Scholarlink

Purpose <p>Smooth crabgrass (<i>Digitaria ischaemum</i>) is a problematic weed in turfgrass systems, posing significant challenges to turfgrass functionality and aesthetics. Herbicides are the most common method of weed control in turfgrass; however, their injudicious use can lead to ecological and environmental issues. Advances in computer vision and deep learning models show great promise for addressing some of these issues through targeted, site-specific weed control (SSWC). However, how model architecture and hyperparameter choices influence detection performance under varying field conditions and plant growth characteristics remains poorly understood.</p> Methods <p>This study evaluates the effect of mowing on detecting crabgrass in bermudagrass turf using YOLOv8 and YOLO11 models. Images were collected under diverse turfgrass management and mowing conditions using DSLR cameras (proximal) and Unmanned Aerial Vehicle-based platforms (airborne). Additionally, two annotation approaches, bounding box and polygon annotations were compared. Model performance was evaluated using metrics such as precision, recall, F1-score, mAP@0.50, and inference time.</p> Results <p>We found that freshly mowed turfgrass did not significantly affect the accuracy of weed detection compared to other conditions, whereas intermediate regrowth phases negatively affected model performance. With respect to annotation methods, bounding box annotations outperformed polygon annotations in detecting smooth crabgrass, achieving high F1-scores (0.87) and mAP@0.50 values (0.93), indicating that simplified bounding box annotations are sufficient for this application. The highest precision, recall, and mAP were recorded for YOLO11s (0.806), YOLOv8l (0.766), and YOLO11l (0.795), respectively. Further, smaller variants like YOLOv8n (2.2 ms) and YOLO11 (2.6 ms) demonstrated superior inference speeds, making them well-suited for real-time detection and robotic weed management applications. Moreover, the models trained on 100% proximal imagery significantly outperformed those that utilized 100% UAV imagery and mixed datasets, achieving the highest F1-score (0.85) and mAP@0.50 (0.92) values. The lower detection accuracies with the UAV imagery could be attributed to lower spatial resolution, background complexity, and scale variation.</p> Conclusions <p>Overall, this research highlights that crabgrass detection performance is strongly influenced by imaging methods, annotation strategies, and mowing conditions. The findings suggest that proximal imaging and bounding box annotations are more effective, while smaller YOLO variants offer an advantage for real-time deployment. These insights are critical for data collection and model selection to enhance the effectiveness of SSWC in turfgrass systems.</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

Effect of mowing, imaging technique, and annotation method on weed detection in turfgrass using YOLO

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Bholuram Gurjar,
  • Ubaldo Torres,
  • Guy Coleman,
  • Chase Straw,
  • Muthukumar Bagavathiannan

摘要

Purpose

Smooth crabgrass (Digitaria ischaemum) is a problematic weed in turfgrass systems, posing significant challenges to turfgrass functionality and aesthetics. Herbicides are the most common method of weed control in turfgrass; however, their injudicious use can lead to ecological and environmental issues. Advances in computer vision and deep learning models show great promise for addressing some of these issues through targeted, site-specific weed control (SSWC). However, how model architecture and hyperparameter choices influence detection performance under varying field conditions and plant growth characteristics remains poorly understood.

Methods

This study evaluates the effect of mowing on detecting crabgrass in bermudagrass turf using YOLOv8 and YOLO11 models. Images were collected under diverse turfgrass management and mowing conditions using DSLR cameras (proximal) and Unmanned Aerial Vehicle-based platforms (airborne). Additionally, two annotation approaches, bounding box and polygon annotations were compared. Model performance was evaluated using metrics such as precision, recall, F1-score, mAP@0.50, and inference time.

Results

We found that freshly mowed turfgrass did not significantly affect the accuracy of weed detection compared to other conditions, whereas intermediate regrowth phases negatively affected model performance. With respect to annotation methods, bounding box annotations outperformed polygon annotations in detecting smooth crabgrass, achieving high F1-scores (0.87) and mAP@0.50 values (0.93), indicating that simplified bounding box annotations are sufficient for this application. The highest precision, recall, and mAP were recorded for YOLO11s (0.806), YOLOv8l (0.766), and YOLO11l (0.795), respectively. Further, smaller variants like YOLOv8n (2.2 ms) and YOLO11 (2.6 ms) demonstrated superior inference speeds, making them well-suited for real-time detection and robotic weed management applications. Moreover, the models trained on 100% proximal imagery significantly outperformed those that utilized 100% UAV imagery and mixed datasets, achieving the highest F1-score (0.85) and mAP@0.50 (0.92) values. The lower detection accuracies with the UAV imagery could be attributed to lower spatial resolution, background complexity, and scale variation.

Conclusions

Overall, this research highlights that crabgrass detection performance is strongly influenced by imaging methods, annotation strategies, and mowing conditions. The findings suggest that proximal imaging and bounding box annotations are more effective, while smaller YOLO variants offer an advantage for real-time deployment. These insights are critical for data collection and model selection to enhance the effectiveness of SSWC in turfgrass systems.