Scholarlink

<p>Weed competition is one of the major biotic stresses that lead to the loss of plant resilience and lower crop productivity. This happens because weeds intensify competition for crucial resources such as water, nutrients, space, and sunlight. With the increasing climate changes and fewer resources, it is essential for agricultural systems to be stable and resilient by timely and correctly assessing weed-induced stresses. In this paper, we introduce an image-driven deep learning framework that combines plant ecological analysis and digital biotechnology to estimate the weed pressure and understand its impact on Proxy-Based Weed-Impact Assessment (PBWIA). The pipeline that we propose integrates two main tasks: weed-species classification with a ResNet50 model plus the addition of a Convolutional Block Attention Module (CBAM) as well as pixel-level weed segmentation by a light SegNet network. We extract from the segmented images biologically sensible indicators, such as weed-density estimation (WDE), weed-spatial entropy (WSE), and weed-growth-rate estimation (WGRE), which together depict the ecological force and spatial variation of the weed competition. The extracted features are subsequently integrated into a Proxy-Based Weed-Impact Assessment (PBWIA) measure and a Multi-Feature Composite Risk Index (MCRI) to assess relative weed-impact conditions and associated crop-risk levels. An XGBoost-based regression framework is further utilized for comparative consistency analysis and categorization of fields into low-, medium-, and high-risk groups, supporting image-driven weed-risk assessment. The framework is tested on several benchmark agricultural datasets such as MH, Weed16, Weed Image Detection, CWFID. The tests results demonstrate a good model generalization to different crop-weed environments. The experiments show that classification on MH, Weed16 reaches an accuracy of 99% while the R for the MCRI, based regression is higher than 0.91, which means our model has strong consistency analysis.</p><p><b>Clinical Trial Registration</b> Not applicable. This study does not report or involve any clinical trial, intervention study, or biomedical experimentation on humans or animals.</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

An image-driven deep learning framework for proxy-based weed-impact assessment and crop-risk analysis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月10日

查看全文
  • G. V. S. Narayana,
  • Murali Gopal Kakita,
  • Punyaban Patel,
  • Sanjay Kumar Kuanar,
  • Sudheer Babu Punuri,
  • Kranthi Kumar Singamaneni

摘要

Weed competition is one of the major biotic stresses that lead to the loss of plant resilience and lower crop productivity. This happens because weeds intensify competition for crucial resources such as water, nutrients, space, and sunlight. With the increasing climate changes and fewer resources, it is essential for agricultural systems to be stable and resilient by timely and correctly assessing weed-induced stresses. In this paper, we introduce an image-driven deep learning framework that combines plant ecological analysis and digital biotechnology to estimate the weed pressure and understand its impact on Proxy-Based Weed-Impact Assessment (PBWIA). The pipeline that we propose integrates two main tasks: weed-species classification with a ResNet50 model plus the addition of a Convolutional Block Attention Module (CBAM) as well as pixel-level weed segmentation by a light SegNet network. We extract from the segmented images biologically sensible indicators, such as weed-density estimation (WDE), weed-spatial entropy (WSE), and weed-growth-rate estimation (WGRE), which together depict the ecological force and spatial variation of the weed competition. The extracted features are subsequently integrated into a Proxy-Based Weed-Impact Assessment (PBWIA) measure and a Multi-Feature Composite Risk Index (MCRI) to assess relative weed-impact conditions and associated crop-risk levels. An XGBoost-based regression framework is further utilized for comparative consistency analysis and categorization of fields into low-, medium-, and high-risk groups, supporting image-driven weed-risk assessment. The framework is tested on several benchmark agricultural datasets such as MH, Weed16, Weed Image Detection, CWFID. The tests results demonstrate a good model generalization to different crop-weed environments. The experiments show that classification on MH, Weed16 reaches an accuracy of 99% while the R for the MCRI, based regression is higher than 0.91, which means our model has strong consistency analysis.

Clinical Trial Registration Not applicable. This study does not report or involve any clinical trial, intervention study, or biomedical experimentation on humans or animals.