A Self-Supervised Approach for the Detection and Classification of Turfgrass Abnormalities

The purpose of this study is to explore the feasibility of employing a self-supervised strategy to detect and classify various levels of abnormalities on turfgrass. We used a sliding-window approach to generate image tiles from RGB image data, which were then processed by a state-of-the-art vision foundation model to generate feature embeddings. The extracted features were clustered into groups, enabling RGB-based ranking for turfgrass anomalies and quality, and evaluated against ground truth derived from NDVI, a vegetation health index. DINOv2 achieved the highest weighted F1 score at 0.67, indicating superior performance in balancing precision and recall compared to ResNet50 (0.65) and EfficientNet (0.57). This workflow successfully created clusters of similar features in the RGB image data, facilitating the assessment of turfgrass conditions over time. These findings highlight the potential for practical applications in sports turf management and other agricultural settings, underscoring the benefits of combining standard camera data with self-supervised learning techniques to improve precision turfgrass management.