Thermal Canopy Segmentation in Tomato Plants: A Novel Approach with Integration of YOLOv8-C and FastSAM

• Published in: Smart agricultural technology 2025-01, p.100806, Article 100806
• Main Authors: P, Hemamalini, MK, Chandraprakash, RH, Laxman, C, Rathinakumari, G, Senthil Kumaran, K, Suneetha
• Format: Article
• Language: English
• Subjects: Detection; Fast Segment Anything Model (SAM); Machine Learning; Segmentation; Thermal imaging; Tomato Plants; YOLOv8
• ISSN: 2772-3755; 2772-3755
• DOI: 10.1016/j.atech.2025.100806

•Holistic thermal canopy analysis, enabling stress assessment.•Compact YOLOv8-C model for faster object detection.•Integrating YOLOv8-C and FastSAM for whole canopy thermal image segmentation of tomato plants.•Masked ROI for precise temperature extraction, enabling targeted analysis of plant regions.•Enhanced understanding of plant stress physiology and precise monitoring of growth dynamics. Imaging has revolutionized plant studies, offering non-invasive insights into stress responses and growth dynamics. Traditional thermal imaging methods in plant studies have primarily focused on analyzing specific regions of interest rather than segmenting the entire canopy. This study introduces an innovative approach combining compact YOLOv8-C detection technology with the Fast Segment Anything Model (FastSAM). The compact YOLOv8-C model differs from the original YOLOv8l (large) model by simplifying the Neck architecture and reducing the number of convolutional and upsampling layers. It enables faster processing and maintains accuracy by achieving superior detection performance, with a mean average precision (mAP50) of 99.2%, mAP95 of 94.6%, precision of 99.3%, recall of 98.7%, and an F1 score of 99%. The methodological innovation emerges through using YOLOv8-C's bounding box outputs as refined input prompts for FastSAM, enabling sophisticated and precise canopy segmentation. Segmentation with FastSAM yielded an Intersection over Union (IoU) score of 92.28%, a Dice Similarity Coefficient (DSC) of 95.99%, and a Global Accuracy (GA) of 98.29%. Advancing this methodology, we introduce a masked region of interest (ROI) temperature extraction technique, enabling targeted temperature extraction of segmented plant regions. This integrated framework, combining YOLOv8-C's detection capabilities with FastSAM's segmentation, is collectively called TCSegNet (Thermal Canopy Segmentation Network). This novel approach leverages an innovative prompt selection strategy for comprehensive thermal canopy segmentation, marking a substantial advancement in plant imaging technology.