Evaluation of Deep Learning Models for Wheat Spike Segmentation Using Hyperspectral-derived Pseudo- RGB Images

Accurate and automated detection of wheat spikes is essential for high-throughput phenotyping and yield prediction, yet traditional manual counting is labor-intensive and error-prone. This study compared two deep learning models, U-Net and FasterViT, for wheat spike segmentation using pseudo-RGB images derived from hyperspectral data (400–1000 nm). A dataset of 400 wheat plants was collected at physiological maturity and annotated pseudo-RGB images were used for model training and testing. U-Net achieved a pixel accuracy of 0.893, a recall of 0.834, and a Dice score of 0.761. FasterViT outperformed U-Net with a pixel accuracy of 0.922, Intersection over Union (IoU) of 0.836, and a Dice score of 0.860, demonstrating better generalization and sharper segmentation of spikes. In terms of computational efficiency, U-Net required 2.5 seconds per image, whereas FasterViT required 6.85 seconds per image, reflecting a trade-off between speed and accuracy. Although the controlled dataset size was limited, the findings highlight the feasibility of low-resolution hyperspectral imagery for spike trait analysis. Future extensions could focus on field-based validation and integration into yield prediction pipelines to advance scalable precision agriculture.