Scholarlink

<p>Pest infestation affects global agriculture, causing massive crop yield losses, ecosystem degradation, and negative economic impacts. High accuracy and efficiency in detecting small insect pests are crucial to avoid pesticide overuse and biodiversity loss. The objective of this research was to evaluate state-of-the-art object detection models (YOLOv7-v13) for multi-pest identification, focusing on small insect (&lt; 5% image area) in mango (fruit flies), maize (fall armyworm), and cotton crops (pink bollworm) for effective decision-making. We developed the Five-Pest dataset comprising 17,251 images captured by IoT-based smart traps and 194,050 pest instances of five economically significant species (three fruit fly species, fall armyworm, and pink bollworm). Images were annotated via Roboflow, then augmented and smoothed (<InlineEquation ID="IEq1"><EquationSource Format="TEX">\(\varepsilon = 0.1\)</EquationSource></InlineEquation>). YOLO models were trained under identical hyperparameters (40 epochs, batch 8, IoU = 0.8, max_det = 550). Performance was assessed by mAP@50, precision, recall, F1-score, and ROC-AUC. YOLOv9 achieved the highest mAP@50 (0.929), an average that includes lower precision such as fruit flies, underscoring the model’s robustness across varied pest types. YOLOv9 was followed by YOLOv8 (0.924), YOLOv12 (0.922), YOLOv10 (0.921), YOLOv11 (0.909), and YOLOv13 (0.909), while YOLOv7 achieved mAP@50 of 0.899 for the Five-Pest dataset. YOLOv12 demonstrated comparable performance to YOLOv8 and YOLOv10 with stable precision and recall, whereas YOLOv13 achieved competitive detection performance but required substantially higher training time, indicating the accuracy-computational cost trade-off in later YOLO generations. All models detected fall armyworm with very high accuracy (mAP 0.96–0.99), while fruit flies (especially <i>B. zonata</i>) remained comparatively more challenging. Pink bollworm recognition remained consistently strong across all variants (mAP 0.91–0.97). To improve per-class detection robustness, an ensemble model averaging YOLOv7-YOLOv13 predictions showing enhanced consistency but requiring more computation. Here we identified the strengths and limitations of YOLO variants for small-insect detection, guiding the selection of models to help reducing pesticide use and enhancing environmental protection 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

Evaluation of YOLOv7-v13 models for multi-class small insect pest detection using the five-pest dataset

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月20日

查看全文
  • Ayesha Hakim,
  • M. Habib Ur-Rahman,
  • Ali Hamza,
  • Munir P. Hoffmann,
  • Vakhtang Shelia,
  • Muhammad Owais,
  • Nimra Khan,
  • Muhammad Saim Ibtesam,
  • Muhammad Rashid,
  • Reimund P. Roetter

摘要

Pest infestation affects global agriculture, causing massive crop yield losses, ecosystem degradation, and negative economic impacts. High accuracy and efficiency in detecting small insect pests are crucial to avoid pesticide overuse and biodiversity loss. The objective of this research was to evaluate state-of-the-art object detection models (YOLOv7-v13) for multi-pest identification, focusing on small insect (< 5% image area) in mango (fruit flies), maize (fall armyworm), and cotton crops (pink bollworm) for effective decision-making. We developed the Five-Pest dataset comprising 17,251 images captured by IoT-based smart traps and 194,050 pest instances of five economically significant species (three fruit fly species, fall armyworm, and pink bollworm). Images were annotated via Roboflow, then augmented and smoothed (\(\varepsilon = 0.1\)). YOLO models were trained under identical hyperparameters (40 epochs, batch 8, IoU = 0.8, max_det = 550). Performance was assessed by mAP@50, precision, recall, F1-score, and ROC-AUC. YOLOv9 achieved the highest mAP@50 (0.929), an average that includes lower precision such as fruit flies, underscoring the model’s robustness across varied pest types. YOLOv9 was followed by YOLOv8 (0.924), YOLOv12 (0.922), YOLOv10 (0.921), YOLOv11 (0.909), and YOLOv13 (0.909), while YOLOv7 achieved mAP@50 of 0.899 for the Five-Pest dataset. YOLOv12 demonstrated comparable performance to YOLOv8 and YOLOv10 with stable precision and recall, whereas YOLOv13 achieved competitive detection performance but required substantially higher training time, indicating the accuracy-computational cost trade-off in later YOLO generations. All models detected fall armyworm with very high accuracy (mAP 0.96–0.99), while fruit flies (especially B. zonata) remained comparatively more challenging. Pink bollworm recognition remained consistently strong across all variants (mAP 0.91–0.97). To improve per-class detection robustness, an ensemble model averaging YOLOv7-YOLOv13 predictions showing enhanced consistency but requiring more computation. Here we identified the strengths and limitations of YOLO variants for small-insect detection, guiding the selection of models to help reducing pesticide use and enhancing environmental protection in precision agriculture.