Paper Details
Full metadata, abstract, citation, and access status.
Overcoming occlusion and overlapping in wheat spike detection: a yolov5s-based framework for accurate phenotypic evaluation
Abstract
In different stages of wheat breeding, timely and accurate counting of wheat spikes is essential for phenotypic evaluation and early yield prediction. Manual spike counting, however, is labor-intensive and time-consuming for wheat-producing regions globally. Aiming at the practical challenges including small spike size, dense distribution, and frequent spike occlusion, this study used the wheat cultivar Zhengmai 1860 as experimental material. Wheat spike images were acquired with mobile devices under potted conditions to build a custom image dataset, and a YOLOv5s object-detection model was developed for wheat-spike counting. The trained YOLOv5s model yielded a precision of 86.7%, recall of 94%, and mean average precision of 85.7%, showing favorable detection performance and generalization capacity. Compared with conventional manual counting, the proposed model delivers superior detection performance, enables rapid wheat-spike counting under potted conditions, and offers an effective technical option for automated wheat phenotypic assessment and yield-related estimation
