A prototype-augmented few-shot segment anything model for field phenotyping of lettuce germplasm from UAV imagery

Accurate phenotypic acquisition determines the effectiveness of germplasm screening, while the efficiency and accuracy of image segmentation directly affect the quality of phenotypic parameter extraction. While foundation models like Segment Anything Model 2 (SAM2) have achieved remarkable success in general tasks, their performance in complex field environments is often limited by a lack of crop-specific visual priors. To achieve high-precision segmentation with minimal data, few-shot learning has gained traction as a strategic alternative to manual-intensive training; however, existing approaches still struggle with weak feature generalization and imprecise boundary delineation when facing highly heterogeneous lettuce germplasm. To address these issues, this study proposes a segmentation model that integrates Prototype Feature-Enhanced Prompting and a Complex Gated Spatial-Channel Attention Mechanism to improve lettuce image segmentation accuracy under complex conditions. Specifically, a Prototype Feature-Enhanced Prompting Module (PFEM) is designed to retrieve historical visual prototypes from an external feature library and fuse them into prompt embeddings, thereby introducing instance-level priors. In addition, a Complex Gated Spatial-Channel Attention Module (CGSC) is embedded in the decoder to enhance boundary perception and key feature awareness. Finally, a Composite Loss Function (CLF) is developed to jointly optimize pixel classification and region overlap. Experimental results showed that, under the 1-shot evaluation setting on a self-built open-field lettuce dataset, the proposed model achieved an accuracy of 99.57%, a precision of 99.17%, a recall of 96.63%, a Dice coefficient of 97.90%, an IoU of 95.85%, and a Boundary IoU of 94.71%. Compared with the evaluated methods, these metrics were higher by 1.16-15.08, 1.16-16.03, 1.21-15.68, 1.56-16.69, 0.72-14.97, and 2.26-14.99 percentage points, respectively. Based on these high-precision segmentation results, fifteen morphological, color, and texture phenotypic parameters were automatically extracted. The proposed method provides a high-precision, low-sample-dependence solution for high-throughput plant phenotyping, offering robust technical support for germplasm screening.