文献类型： 外文期刊

作者： Sun, Guangyao ¹ ; Zhang, Yong ² ; Wang, Lei ² ; Zhou, Longyu ³ ; Fei, Shuaipeng ³ ; Han, Shiteng ¹ ; Xiao, Shunfu ³ ; Che, Yingpu ⁵ ; Yan, Long ⁴ ; Xu, Yun ¹ ; Li, Yinghui ⁵ ; Qiu, Lijuan ⁵ ; Ma, Yuntao ¹ ;

作者机构： 1.China Agr Univ, Coll Informat & Elect Engn, East Campus, Beijing, Peoples R China

2.Heilongjiang Acad Agr Sci, Keshan Branch, Qiqihar, Peoples R China

3.China Agr Univ, Coll Land Sci & Technol, Beijing 100193, Peoples R China

4.Hebei Acad Agr & Forestry Sci, Inst Cereal & Oil Crops, Shijiazhuang 050035, Hebei, Peoples R China

5.Chinese Acad Agr Sci, Inst Crop Sci, Beijing 100081, Peoples R China

关键词： Unmanned Aerial Vehicle; Deep Learning; Resnet; Soybean Yield and Lodging Prediction

期刊名称：COMPUTERS AND ELECTRONICS IN AGRICULTURE （ 影响因子：8.9； 五年影响因子：9.3 ）

ISSN： 0168-1699

年卷期： 2025 年 230 卷

页码：

收录情况： SCI

摘要： Yield and lodging are crucial indicators in soybean breeding. The development of unmanned aerial vehicle (UAV) equipped with hyperspectral imaging technologies provides high-throughput data for estimating these factors. Previous studies have primarily focused on using hand-crafted band reflectance information, vegetation indices, and texture features to construct empirical models for yield and lodging estimation. However, few studies have directly employed deep learning techniques to automatically extract features from raw hyperspectral images in this context. The objectives were to investigate the potential of combining hyperspectral images with deep learning for soybean yield and lodging prediction, and to avoid the complex process of traditional feature extraction. A novel Prototype Contrastive Learning (PCL) network was proposed to learn representations from raw images. For comparison, hand-crafted vegetation indices and texture features, selected for their effectiveness in crop growth monitoring, were extracted and input into the same machine learning model. The impact of different growth stages on yield and lodging prediction was then investigated. Results demonstrated that the PCL network can effectively capture the similarities within the same class and the differences between different classes. The PCL representations exhibited more distinct clusters according to class labels compared to handcrafted features. At 86 days after emergence (DAE), the PCL method achieved optimal yield prediction accuracy (R2 = 0.65, RMSE = 507.56 kg/ha) and was significantly higher than the hand-crafted features method (R2 = 0.55, RMSE = 581.37 kg/ha). The highest performance of lodging grades classification was achieved at 65 DAE, and the PCL representations (F1-score = 0.80) achieved a 48 % accuracy improvement compared to handcrafted features (F1-score = 0.54). This study pioneered the use of deep learning for automatic hyperspectral feature extraction in real-world breeding scenarios, providing valuable insights and strategies to improve yield prediction and lodging classification, thereby more effectively supporting soybean breeding and field management.