Your new experience awaits. Try the new design now and help us make it even better

ORIGINAL RESEARCH article

Front. Plant Sci.

Sec. Sustainable and Intelligent Phytoprotection

Volume 16 - 2025 | doi: 10.3389/fpls.2025.1682243

This article is part of the Research TopicInnovative Approaches in Remote Sensing for Precise Crop Yield Estimation: Advancements, Applications, and Future DirectionsView all 5 articles

YOLOv8-FDA: Lightweight Wheat Ear Detection and Counting in Drone Images Based on Improved YOLOv8

Provisionally accepted
  • Nanjing Forestry University, Nanjing, China

The final, formatted version of the article will be published soon.

Wheat is an important staple crop worldwide, and accurate detection and counting of wheat ears play a crucial role in crop yield prediction, phenotypic analysis, and field management. However, due to the complexity of field environments, lightweight and high-precision wheat ear detection and counting still faces significant challenges in practical applications. To tackle this issue, this study puts forward YOLOv8-FDA, a YOLOv8-based lightweight approach for wheat ear detection and counting. The method combines YOLOv8 for object detection, RFAConv for feature extraction, DySample for dynamic upsampling, HWD for wavelet-based downsampling, and the SDL loss function. First, the RFAConv module enhances the feature extraction ability of wheat ears, effectively solving the long-range dependency and adaptive spatial aggregation problems inherent in traditional convolutions. Second, the DySample module replaces the conventional upsampling module, optimizing multi-scale feature fusion and detail preservation while reducing computational redundancy. By introducing the HWD module, the model's parameters are further compressed, and training convergence is accelerated. Additionally, the SDL loss function is utilized to optimize the regression of bounding boxes, improving the model's localization accuracy in complex scenarios. Experimental results show that YOLOv8-FDA achieves a precision of 86.3%, recall of 77.5%, and mAP@0.5 of 84.9% on the GWHD dataset, representing improvements of 1.4%, 3.3%, and 2.3%, respectively, compared to the original YOLOv8n model. Meanwhile, the YOLOv8-FDA has a size of 2.96MB and a computational cost of 8.3GFLOPs, significantly reducing the overall model size. The model's real-time counting performance was also validated, achieving an average accuracy exceeding 97.5% with the cross-row segmentation counting method and operating at 19.2 frames per second, enabling stable real-time wheat ear counting. These results demonstrate the model's strong performance and suitability for deployment in real-world agricultural environments.

Keywords: wheat ear detection and counting, deep learning, Lightweight, HWD, YOLOv8-FDA

Received: 08 Aug 2025; Accepted: 28 Aug 2025.

Copyright: © 2025 Lin, Xiao and Lin. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

* Correspondence: Haifeng Lin, Nanjing Forestry University, Nanjing, China

Disclaimer: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.