摘要
【目的】设计一种结合高光谱遥感、卷积神经网络和迁移学习技术的田间水稻秸秆覆盖度(RRC)信息提取方法。【方法】1)实验室测量了多种土壤含水量、水稻秸秆含水量和RRC的“土壤-水稻秸秆”混合光谱。2)基于视觉几何组网络设计了水稻秸秆覆盖度高光谱网络(RRChyperNet)模型,该模型融合深层和浅层网络特征开展RRC估算。3)基于实验室实测和田间实测“土壤-水稻秸秆”混合光谱评估了使用RRChyperNet模型开展田间RRC估算的可行性。使用决定系数(R^(2))和均方根误差(RMSE)评估田间RRC信息提取的准确性。【结果】1)RRChyperNet可用于开展田间RRC信息高精度估算(R^(2)=0.953,RMSE=0.085)。2)基于预训练的RRChyperNet结合迁移学习方法能够实现对研究区RRC高精度估算(R^(2)=0.867,RMSE=0.093),其精度明显高于广泛使用的随机森林和支持向量机回归模型(R^(2)=0.686~0.691,RMSE=0.122~0.128)。3)研究仅基于水稻秸秆开展了RRChyperNet模型训练和性能测试,其针对小麦、玉米等秸秆覆盖度信息提取精度仍需要未来更多的试验来验证。【结论】RRChyperNet能够提供高精度的田间水稻秸秆覆盖度信息,为动态掌握农田保护性耕作实施进度和落实农业生态环境保护建设提供技术支持。
【Objective】A method for extracting rice residue cover(RRC)information in the field was designed by combining hyperspectral remote sensing,convolutional neural network and transfer learning.【Method】The work mainly includes three parts:1)Various soil moisture content,rice residue moisture content and RRC“soil-rice straw”mixed spectrum were measured in the laboratory;2)A rice residue cover Hyperspectral network(RRChyperNet)model was designed based on the visual geometry group network,which combined the features of deep and shallow networks to carry out RRC estimation;3)The feasibility of using RRChyperNet model for field RRC estimation was evaluated based on laboratory and field measurements of soil-rice straw spectra.Coefficient of determination(R2)and root mean square error(RMSE)were used to evaluate the accuracy of RRC information extraction in the field.【Result】1)The RRChyperNet can be used to estimate field RRC information with high precision(R^(2)=0.953,RMSE=0.085);2)The RRChyperNet based on pre-training combined with transfer learning method can achieve a highly accurate estimation of RRC in the study area(R^(2)=0.867,RMSE=0.093)and its accuracy is significantly higher than that of the widely used random forest and support vector machine regression models(R^(2)=0.686-0.691,RMSE=0.122-0.128);3)This study only conducted RRChyperNet model training and performance testing based on rice residue dataset;The estimation potential of RRChyperNet for wheat,corn,and other straw types still requires further experimentation in the future.【Conclusion】RRChyperNet model can provide high-precision rice straw coverage information in the field,and provide technical support for dynamically grasping the implementation progress of farmland conservation tillage and implementing the construction of agricultural ecological environment protection.
作者
岳继博
李婷
宋洁
田庆久
刘杨
冯海宽
YUE Jibo;LI Ting;SONG Jie;TIAN Qingjiu;LIU Yang;FENG Haikuan(College of Information and Management Science,Henan Agricultural University,Zhengzhou 450046,China;School of Foreign Studies,Henan Agricultural University,Zhengzhou 450046,China;International Institute for Earth System Science,Nanjing University,Nanjing 210023,China;Key Lab of Smart Agriculture System,Ministry of Education,China Agricultural University,Beijing 100083,China;Key Laboratory of Agricultural Remote Sensing Mechanism and Quantitative Remote Sensing,Ministry of Agriculture and Rural Affairs,Information Technology Research Center,Beijing Academy of Agriculture and Forestry Sciences,Beijing 100097,China)
出处
《河南农业大学学报》
CAS
CSCD
北大核心
2024年第5期838-851,共14页
Journal of Henan Agricultural University
基金
国家自然科学基金项目(42101362,42101321)
河南省科技攻关计划项目(232102111123)。
关键词
卷积神经网络
水稻秸秆覆盖度
深度学习
迁移学习
convolutional neural network
rice residue cover(RRC)
deep learning
transfer learning
