文献类型： 外文期刊

作者： Xiang, Yun ¹ ; Liang, Tian ¹ ; Bu, Yuanpeng ³ ; Cai, Shiqiang ¹ ; Guo, Jingjie ⁴ ; Su, Zhongjing ¹ ; Hu, Jinxuan ¹ ; Cai, Chang ¹ ; Wang, Bin ³ ; Feng, Zhijuan ³ ; Zhang, Guwen ³ ; Liu, Na ³ ; Gong, Yaming ³ ;

作者机构： 1.Zhejiang Univ Technol, BinJiang Inst Artificial Intelligence, Hangzhou 310051, Peoples R China

2.Zhejiang Univ Technol, Inst Cyberspace Secur, Hangzhou 310014, Peoples R China

3.Zhejiang Acad Agr Sci, Inst Vegetables, Key Lab Vegetable Legumes Germplasm Enhancement &, Minist Agr & Rural Affairs, Hangzhou 310021, Peoples R China

4.Zhejiang Shuren Univ, Coll Biol & Environm Engn, Key Lab Pollut Exposure & Hlth Intervent Zhejiang, Hangzhou 310015, Peoples R China

关键词： spectral range; herbicide phytotoxicity; early stress detection; deep learning; soybean-corn intercropping

期刊名称：AGRONOMY-BASEL （ 影响因子：3.4； 五年影响因子：3.8 ）

ISSN：

年卷期： 2025 年 15 卷 7 期

页码：

收录情况： SCI

摘要： Herbicide phytotoxicity represented a critical constraint on crop safety in soybean-corn intercropping systems, where early detection of herbicide stress is essential for implementing timely mitigation strategies to preserve yield potential. Current methodologies lack rapid, non-invasive approaches for early-stage prediction of herbicide-induced stress. To develop and validate a spectral-feature-based prediction model for herbicide concentration classification, we conducted a controlled experiment exposing three-leaf-stage vegetable soybean (Glycine max L.) seedlings to aqueous solutions containing three concentrations of nicosulfuron herbicide (0.5, 1, and 2 mL/L) alongside a water control. Hyperspectral imaging of randomly selected seedling leaves was systematically performed at 1, 3, 5, and 7 days post-treatment. We developed predictive models for herbicide phytotoxicity through advanced machine learning and deep learning frameworks. Key findings revealed that the ResNet-18 deep learning model achieved exceptional classification performance when analyzing the 386-1004 nm spectral range at day 7 post-treatment: 100% accuracy in binary classification (herbicide-treated vs. water control), 93.02% accuracy in three-class differentiation (water control, low/high concentration), and 86.53% accuracy in four-class discrimination across specific concentration gradients (0, 0.5, 1, 2 mL/L). Spectral analysis identified significant reflectance alterations between 518 and 690 nm through normalized reflectance and first-derivative transformations. Subsequent model optimization using this diagnostic spectral subrange maintained 100% binary classification accuracy while achieving 94.12% and 82.11% accuracy for three- and four-class recognition tasks, respectively. This investigation demonstrated the synergistic potential of hyperspectral imaging and deep learning for early herbicide stress detection in vegetable soybeans. Our findings established a novel methodological framework for pre-symptomatic stress diagnostics while demonstrating the technical feasibility of employing targeted spectral regions (518-690 nm) in field-ready real-time crop surveillance systems. Furthermore, these innovations offer significant potential for advancing precision agriculture in intercropping systems, specifically through refined herbicide application protocols and yield preservation via early-stage phytotoxicity mitigation.