Quantitative grading of hydroponic lettuce seedlings using an improved BlendMask framework with GNN-assisted canopy optimization

In hydroponic leafy vegetable production, variation in seedling size can affect transplanting efficiency and growth uniformity, ultimately reducing market quality. Although machine vision has been widely applied to non-destructive seedling monitoring and grading, accurate seedling segmentation and quantitative classification remain challenging because of severe leaf overlap and complex canopy structures in densely planted trays. To address these challenges, this study proposes a quantitative grading framework for hydroponic lettuce seedlings based on an improved BlendMask network and graph-neural-network-assisted canopy optimization. The framework employs a ResNet50 backbone integrated with deformable convolutions (DCNs) and a DeepLabV3+ decoder to enhance multi-scale feature representation and leaf boundary extraction. A graph-neural-network-based canopy optimization algorithm was further developed to model spatial relationships among canopy fragments, suppress over-segmentation caused by overlapping leaves, and reconstruct biologically consistent canopy structures. The 95% confidence interval of the canopy projection area was adopted as the grading criterion for whole-tray quantitative classification. Experimental results showed that the proposed BlendMask-R50 + DCNi3 configuration achieved a precision of 98.2% and an AP of 72.0. After canopy optimization, seedling counting accuracy reached 98.0%–100% under different occlusion conditions, while grading accuracy remained above 93%. Repeated experiments on hundreds of seedling trays, each containing approximately 200 planting positions, demonstrated that the complete grading process was completed within 2.67 s per tray. These results indicate that the proposed framework provides an efficient and practical approach for automated lettuce seedling counting, canopy reconstruction, and quantitative grading in plant-factory production systems.