Semi-Supervised Learning Method for Basil Crop Images Segmentation and Growth Monitoring

Artificial Intelligence (AI) and Internet of Things (IoT) technologies are evolving traditional agriculture into the concept of Smart Agriculture (SA), enabling timely and effective decision-making. Nevertheless, several challenges persist. A relevant example is crop segmentation from agricultural images, which is particularly difficult because of foreground and background appearance, lighting, and camera conditions changing across time, making robust segmentation hard over long periods. To this end, this paper integrates IoT-based data collection, semi-supervised learning, and Digital Image Processing (DIP) techniques to segment basil cultivation images at the row level, across different growth stages, considering both day and night acquisitions. Using simple low-cost cameras to collect images, the proposed method applies multi-step preprocessing and color-guided GrabCut segmentation. In detail, standard rectangular GrabCut initialization is replaced by a foreground/background initialization derived from color cues and intensity priors, which is then used as a GrabCut mask, obtaining improved stability under shadows, glare, and background clutter. The method produces binary segmentation masks and overlays enabling continuous analysis of plant canopy coverage over time. When compared with other methods (e.g., Otsu thresholding and K-means clustering), the proposed approach achieves better segmentation performance, with a mean Intersection over Union (mIoU) of 0.96 and a Dice Coefficient of 0.98 across the dataset, demonstrating its robustness across varying illumination conditions and growth stages, and providing consistent segmentation for downstream analysis.