Scholarlink

<p>Weeds compete with crop plants for space, nutrients, sunlight, and soil moisture, reducing crop yields during early weeks after emergence. Controlling weeds in perennial crops like sugarcane is challenging and typically addressed by herbicides and mechanical tillage. This work focuses on weed detection in sugarcane crops. We provide an in-field dataset as a benchmark and evaluate deep learning architectures for object detection, classification, and conduct a bounding-box-guided segmentation study with both qualitative and quantitative evaluation. For detection, a 44.2 AP50 score was achieved combining RTMDeT, an architecture employing large-kernel depth-wise convolution, a loss function incorporating geometric constraints, and feature pyramid networks. For classification, we leveraged Swin Transformer with self-supervised pre-training achieving 99% accuracy. We compared segmentation performance among SAM, ExGR, and S2C approaches both qualitatively and quantitatively, using manually annotated pixel-level ground truth on the test set. Although significant progress was made, precisely detecting weeds in perennial crops remains unsolved under real-world conditions.</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

Toward autonomous weed management systems in sugarcane crops and an assessment of technological readiness

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月6日

查看全文
  • João P. Papa,
  • João R. R. Manesco,
  • Michael Schoder,
  • Cody Jacobucci,
  • Jiangpeng He,
  • Douglas J. Spaunhorst,
  • Richard M. Johnson,
  • Hermano I. Krebs

摘要

Weeds compete with crop plants for space, nutrients, sunlight, and soil moisture, reducing crop yields during early weeks after emergence. Controlling weeds in perennial crops like sugarcane is challenging and typically addressed by herbicides and mechanical tillage. This work focuses on weed detection in sugarcane crops. We provide an in-field dataset as a benchmark and evaluate deep learning architectures for object detection, classification, and conduct a bounding-box-guided segmentation study with both qualitative and quantitative evaluation. For detection, a 44.2 AP50 score was achieved combining RTMDeT, an architecture employing large-kernel depth-wise convolution, a loss function incorporating geometric constraints, and feature pyramid networks. For classification, we leveraged Swin Transformer with self-supervised pre-training achieving 99% accuracy. We compared segmentation performance among SAM, ExGR, and S2C approaches both qualitatively and quantitatively, using manually annotated pixel-level ground truth on the test set. Although significant progress was made, precisely detecting weeds in perennial crops remains unsolved under real-world conditions.