摘要
[目的]北方寒地农业生产中普遍存在氮肥过量施用等不合理问题,造成了严重的资源浪费与土壤污染。为实现精准施肥,高光谱技术凭借其无损检测、快速响应的优势,可通过监测作物冠层氮素含量动态为变量施肥决策提供定量依据,已成为破解上述难题的核心技术支撑。但在实际应用中仍存在光谱数据获取难度大、质量要求高,数据迁移性差,模型的泛化能力和准确性低等问题。[方法]开发一种深度学习迁移模型用于水稻氮素高光谱反演与迁移学习。以寒地水稻为研究对象,采用手持式高光谱成像仪采集水稻高光谱数据。为实现模型在不同作物间的泛化能力,建立水稻-小麦跨作物数据集,进一步探究水稻与小麦生理规律共性。通过比较反向传播网络、随机森林和一维卷积神经网络3种回归模型在水稻氮素反演中的应用,发现1DCNN性能最佳。基于1DCNN网络结构,结合迁移学习与深度学习技术提出了Nitro-DTL模型,用于探究不同作物间的氮素预测精度。[结果]Ni⁃tro-DTL以小麦为源域的最佳预处理下的R^(2)较直接迁移学习提高59.39%。结合边缘分布自适应对深度迁移学习模型进一步优化,其协同模型(TCA-Nitro-DFTL)表现最优,小麦为源域时其最佳预处理下的R^(2)达到0.690,较直接迁移学习提升109.09%。[结论]该成果对于处理特征分布不同的数据集具有重要意义,为跨作物种类的精准农业实践提供了技术支持与理论参考。
[Objective]In the agricultural production of cold regions in northern China,there is a widespread issue of excessive nitrogen fertilizer application,leading to severe resource waste and soil pollution.[Methods]To achieve precise fertilization,hyperspectral technology,with its advantages of non-destructive detection and rapid response,can provide quantitative evidence for variable fertilization decisions by monitoring the dynamic nitrogen content in crop canopies.This has become the core technological support for solving these problems.However,in practical applications,challenges remain,such as the difficulty in obtaining spectral data,high quality requirements,poor data transferability,and low generalization ability and accuracy of models.In response to these issues,a deep learning transfer model for rice nitrogen hyperspectral inversion and transfer learning has been developed.Using handheld hyperspectral imagers to collect high-spectral data from cold-region rice,a cross-crop dataset combining rice and wheat was established to explore the common physiological patterns between rice and wheat.By comparing three regression models-back propagation network,random forest,and one-dimensional convolutional neural network-in rice nitrogen inversion,it was found that the 1DCNN performed best.Based on the 1DCNN architecture,combined with transfer learning and deep learning techniques,the Nitro-DTL model was proposed to investigate nitrogen prediction accuracy across different crops.[Results]The results from the cross-crop dataset show that under optimal preprocessing using wheat as the source domain,Nitro-DTL improves R^(2) by 59.39%compared to direct transfer learning.Combining edge distribution adaptive techniques to further optimize deep transfer learning models,the collaborative model(TCA-Nitro-DFTL)performs optimally.When wheat is the source domain,its best preprocessing achieves an R^(2) of 0.690,representing a 109.09%improvement over direct transfer learning.[Conclusion]This achievement is significant for handling datasets with different feature distributions,providing technical support and theoretical references for precision agriculture across crop types.
作者
回彦霖
王金峰
王瑞东
初宇航
HUI Yanin;WANG Jinfeng;WANG Ruidong;CHU Yuhang(Propaganda Department,Northeast Agricultural University,Harbin 150030,China;College of Engineering,Northeast Agricultural University,Harbin 150030,China)
出处
《沈阳农业大学学报》
北大核心
2025年第4期126-137,共12页
Journal of Shenyang Agricultural University
基金
黑龙江省自然科学基金重点项目(ZL2024E001)
国家重点研发计划项目(2024YFD2001103)。
关键词
水稻
氮素
高光谱
迁移学习
rice
nitrogen
hyperspectral
transfer learning
