【作者】 戴震；

【导师】 张东彦； 徐新刚；

【作者基本信息】 安徽大学 ， 电子与通信工程（专业学位）， 2021， 硕士

【摘要】 玉米是我国主产粮食作物,对维持国家粮食安全具有重要意义。氮素是植物生命活动过程中必须元素,对植被的光合作用以及生长发育具有指示性及决定性作用。如何及时、准确地监测玉米各生育期叶片氮含量,制定相应的氮肥施用方案,是提高玉米产量和品质的关键。传统的玉米氮素监测方法,是通过人工进行实地取样,然后将样本带回实验室进行测量。该方法消耗大量人力物力,且效率低下,还会对田间作物造成损坏。在进行人工采样时,难免出现人为误差,并且该方法很难形成统一标准。随着遥感技术的发展,通过遥感展开农情监测优势明显,借助无人机遥感平台,可以实现快速,高效的数据采集,方便后期进行数据处理和各种作物不同指标的反演研究。本文以北京市昌平区小汤山国家精准农业研究示范基地为研究区。通过无人机遥感平台搭载数码相机和多光谱传感器,采集2017年、2019年玉米抽雄吐丝期和灌浆期无人机遥感影像,开展基于无人机多源影像信息融合的玉米叶片氮含量(Leaf Nitrogen Concentration,LNC)反演研究。论文的主要研究内容和结论如下:(1)基于无人机多源影像像素级融合的玉米氮素遥感监测。通过GS(Gram-schmidt)融合方法将无人机高清数码影像和多光谱影像进行像素级融合。对无人机高清数码影像、无人机多光谱影像和融合后影像进行典型冠层光谱指数的构建,并将冠层光谱指数与玉米实测LNC进行相关性分析,筛选相关性最好的前五个光谱指数。利用随机森林(Random Forest,RF)分类方法,剔除土壤和阴影等背景噪声,探讨土壤背景因素对玉米LNC模型精度的影响。最后采用拉索回归(Least Absolute Shrinkage and Selection Operator,LASSO)和偏最小二乘回归(Partial Least Squares Regression,PLSR)两种算法,综合评估不同条件下玉米LNC反演模型精度。结果表明,数码影像和多光谱影像融合后,玉米LNC反演模型精度得到提升,其中多光谱影像融合后PLSR模型在抽雄吐丝期去除土壤背景条件下R~2提升0.25,RMSE降低0.03,NRMSE降低2.20%;玉米LNC反演模型在灌浆期反演效果最好,多光谱影像灌浆期去除土壤背景后LASSO模型和PLSR模型的R~2比抽雄吐丝期分别提升了0.15和0.10;无人机影像在剔除土壤背景因素后,数码影像和融合后影像的各模型精度得到提升,并且LASSO模型略好于PLSR模型。(2)基于无人机多源影像信息的特征级融合和深度学习方法的玉米氮素遥感监测。对构建的无人机高清数码影像和无人机多光谱影像的冠层光谱指数,通过灰色关联度分析法,筛选与玉米LNC关联度最高的五种光谱指数,作为玉米冠层光谱特征信息;通过灰度共生矩阵(Gray-level Co-occurrence Matrix,GLCM)提取数码影像的纹理特征信息;将玉米冠层光谱特征信息和纹理特征信息通过不同线性组合作为变量输入,利用随机森林回归(Random Forest Regression,RFR)、支持向量机回归(Support Vector Regression,SVR)和深度神经网络(Deep Neural Networks,DNN)算法建立玉米LNC反演模型。综合评估不同条件下玉米LNC反演模型精度。结果表明,纹理信息对模型反演精度提升效果明显,多光谱冠层光谱特征加上纹理信息后,RFR、SVR、DNN模型的R~2分别提升0.08、0.13、0.09;DNN回归模型对玉米LNC预测结果优于RFR和SVR模型,其中DNN-F2模型的R~2达到0.85,RMSE为0.27,NRMSE为10.07%。本文从无人机多源影像信息的像素级和特征级融合出发,分析筛选玉米冠层光谱特征和纹理特征,建立玉米LNC遥感反演模型,实现无人机多源影像信息融合下玉米LNC高效、无损、便捷的遥感监测,为玉米LNC获取提供新的手段,也为田间农情信息精准获取提供科学有效的遥感技术支持。更多还原

【Abstract】 Corn is a high-yield food crop in China,which is meaningful to maintaining food security.Especially,nitrogen is an essential element in the process of plant life activities,and it has an indicative and decisive effect on the photosynthesis,growth and development of vegetation.Therefore,how to timely and accurately monitor the nitrogen content of corn in each growth period and formulate the corresponding nitrogen fertilizer application plan is the key to improve the quality of corn.The traditional method of corn nitrogen monitoring is to manually take field samples,and then take the samples back to the laboratory for measurement.This method which consumes a lot of manpower and material resources is inefficient,and can cause damage to field crops.When surveyors conduct manual sampling,in manual sampling,human error is inevitable,and it is difficult to form a uniform standard with this method.Therefore,the use of remote sensing technology in agriculture has advantages.Through with the development of remote sensing technology,the advantages of using remote sensing to monitor agricultural conditions are obvious.With the help of UAV remote sensing platform,fast and efficient data collection can be realized,which is convenient for later data processing and inversion research of various crop indicators.This thesis uses the Xiaotangshan National Precision Agriculture Research Demonstration Base in Changping District,Beijing as the research area to conduct experiments.The UAV remote sensing platform is equipped with digital cameras and multi-spectral sensors to collect UAV remote sensing images of corn tasseling and filling periods in 2017 and 2019.Launched the inversion study of Leaf Nitrogen Concentration(LNC)based on the fusion of Unmanned Aerial Vehicle multi-source image information.The main research content and conclusions of the thesis are as follows:(1)Research on remote sensing monitoring of corn nitrogen based on UAV multi-source image fusion.Through the GS(Gram-schmidt)fusion method,the UAV high definition digital image and multi-spectral image are fused at the pixel level.Construct a typical canopy spectral index for UAV HD digital images,UAV multi-spectral images and fusion images,and analyze the correlation between the canopy spectral index and the measured LNC of corn,and select the top five spectral indices with the best correlation.Random Forest(RF)classification method is used to eliminate background noises such as soil and shadows,and the influence of soil background factors on the accuracy of the corn LNC model is discussed.Finally,two algorithms,Least Absolute Shrinkage and Selection Operator(LASSO)and Partial Least Squares Regression(PLSR)are used to comprehensively evaluate the accuracy of the corn LNC inversion model under different conditions.The results show that the accuracy of the corn LNC inversion model is improved after the digital image and the multi-spectral image are fused.After the multi-spectral image is fused,the PLSR model will increase the R~2 by 0.25,reduce the RMSE by 0.03,and reduce the NRMSE by 2.20% when the soil background is removed during the tasseling and filling period.The corn LNC inversion model has the best inversion effect in the filling period,the R~2 of the LASSO model and the PLSR model were increased by0.15 and 0.10 respectively compared with the tasseling and filling period;and the LASSO model was better than the PLSR model.(2)Research on remote sensing monitoring of corn based on the feature-level fusion of UAV multi-source image and deep learning methods.The canopy spectral indices of the constructed UAV HD digital images and UAV multi-spectral images were screened by gray correlation analysis for the five spectral indices with the highest correlation with corn LNC as corn canopy spectral feature information;through the gray co-occurrence matrix Extract the texture feature information of digital images;pass spectral feature information and texture feature information through Random Forest Regression(RFR),Support Vector Regression(SVR)and Deep Neural Networks(DNN)algorithm establishes the corn LNC inversion model.Comprehensive evaluation of maize LNC inversion model accuracy under different conditions.The results show that the texture information of digital images contributes to the accuracy of model inversion.After the multi-spectral canopy spectral features plus texture information,the R~2 of the RFR,SVR,and DNN models are increased by 0.08,0.13,and 0.09 respectively,the DNN regression model predicts the results of the corn LNC It is better than the RFR and SVR models.Among them,the R~2 of the DNN-F2 model reaches 0.85,the RMSE is 0.27,and the NRMSE is 10.07%.In summary,this paper uses the pixel-level and feature-level fusion of UAV multi-source image feature information by analyzes and selects the spectral and texture features of the corn canopy,and constructs The corn LNC remote sensing quantitative inversion model to achieve the efficient,non-destructive and convenient nitrogen remote sensing monitoring under the multi-source image information fusion of the corn LNC which improves the method of obtaining corn LNC information and provides accurate acquisition of field agricultural information scientific and effective remote sensing technology support.更多还原