SAM2-Agri: Parameter-efficient fine-tuning of segment anything model 2 for symptomatic-leaf segmentation of sweet potato virus disease in field conditions
Physiological and Molecular Plant Pathology, cilt.145, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 145
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.pmpp.2026.103388
- Dergi Adı: Physiological and Molecular Plant Pathology
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, BIOSIS, Academic Search Ultimate (EBSCO)
- Anahtar Kelimeler: LoRA, Parameter-efficient fine-tuning, Plant disease detection, SAM2, Sweet potato virus disease
- Kayseri Üniversitesi Adresli: Evet
Özet
Sweet potato virus disease (SPVD), caused by co-infection of sweet potato chlorotic stunt virus and sweet potato feathery mottle virus, is the most damaging viral disease of sweet potato, and locating symptomatic leaves in field images is an important step toward automated monitoring. We present SAM2-Agri, a parameter-efficient adaptation of the Segment Anything Model 2 (SAM2) for segmenting SPVD-symptomatic sweet potato leaves on the SPVD-SEG benchmark. SAM2-Agri keeps the Hiera-S vision encoder almost entirely frozen (97.6%), injects Low-Rank Adaptation (LoRA) modules into its attention and MLP projections, and fully fine-tunes the lightweight mask decoder, updating only 12.77% of the parameters; it is trained with a composite Dice + Focal + IoU loss and bounding-box prompts. A fully automatic two-stage pipeline, in which a lightweight YOLOv8n detector supplies the prompt, segments symptomatic leaves at 96.39% mIoU without any ground truth at inference, rising to 98.46% mIoU when an oracle box is provided. On the same split, SAM2-Agri is competitive with pre-trained U-Net, DeepLabV3+, SegFormer and SAM2-UNet baselines, while remaining lightweight and running in real time on a single laptop GPU. These results indicate that parameter-efficient adaptation of foundation segmentation models is a practical, low-cost option for field-level SPVD leaf monitoring.