东北地区稻田土壤氮含量无人机高光谱反演建模研究  

Study on Modeling of UAV Hyperspectral Inversion of Soil-borne Nitrogen Content in Rice Field in Northeast China

作　　者：许童羽[1,2] 方健羽 郭忠辉 白驹驰 金忠煜 于丰华[1,2] XU Tong-yu;FANG Jian-yu;GUO Zhong-hui;BAI Ju-chi;JIN Zhong-yu;YU Feng-hua(School of Information and Electrical Engineering,Shenyang Agricultural University,Shenyang 110161,China;Liaoning Provincial Key Laboratory of Intelligent Agricultural Technology,Shenyang Agricultural University,Shenyang 110161,China)

机构地区：[1]沈阳农业大学信息与电气工程学院,沈阳110161 [2]沈阳农业大学辽宁省智慧农业技术重点实验室,沈阳110161

出　　处：《沈阳农业大学学报》2023年第6期759-768,共10页Journal of Shenyang Agricultural University

基　　金：辽宁省教育厅重点攻关项目(LSNZD202005)。

摘　　要：土壤氮元素是土壤肥力的一个重要指标,掌握土壤氮元素含量的变化是观测农作物的发育状况及时空变化规律等的基础。利用高光谱对土壤氮元素进行反演,可为精准农业地表土壤元素快速测定提供参考。为实现对土壤中氮元素含量的快速测定,以沈阳农业大学海城试验田为例,对土壤原始反射率进行了一阶导数变换,运用等距特征映射算法(isometric feature mapping,lsomap)、竞争性自适应重加权采样法(competitive adapative reweighted sampling, CARS)对一阶导数光谱数据进行降维并提取出相关特征。运用BP神经网络(back propagation neural network,BPNN)、GA优化(GA-BPNN)以及经过NSGA-Ⅲ优化后的BP神经网络(NSGA-Ⅲ-BPNN)3种分析方法建立土壤全氮的高光谱反演模型,并利用决定系数(R^(2))和均方根误差(root mean square error,RMSE)对反演模型进行评价。结果表明:经过ISOMAP进行的降维相对于CARS能有效的对特征进行提取。优化后的神经网络模型建立的土壤养分含量预测模型优于未优化的神经网络,能极好地预测土壤中的氮元素含量。基于Isomap降维后的NSGA-Ⅲ-BPNN的模型的反演模型预测效果最好,最终预测土壤全氮含量训练集为R^(2)=0.842、RMSE=0.077,测试集R^(2)=0.826、RMSE=0.089。反演精度高于GA-BPNN和BPNN的模型反演精度,与其他模型组合相比,该组合可以为土壤氮元素含量的反演研究提出一种新的方法。Soil nitrogen is an important indicator of soil fertility.Obtaining the change of soil nitrogen content is the basis for observing the development and the laws of temporal and spatial changes of crops.The retrieval of soil nitrogen by hyperspectral method can provide reference for rapid determination of surface soil elements in precision agriculture.In order to realize the rapid determination of nitrogen content in soil,the Haicheng experimental field of Shenyang Agricultural University was taken as an example to carry out a first-order differential transformation of the original soil reflectance,and the isometric feature mapping algorithm(lsomap)and competitive adaptive reweighted sampling(CARS)were used to reduce the dimensions of first-order differential spectral data and extract relevant features.BP neural network(BPNN),GA optimization(GA-BPNN)and NSGA-Ⅲoptimized BP neural network(NSGA-Ⅲ-BPNN)were used to establish the hyperspectral inversion model of soil total nitrogen,and the determination coefficient(R^(2))and root mean square error(RMSE)were used to evaluate the inversion model.The results show that the dimensionality reduction through ISOMAP can effectively extract features compared with CARS.The prediction model of soil nutrient content established by the optimized neural network model is better than that of the non-optimized neural network model,which can predict the nitrogen content in soil very well.The inversion model of NSGA-Ⅲ-BPNN model based on Isomap dimensionality reduction has the best prediction effect.The final predicted soil total nitrogen content training set is R^(2)=0.842,RMSE=0.077,and the test set is R^(2)=0.826,RMSE=0.089,and the inversion accuracy is higher than that of GA-BPNN and BPNN models.Compared with other model combinations,this combination can provide a new method for the inversion of soil nitrogen content.

关 键 词：无人机高光谱 土壤氮元素反演 NSGA-Ⅲ优化算法 Isomap降维 BP神经网络 

分 类 号：S252[农业科学—农业机械化工程]