Work

Extracting leaf morphological traits within occluded lettuce canopies using instance segmentation and supervised conditional GAN

Year: 2025

Type: article

Abstract: Precise measurement of leaf morphology in densely occluded canopies is critical for optimizing plant growth management and enhancing crop yield in agricultural systems. Leaf completion under occlusion is key to improving phenotypic measurement accuracy. Although RGB-based conditional generative adversarial networks (CGANs) approaches offer cost-effective and efficient leaf completion, they struggle to capture individual leaf characteristics, establish accurate pairwise correspondences, and quantify occlusion impact. Using butterhead lettuce as research subject, this study presents an automated pipeline that integrates instance segmentation for leaf extraction with a supervised CGAN framework for leaf completion. We developed a data acquisition and image processing workflow to establish in vivo–ex vivo correspondences for training the supervised CGANs. Comprehensive evaluation identified YOLOv8s-Seg as the optimal segmentation model and pix2pix as the most effective completion network. For leaf completion, pix2pix achieved an R 2 of 0.948 and an RMSE of 2.851 cm 2 for leaf area estimation, and a SAMScore (indicating semantic similarity) of 0.9744, demonstrating accurate reconstruction of leaf morphology from external contours and internal anatomical features. Furthermore, the findings demonstrate the advantages of supervised CGANs utilizing exclusively paired training data against unsupervised CGANs using unpaired data and uncover an ascend-descend performance pattern in leaf completion accuracy relative to completeness levels, with optimal performance achieved at approximately 60% leaf completeness. Our automated, cost-effective and highly scalable pipeline holds promise to become a new paradigm for intelligent plant monitoring and management in modern agricultural systems and occlusion analysis in phenotyping research. • Providing a reference for paired plant dataset construction to train supervised CGANs. • Assessing various models to identify the optimal approach for leaf segmentation and completion. • Confirming superiority of supervised CGANs relative to unsupervised CGANs in leaf completion. • Uncovering how occlusion severity and distribution shape the completion quality.

Source: Journal of Agriculture and Food Research

DOI: 10.1016/j.jafr.2025.102518