Automated leaf segmentation for robust phenotyping leveraging segmentation foundation models with weak supervision

Accurate plant phenotyping is essential for crop breeding and the development of health monitoring systems. Traditional phenotyping methods, such as visual observation and leaf counting, are inefficient and labor-intensive. Although automated approaches are more efficient, they still require extensive labeled datasets for training segmentation networks, which can be costly and time-consuming to prepare. This paper builds on Williams et al. and uses SAM for initial segmentation, then compares two approaches for selecting leaf segments: geometric and color-based filtering with automatically determined thresholds, and a lightweight CNN trained on minimal data. The CNN method delivers superior performance, with an Average Recall AR$_{75}$ of 63 % and an Average Precision AP$_{75}$ of 58 % (IoU threshold = 75 %) using only four training images, and minimal annotation work. These findings highlight the potential of combining SAM with CNN-based filtering for robust plant phenotyping applications, offering a scalable solution that makes advanced phenotyping more accessible and less dependent on extensive data preparation.