Feng et al. (2025) Field-deployable lightweight YOLOv8n for real-time detection and counting of Maize seedlings using UAV RGB imagery

Identification

• Journal: Frontiers in Plant Science
• Year: 2025
• Date: 2025-09-08
• Authors: Pengbo Feng, Zhigang Nie, Guang Li
• DOI: 10.3389/fpls.2025.1639533

Research Groups

• College of Information Science and Technology, Gansu Agricultural University, Lanzhou, China
• State Key Laboratory of Aridland Crop Science, Gansu Agricultural University, Lanzhou, Gansu, China
• Hexi University, Zhangye, Gansu, China

Short Summary

This study proposes YOLOv8-FLY, a lightweight deep learning model for real-time detection and counting of maize seedlings using UAV RGB imagery. The model achieves 96.5% detection accuracy while significantly reducing model size, parameters, and computational cost, making it suitable for resource-constrained edge devices.

Objective

• To develop a lightweight, real-time YOLOv8n-based maize seedling detection algorithm (YOLOv8-FLY) for UAV RGB imagery, addressing issues of large model size, high computational cost, and limited real-time performance in existing models, while maintaining high detection accuracy.

Study Configuration

• Spatial Scale: Experimental site in Huarui Ranch, Minle County, Gansu Province, China (38°44′3.32″N, 100°42′5.03″E; 1,683 m above sea level). UAV flight altitude of 3 m, resulting in a ground sampling distance (GSD) of approximately 0.07 cm/pixel. Images have a resolution of 5280 × 2970 pixels, with target maize seedlings typically ranging from 106 to 370 pixels in width and 59 to 297 pixels in height.
• Temporal Scale: Data collected on 3 and 4 May 2024, between 9:00 and 14:00, targeting maize seedlings at the 2-leaf and 1-center stage (approximately 20 days after sowing). Data collection for the experimental plot took approximately 45–60 minutes.

Methodology and Data

• Models used:
  • Base model: YOLOv8n
  • Proposed model: YOLOv8-FLY, incorporating:
    • Rep_HGBlock: A lightweight multi-scale backbone module designed by fusing RepConv with HGNetV2.
    • BiFPN (Bidirectional Feature Pyramid Network): Introduced into the neck network for enhanced multi-scale information fusion.
    • TDADH (Task Dynamically Aligned Detection Head): A lightweight detection head based on GroupNorm, shared convolution, and task-decoupled interaction mechanisms.
  • Visualization technique: Grad-CAM++ for model interpretability.
  • Comparison models: YOLOv3, YOLOv5s, YOLOv6, YOLOv8s, YOLOv8m, YOLOv8l, YOLOv8x, YOLOv9s, YOLOv10s, Faster R-CNN, SSD.
• Data sources:
  • Custom-built maize seedling dataset collected using a DJI Mavic Classic 3 drone with an RGB camera.
  • 1,213 original images (5280 × 2970 pixels) collected from a drip-irrigated field, with 993 high-quality images selected.
  • 21,974 manually annotated instances using LabelImg 1.8.6 in YOLO format.
  • Dataset split: 70% for training, 20% for validation, and 10% for testing.
  • Data augmentation techniques: Horizontal flip, vertical flip, 90-degree rotation, Gaussian noise, salt and pepper noise, and brightness adjustment, generating 65,600 augmented images (one representative image retained per original ID).

Main Results

• The YOLOv8-FLY model achieved a detection accuracy (mAP@0.5) of 96.5%.
• Model weight size was reduced to 3.5 MB, a 43% reduction compared to the original YOLOv8n (6.3 MB).
• The number of parameters was reduced to 1.58 M, a 47% reduction compared to the original YOLOv8n (3.01 M).
• Computational FLOPs were reduced to 7.4 G, an 8.6% reduction compared to the original YOLOv8n (8.1 G).
• The inference speed (FPS) was 146.3, a minor reduction of 2.4% from the original YOLOv8n (149.98 FPS).
• Grad-CAM++ visualization demonstrated that YOLOv8-FLY exhibited improved feature fusion and more accurate attention focusing on target objects compared to the original model.
• The developed real-time detection system, integrating YOLOv8-FLY, supports image and video stream processing, displaying real-time plant positions, target counts, and frame rates, and is compatible with live UAV camera input.

Contributions

• Proposed YOLOv8-FLY, a novel lightweight object detection model specifically tailored for real-time maize seedling monitoring using UAV RGB imagery in complex field environments.
• Designed Rep_HGBlock, a lightweight multi-scale backbone module, by fusing RepConv with HGNetV2 to enhance feature representation for small targets while reducing network complexity.
• Integrated BiFPN into the neck layer to improve multi-scale information fusion and reduce computational burden, enhancing robustness in UAV images with occlusions and lighting changes.
• Developed TDADH, a lightweight detection head based on GroupNorm, shared convolution, and task-decoupled interaction mechanisms, to achieve high detection capability with fewer parameters, facilitating edge device deployment.
• Constructed a comprehensive maize seedling dataset using UAV-acquired RGB imagery from a drip-irrigated field, specifically for small and densely planted targets.
• Developed and validated a practical real-time detection and counting system, demonstrating the model's deployability and efficiency for early-stage crop management and precision agriculture.
• Achieved a superior balance of detection accuracy, model size, computational cost, and inference speed compared to various mainstream lightweight detectors.

Funding

• 2024 Central Guided Local Science and Technology Development Funds Project (Grant No. 24ZYQA023)
• Gansu Provincial Industry Support Programme Project (2025CYZC-042)
• Gansu Provincial Key Science and Technology Special Project (24ZD13NA019)

Citation

@article{Feng2025Fielddeployable,
  author = {Feng, Pengbo and Nie, Zhigang and Li, Guang},
  title = {Field-deployable lightweight YOLOv8n for real-time detection and counting of Maize seedlings using UAV RGB imagery},
  journal = {Frontiers in Plant Science},
  year = {2025},
  doi = {10.3389/fpls.2025.1639533},
  url = {https://doi.org/10.3389/fpls.2025.1639533}
}