Climate change, particularly increasing frequency and intensity of spring frost events, poses a serious threat to viticulture by reducing yield and product quality. This study proposes an image processing and machine learning-based framework for early, rapid, and accurate segmentation of frost damage in vineyards using YOLOv11s enhanced with Atrous Spatial Pyramid Pooling (ASPP). A unique dataset called FGVL dataset from Sultana seedless grape vineyards in Manisa, Türkiye, following a severe frost event in April 2025. FGVL includes 418 frost-damaged grapes, 510 frost-damaged leaves, 395 healthy grapes, and 698 healthy leaves, all manually annotated by experts under natural field conditions. By integrating ASPP into YOLOv11s, proposed model improved multi-scale contextual feature extraction and achieved mAP@50 of 0.7686, demonstrating stronger performance in instance segmentation of small, overlapping, and visually similar grapevine organs. In addition, Dynamic Confidence Thresholding (DCT) strategy was introduced to improve prediction reliability in dense and visually complex vineyard scenes. Despite challenges such as background clutter, object overlap, and small target structures, model maintained stable performance with low computational demand, requiring only 6.45 GB of GPU memory. Proposed framework offers an accurate, efficient, and practically deployable early recognition system for frost damage assessment in viticulture.
Arık et al. (Sat,) studied this question.