Scholarlink

<p>Accurate soybean seed counting is crucial for yield prediction, yet traditional image processing methods often fail due to their sensitivity to lighting, complex feature extraction, and instability. Salt stress, as a&#xa0;major abiotic factor limiting soybean yield, significantly affects the number of seeds. Therefore, the development of efficient seed counting methods is important for salt stress response assessment. The SPCNet, a&#xa0;seed phenotype counting network proposed in this study utilizes nighttime imaging under fixed light sources to reduce the impact of natural light variations, thereby accurately identifying and counting seeds under two salt concentration conditions. This provides data support for quantitatively assessing the effects of stress and screening salt-tolerant varieties. SPCNet innovatively integrates VGG19_BN and Transformer to enhance feature representation, while incorporating the SE module to refine channel attention across three feature scales. By introducing an overlap function and customized loss functions, the model simultaneously enables seed classification, regression, and counting. A&#xa0;unique point-merging strategy is applied to eliminate redundancy, further improving counting accuracy. Experimental results demonstrate SPCNet’s superior performance, with a&#xa0;MAE of 6.05, Relative MAE of 7.87, and R<sup>2</sup> of 0.92, outperforming existing models such as MCNN (MAE:60.63) and P2PNet (MAE:41.74). The model shows remarkable adaptability across different soybean varieties, with R<sup>2</sup> values ranging from 0.71 to 0.94. Regression analysis confirms a&#xa0;strong fit between actual and predicted values, validating SPCNet’s high accuracy. Principal component analysis indicates a&#xa0;strong correlation between salt concentration and seed count, with different soybean varieties exhibiting distinct responses. Notably, variety AFA2 demonstrates strong adaptability. In conclusion, SPCNet provides a&#xa0;novel and field-deployable solution for soybean pod enumeration, and an efficient and scalable technological pathway for conducting soybean phenotype monitoring and breeding screening under salt stress 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

SPCNet: an Intelligent Field-Based Soybean Seed Counting Algorithm for Salinity Stress Response Evaluation

  • 期刊论文

  • 发表日期：

    2025年9月1日

查看全文
  • Qiong Wu,
  • Fei Liu,
  • Zhongzhi Han,
  • Haoyu Wang,
  • Hang Liu,
  • Nan Xin,
  • Sijie Xiong,
  • Longgang Zhao

摘要

Accurate soybean seed counting is crucial for yield prediction, yet traditional image processing methods often fail due to their sensitivity to lighting, complex feature extraction, and instability. Salt stress, as a major abiotic factor limiting soybean yield, significantly affects the number of seeds. Therefore, the development of efficient seed counting methods is important for salt stress response assessment. The SPCNet, a seed phenotype counting network proposed in this study utilizes nighttime imaging under fixed light sources to reduce the impact of natural light variations, thereby accurately identifying and counting seeds under two salt concentration conditions. This provides data support for quantitatively assessing the effects of stress and screening salt-tolerant varieties. SPCNet innovatively integrates VGG19_BN and Transformer to enhance feature representation, while incorporating the SE module to refine channel attention across three feature scales. By introducing an overlap function and customized loss functions, the model simultaneously enables seed classification, regression, and counting. A unique point-merging strategy is applied to eliminate redundancy, further improving counting accuracy. Experimental results demonstrate SPCNet’s superior performance, with a MAE of 6.05, Relative MAE of 7.87, and R2 of 0.92, outperforming existing models such as MCNN (MAE:60.63) and P2PNet (MAE:41.74). The model shows remarkable adaptability across different soybean varieties, with R2 values ranging from 0.71 to 0.94. Regression analysis confirms a strong fit between actual and predicted values, validating SPCNet’s high accuracy. Principal component analysis indicates a strong correlation between salt concentration and seed count, with different soybean varieties exhibiting distinct responses. Notably, variety AFA2 demonstrates strong adaptability. In conclusion, SPCNet provides a novel and field-deployable solution for soybean pod enumeration, and an efficient and scalable technological pathway for conducting soybean phenotype monitoring and breeding screening under salt stress conditions.