Research on detection and counting method of wheat ears in the field based on YOLOv11-EDS

Jing Li, Zhenchang Wang, Xiaoqing Luo, Bo Feng, Kuijie Gong, Xia Zhang, Zheng JiyeFrontiers in Plant Science 2025 · DOI

TL;DR

Researchers have developed YOLOv11-EDS, a lightweight deep learning model specifically optimized for detecting and counting wheat ears in complex field environments. By integrating dynamic upsampling and direction-aware attention, the model achieves higher accuracy than standard YOLOv11 while reducing the computational footprint, making it ideal for deployment on mobile agricultural devices.

What Problem Does This Paper Address?

Estimating crop yield is a cornerstone of smart agriculture, and for wheat, this starts with counting "ears" (the grain-bearing tips). However, doing this automatically in a field is surprisingly difficult. Wheat ears are often densely packed, overlap one another, and vary wildly in size and orientation. Furthermore, they often blend into the background of stalks and leaves.

While existing object detection models (like previous YOLO versions) are powerful, they often struggle with:

  1. Small-scale features: Tiny or distant wheat ears get lost during the image upsampling process.
  2. Geometric complexity: Wheat ears grow in long strips at various angles, which standard "square-focused" attention mechanisms don't capture well.
  3. Computational cost: High-accuracy models are often too "heavy" to run on the handheld devices or drones that farmers actually use in the field.

Methods

The authors propose YOLOv11-EDS, which introduces three key architectural enhancements to the baseline YOLOv11 framework:

  • DySample (Dynamic Upsampling): Instead of using traditional static upsampling (like bilinear interpolation) which can blur details, DySample learns sampling offsets based on the image content. This preserves the sharp details of small wheat ears as the feature maps are enlarged.
  • OrientedECA (Direction-aware Oriented Efficient Channel Attention): Wheat ears aren't just blobs; they have a distinct "long strip" morphology and often grow in rows. This custom attention mechanism performs horizontal and vertical pooling to capture direction-sensitive features (like growth angles and row spacing), helping the model distinguish wheat ears from similar-looking stalks.
  • Slim-Neck (GSConv): To make the model "lightweight," the authors replaced standard feature fusion layers with a "Slim-Neck" structure using GSConv. This method reduces redundant calculations and lowers the parameter count without sacrificing the model's ability to process features at different scales.

Datasets

The study utilized two primary data sources:

  1. Global Wheat Head Detection (GWHD) 2021: A benchmark dataset. The authors selected a subset of 1,140 images (containing over 40,000 annotated ears) to ensure visual diversity in density, scale, and lighting.
  2. Self-built iPhone-perspective Dataset: To test real-world applicability, the team collected 56 high-resolution images using an iPhone at a 15° depression angle. This dataset captured wheat at different growth stages (milk stage and ripening stage) to verify environmental adaptability.

Key Insights and Ideas

  • Geometry Matters: One of the most insightful aspects of this paper is the move toward "direction-aware" AI. By acknowledging that wheat ears have a specific orientation and growth pattern, the researchers were able to suppress background noise (like vertical stalks) that typically confuses general-purpose detectors.
  • Efficiency vs. Accuracy is Not a Zero-Sum Game: Often, making a model "lightweight" hurts performance. Here, the specialized modules (DySample and OrientedECA) actually improved accuracy while the Slim-Neck reduced the model's size.
  • Robustness Across Growth Stages: The model remained effective even as wheat changed color and texture from the "grouting" (green) stage to the "maturity" (golden) stage, suggesting it can be used throughout the harvest season.

Key Results

The YOLOv11-EDS model outperformed the baseline YOLOv11 and several other mainstream models (including Faster-RCNN, SSD, and YOLOv8):

  • Accuracy: Precision increased by 2.0%, Recall by 3.5%, and [email protected] (mean Average Precision) reached 96.8%.
  • Efficiency: The model parameters were reduced to just 2.5 Million, and floating-point operations (FLOPs) dropped to 5.8G—both lower than the baseline YOLOv11.
  • Counting Performance: In low-light and blurry conditions, YOLOv11-EDS showed the closest match to manual ground-truth counts, significantly reducing the "missed detections" common in other models.

Limitations and Future Work

While highly effective, the authors noted that the model still struggles with extreme occlusion—specifically when more than 60% of a wheat ear is hidden by leaves.

Future research directions include:

  • Multimodal Fusion: Combining RGB images with other data types (like thermal or multispectral data) to "see" through dense foliage.
  • Extreme Occlusion Handling: Developing specific sub-modules to infer the presence of partially hidden ears.
  • Broader Deployment: Further optimizing the code for seamless integration into mobile apps and autonomous agricultural robots.

Summary updated

Abstract

This study provides an efficient solution for the automated analysis of wheat phenotypic parameters in complex farmland environments, which is of great value for promoting the development of smart agriculture.
BibTeX
@inproceedings{li2025research,
  author = {Jing Li and Zhenchang Wang and Xiaoqing Luo and Bo Feng and Kuijie Gong and Xia Zhang and Zheng Jiye},
  title = {{Research on detection and counting method of wheat ears in the field based on YOLOv11-EDS}},
  booktitle = {Frontiers in Plant Science},
  year = {2025},
  doi = {10.3389/fpls.2025.1672425},
}
Figure