Scholarlink

Purpose <p>This study addresses the limitations of manual strawberry monitoring and the lack of greenhouse-tailored image acquisition systems. The proposed automated monitoring system, built on a custom-made rail-camera platform, was developed by evaluating recent YOLO variants and Detectron2. The system detects seven strawberry growth stages and provides a benchmark combining accuracy with hardware-specific metrics.</p> Methods <p>Images were collected during two different growing periods — early winter and later winter — in a controlled greenhouse environment. The early winter dataset was partitioned into three different subsets to assess model performance based on different splits. The late winter dataset was used to assess the generalization error of the trained models. Models were evaluated for detection accuracy, generalization, and computational efficiency.</p> Results <p>The results show that model performance varied depending on the dataset partitioning. Split-set 2 yielded higher mAP and F1-scores. YOLOv8 had the lowest generalization error of 0.126, while YOLOv7 had the highest generalization error of 0.243. Regarding computational efficiency, YOLOv5 achieved 136 FPS with 3% GPU utilization, while Detectron2 reached 46 FPS with 25%. Based on these comparisons, YOLOv8 was selected as the most suitable model, achieving an mAP of 0.666 and F1-score of 0.677 on the Early-Winter dataset with 108 FPS and a parameter size of 98.59&#xa0;MB.</p> Conclusion <p>Overall, these results suggest that data timing, partitioning, performance, and efficiency should be considered when developing automated strawberry monitoring systems, and they are expected to support the commercialization and further development of deep learning-based monitoring in precision agriculture.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Development and Evaluation of Automated Monitoring Systems for Multi-stage Strawberry Growth Detection Using Advanced Deep Learning Models

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Hyeonji Park,
  • Nala Shin,
  • Seungwoo Kum,
  • Sung Kyeom Kim,
  • Yoel Kim,
  • Seungwook Choi,
  • Do Yeon Won,
  • Hyun Kwon Suh

摘要

Purpose

This study addresses the limitations of manual strawberry monitoring and the lack of greenhouse-tailored image acquisition systems. The proposed automated monitoring system, built on a custom-made rail-camera platform, was developed by evaluating recent YOLO variants and Detectron2. The system detects seven strawberry growth stages and provides a benchmark combining accuracy with hardware-specific metrics.

Methods

Images were collected during two different growing periods — early winter and later winter — in a controlled greenhouse environment. The early winter dataset was partitioned into three different subsets to assess model performance based on different splits. The late winter dataset was used to assess the generalization error of the trained models. Models were evaluated for detection accuracy, generalization, and computational efficiency.

Results

The results show that model performance varied depending on the dataset partitioning. Split-set 2 yielded higher mAP and F1-scores. YOLOv8 had the lowest generalization error of 0.126, while YOLOv7 had the highest generalization error of 0.243. Regarding computational efficiency, YOLOv5 achieved 136 FPS with 3% GPU utilization, while Detectron2 reached 46 FPS with 25%. Based on these comparisons, YOLOv8 was selected as the most suitable model, achieving an mAP of 0.666 and F1-score of 0.677 on the Early-Winter dataset with 108 FPS and a parameter size of 98.59 MB.

Conclusion

Overall, these results suggest that data timing, partitioning, performance, and efficiency should be considered when developing automated strawberry monitoring systems, and they are expected to support the commercialization and further development of deep learning-based monitoring in precision agriculture.