmentation model specifically designed for precise leaf area estimation in small-seeded broccoli seedlings. The model incorporates an EfficientNetV2 backbone, Focal Modulation, C2PSA-iRMB attention, LDConv, and CCFM modules, optimizing spatial sensitivity, multiscale fusion, and computational efficiency. A total of 6,192 germination-stage images were captured using a custom phenotyping system, from which 2,000 were selected and augmented to form a 5,000-image training set. Post-processing techniques-including morphological optimization, edge enhancement, and watershed segmentation-were employed to refine leaf boundaries and compute geometric area. RESULTS: Compared to the original YOLOv11 model, YOLOv11-AreaNet achieves comparable segmentation accuracy while significantly reducing the number of parameters by 57.4% (from 2.84M to 1.21M), floating point operations by 25.9% (from 10.4G to 7.7G), and model weight size by 51.7% (from 6.0MB to 2.9MB), enabling real-time deployment on edge devices. Quantitative validation against manual measurements showed high correlation (R² = 0.983), confirming the system's precision. Additionally, dynamic tracking revealed individual growth differences, with relative leaf area growth rates reaching up to 26.6% during early germination. DISCUSSION: YOLOv11-AreaNet offers a robust and scalable solution for automated leaf area measurement in small-seeded crops, supporting high-throughput screening and intelligent crop monitoring under real-world agricultural conditions.