Abstract Accurate segmentation of leaf area is a critical task in plant phenotyping and precision agriculture, as it directly impacts yield estimation, disease monitoring, and weed management. Conventional Convolutional Neural Networks (CNNs), such as UNet and its variants, often struggle with capturing long range contextual dependencies and preserving fine structural boundaries, while pure transformer based architectures like the Vision Transformer (ViT) suffer from poor inductive bias and limited data efficiency. To overcome these challenges , we propose a SegFormer inspired model that integrates Edge Gated Multi Head Spectral Attention (EG MHSA) for robust leaf area segmentation. The spectral attention mechanism captures discriminative frequency domain representations across spectral bands, while the edge gating module enhances boundary preservation by adaptively fusing multiscale edge features. Evaluated on the benchmark CWFID dataset, the proposed model achieves superior performance with an F1score of 97.33%, IoU of 95.84%, and the lowest loss of 0.0395, outperforming UNet variants and transformer based baselines. Qualitative analysis further demonstrates its effectiveness in accurately delineating fine leaf boundaries under complex field conditions. The ablation results highlight the complementary contributions of spectral attention and edge gating in boosting segmentation performance. With its lightweight architecture, edge focused refinement, and strong generalization capability, the proposed approach sets a new benchmark for leaf area segmentation and provides a practical, scalable solution for agricultural applications.
Banu et al. (Fri,) studied this question.