基于无人机高光谱遥感的东北粳稻生长信息反演建模研究

发布时间：2018-03-29 23:17

本文选题：东北粳稻　切入点：高光谱遥感　出处：《沈阳农业大学》2017年博士论文

【摘要】：当前我国东北水稻生产过程中化肥施用量连年增加,在一定程度上造成了部分土地退化、环境污染、病虫害增加等多个生态问题。因此在水稻生长过程中对其生长信息进行快速、无损检测,对于辅助开展水稻营养诊断、精准施肥提高化肥利用率,降低环境污染等具有重要意义。目前,由于空间分辨率和光谱信息的限制基于卫星遥感和地面遥感建立的水稻生长信息反演模型,难以符合区域级水稻生长信息精准反演的需求。随着无人机和高光谱技术的快速发展,为解决区域级水稻生长信息精准反演提供了新的方法和技术支撑。本研究以东北粳稻为主要研究对象,于2015年至2016年连续两年在沈阳农业大学道南试验田开展基于"3414"肥料设计的水稻田间栽培试验。集成无人机高光谱遥感平台,在水稻不同生长期内获取水稻冠层高光谱遥感影像,采用光谱角填图法、最大似然分类法、Fisher判别法、支持向量机法、第二代小波融合算法等提取水稻特征信息并分类,通过分类结果可知,采用基于期望最大算法优化的二代小波分类法提取的纯净水稻高光谱信息分类精度为90.36%,高于其他分类算法提取的水稻高光谱信息。分类结果表明,采用二代小波分类算法对于复杂的稻田环境能够比较理想的提取纯净的水稻高光谱信息。通过分析东北粳稻的高光谱特性可知,在400nm～750nm的可见光区域,水稻高光谱主要是由叶片内色素含量决定光谱信息的变化,水稻对蓝光波段和红光波段的吸收性较好,而对绿光波段吸收弱于蓝光和红光波段,在750nm～1000nm的近红外范围内水稻叶片光谱特征主要受细胞结构变化的影响。在近红外波段水稻对光的反射性较强,而吸收性较弱。通过分析叶片高光谱特性可知,水稻正面与背面的反射率变化趋势是一致的,在可见光区域并无明显差异,在近红外区域内,水稻正面反射率要稍稍大于背面反射率。鲜叶与干叶比较,鲜叶反射率整体上明显小于干叶的反射率,水稻生长含量不同所引起光谱信息的变化是本研究基于高光谱信息反演水稻生长信息的重要理论基础。利用多光谱植被指数和高光谱特征信息采用回归分析的方法建立水稻叶绿素、氮素、叶面积指数(LAI)、生物量等水稻生长信息反演模型,模型反演结果表明,基于多光谱植被指数的叶绿素、鲜生物量、干生物量、LAI、氮素反演模型的决定系数分别为0.498、0.485、0.414、0.599、0.542。基于高光谱特征信息的叶绿素、鲜生物量、干生物量、LAI、氮素反演模型的决定系数分别为0.617、0.569、0.615、0.690、0.668。高光谱特性信息反演的水稻生长信息模型效果要优于通过传统多光谱建立的水稻生长信息反演模型,但通过回归分析方法建立的水稻生长信息反演模型容易受到外界因素的干扰。本研究通过优化作物冠层辐射传输机理模型PROSAIL,提出N-PROSAIL水稻生长信息反演模型,相比现有的水稻生长信息PROSAIL反演模型,N-PROSAIL模型能够弥补现有PROSAIL模型无法反演水稻氮素的不足。在此基础上,分别利用查找表法和数值优化的方法反演水稻生长信息。采用N-PROSAIL模型反演水稻生长信息的模型决定系数为叶绿素0.712、鲜生物量0.565、干生物量0.696、LAI0.696、氮素0.709。模型的反演效果要优于采用回归分析的反演精度。采用遗传神经网络、高斯过程回归、核岭回归、随机森林四种机器学习算法建立水稻生长信息反演模型。其中本研究采用高斯径向基核函数对岭回归算法进行核化,转换为核岭回归算法,在降低模型输入参数的同时,还能够比较理想的处理水稻高光谱这种非线性的问题。本研究利用核岭回归反演水稻生长信息模型的精度要优于其他三种机器学习算法,并且鲜生物量和叶面积指数的模型决定系数分别为0.723和0.786,高于其他方法建立的反演模型。本研究通过无人机高光谱遥感平台获取东北粳稻冠层高光谱信息,建立高光谱信息与水稻生长之间的反演模型及研究结果能够为水稻生长信息快速、无损检测、科学施肥提供一定的理论基础和技术支撑。
[Abstract]:The amount of chemical fertilizer in Northeast of rice in China in the process of production has increased year by year, resulting in a part of land degradation, environmental pollution in a certain extent, pests and diseases increased many ecological problems. So in the process of rice growth in the growth of information fast, nondestructive testing, to carry out auxiliary rice nutritional diagnosis, improve the utilization rate of fertilizer fertilizer plays an important role in reducing environmental pollution. At present, due to the limited spatial resolution and spectral information based on growth model of information retrieval of satellite remote sensing and ground remote sensing of rice, rice growth is difficult to meet the region level accurate information inversion needs. With the rapid development of UAV and hyperspectral technology, provides methods and techniques support new growth information accurate inversion to solve regional level. In this study, the Northeast japonica rice as the main object of study, from 2015 to 2016 for two consecutive years in the Shenyang Agricultural Uinversity Daonan experimental field development based on the "3414" Fertilizer Design of rice field cultivation experiment. Integrated UAV hyperspectral remote sensing platform, obtaining the rice canopy hyperspectral remote sensing image in different rice growth period, using spectral angle mapping method, maximum likelihood classification, Fisher discriminant analysis, support vector machine algorithm, feature extraction of rice information fusion and classification of the second Dai Xiaobo, through the classification results, the classification accuracy of Hyperspectral Information Extraction of two pure rice Dai Xiaobo classification expectation maximization algorithm based on Optimization of rice was 90.36%, higher than other high spectral information classification algorithm. The classification results show that the two Dai Xiaobo classification algorithm for extraction of complex environment to paddy field the ideal of pure rice high spectral information. Through the analysis of spectral characteristics of japonica, 400nm ~ 750nm visible Light area, rice high spectrum is dominated by changes in leaf pigment content determines the spectral information, rice absorption of the blue band and red band better, and to the green light absorption weak blue and red band effect in the near infrared range 750nm ~ 1000nm within the cell structure is mainly affected by the changes of the rice Ye Pianguang spectrum characteristics in. Near infrared light reflection on rice is strong, and weak absorption. By analyzing the spectral characteristics of the blade, the front and back of the rice albedo variation trend is consistent, no apparent difference in the visible light region, in the near infrared region, rice positive reflectivity to be slightly larger than the back reflectivity of fresh leaves and dry leaves compared. On the whole, fresh leaf reflectance was smaller than the dry leaf reflectance of rice growth in different caused by the changes of the spectrum information is the research of hyperspectral data inversion based on Rice An important theoretical basis for long information. Using multi spectral vegetation index and hyperspectral characteristics of information to establish the chlorophyll, by using the method of regression analysis of nitrogen, leaf area index (LAI), the biomass of rice growth information retrieval model, model inversion results show that the multi spectral vegetation index based on chlorophyll, biomass, stem biomass the amount of LAI, the coefficient of determination of nitrogen, the inversion model respectively based on 0.498,0.485,0.414,0.599,0.542. Hyperspectral Feature Information of chlorophyll, biomass, stem biomass, LAI, nitrogen determination coefficient inversion model respectively. The growth model of information 0.617,0.569,0.615,0.690,0.668. inversion of hyperspectral characteristic information retrieval rice growth information model is better than established by traditional multispectral rice but, through the regression analysis method to establish the rice growth information retrieval model is vulnerable to interference from external factors. This research The PROSAIL through the optimization of crop canopy radiative transfer mechanism model, N-PROSAIL information retrieval model is proposed for rice growth, compared with the existing rice growth information PROSAIL inversion model, to supplement the existing PROSAIL model can not inversion of nitrogen in rice N-PROSAIL model. On this basis, respectively, using look-up table method and the method of numerical optimization inversion of rice growth growth information. Using N-PROSAIL model inversion rice model decision coefficient 0.712 for chlorophyll, 0.565 fresh biomass, stem biomass of 0.696, LAI0.696, 0.709. model of nitrogen inversion effect is superior to the inversion accuracy of regression analysis. Using genetic neural network, Gauss regression, kernel ridge regression, random forest four machine learning algorithm of rice growth the information retrieval model. Which were used in the study of kernel ridge regression algorithm Gauss radial basis kernel function conversion The kernel ridge regression algorithm, reduce the input parameters of the model at the same time, the nonlinear processing of rice in high spectrum of this can also be an ideal problem. This study uses kernel ridge regression inversion rice growth information model is more accurate than other three kinds of machine learning algorithm, and the model decision coefficient and leaf area index of fresh biomass was 0.723 and 0.786, higher than the other inversion model established method. Through the study of UAV hyperspectral remote sensing platform for the northeast rice canopy spectral information, the establishment of the high spectral information between rice growth model and inversion results can for rice growth information fast, nondestructive testing, and provide a certain theoretical foundation and technical support of scientific fertilization.

【学位授予单位】：沈阳农业大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：S511.22;S127

【参考文献】