Abstract
Citrus production faces significant economic threats from sunburn caused by extreme heat and solar radiation. Traditional identification methods often fail in complex orchard environments. This study introduces YOLOv8-Scm, a lightweight model based on YOLOv8n, optimized for real-time sunburn detection and grading. By incorporating Distribution-Shifted Convolution (DSConv), a Global Attention Mechanism (GAM), and the EIoU loss function, the model achieves a mAP50 of 92.7% and a Precision of 86.6%. These enhancements allow for superior performance in challenging natural scenarios with minimal computational overhead, providing a robust tool for smart orchard management.
Introduction
Citrus is a vital global crop, particularly in southern China. However, climate change has increased the frequency of "citrus sunburn," a physiological disorder that damages fruit skin and halts development. Automating the assessment of this damage is difficult due to environmental noise like variable lighting and leaf occlusion, the need for fine-grained severity grading, and the limited processing power of mobile devices used in orchards.
Methodology
The researchers developed YOLOv8-Scm by enhancing the YOLOv8n (nano) architecture with three specific components designed to balance efficiency and accuracy.
- DSConv (Distribution-Shifted Convolution): Replaces standard convolutions in the backbone to reduce memory and computational overhead while maintaining feature quality.
- GAM (Global Attention Mechanism): Captures channel and spatial contextual cues to improve detection of occluded or multi-scale targets.
- EIoU (Enhanced Intersection over Union) Loss: Improves bounding box regression, leading to more accurate localization of fruits across various sizes.
Data Collection and Classification
The study utilized a custom dataset of 2,099 high-resolution images of Orah mandarins, captured under diverse conditions to ensure model robustness.
The dataset includes 5,390 labeled instances categorized into a 5-level grading system:
- Hfruits: Healthy (pure green).
- ODB-fruit: Level 1 (slight yellowing).
- SDB-fruit: Level 2 (discoloration < 25%).
- TDB-fruit: Level 3 (burns covering 25%–66.7%).
- FDB-fruit: Level 4 (severe burns/scabs > 66.7%).
Results and Discussion
YOLOv8-Scm demonstrates significant improvements over the baseline and other state-of-the-art models:
- Performance: Achieved 92.7% mAP50, a 2.0% increase over YOLOv8n.
- Efficiency: The model uses only 3.2M parameters, maintaining an inference speed of 117.6 FPS on a standard GPU.
- Comparative Advantage: It outperformed SSD, Faster R-CNN, and YOLOv10n in terms of precision and recall for this specific agricultural task.
The inclusion of the Global Attention Mechanism was particularly effective in reducing false positives caused by complex backgrounds, such as mistaking green leaves for healthy fruit.
Conclusion
YOLOv8-Scm provides a highly efficient and accurate solution for monitoring citrus sunburn in real-time. Its lightweight design makes it ideal for deployment on mobile devices and drones, offering critical technical support for precision agriculture. Future work will focus on expanding the model's generalization to other citrus varieties and improving performance under extreme occlusion.
Summary updated