节点文献
基于高光谱成像技术的作物叶绿素信息诊断机理及方法研究
Diagnosis Mechanism and Methods of Crop Chlorophyll Information Based on Hyperspectral Imaging Technology
【作者】 张东彦;
【导师】 王纪华;
【作者基本信息】 浙江大学 , 农业遥感与信息技术, 2012, 博士
【摘要】 如何利用光谱技术快速、无损、准确地诊断作物的养分状况,一直是农业定量遥感关注的焦点。受遥感仪器性能及价格的限制,目前的研究大多利用非成像高光谱进行作物长势监测与机理探测研究。近年来,随着材料科学、传感技术及精密仪器的飞速发展和对田块尺度作物精细管理的需求,低成本、高性能的地面成像光谱仪的开发与应用已成为农业定量遥感的一个重要方向。论文以地面成像光谱仪为观测工具,采集了不同作物(小麦、玉米、大豆)在不同生育期、不同尺度(叶片与冠层)、多角度下的高光谱影像数据,系统地分析从影像中提取的高光谱数据及其与作物叶绿素信息的关系,尝试在下述三个方面进行研究:1)验证自主研发的地面成像光谱仪(PIS)获取的光谱数据的可靠性;2)挖掘成像高光谱“图谱合一”数据在作物叶绿素信息遥感诊断机理及反演中的优势;3)探讨多角度观测及影像分割在作物BRDF特性定量解析及叶绿素密度反演中的作用。主要工作及进展如下:1)提出一种基于不同参照布的反射率场地标定方法。它是通过建立可见-近红外成像光谱仪(VNIS)与地物光谱仪(ASD)同时获取的DN值之间的转换关系,将ASD获取的标准白板的DN值转换为VNIS的DN值,再通过定标公式实现影像反射率提取。研究发现,该方法计算得出的相对反射率与当场的A SD反射率曲线有较高的一致性,由此证明这种方法能满足成像传感器反射率场地标定的要求。2)在作物的单叶尺度,开展PIS光谱数据的可靠性验证、作物叶绿素信息反演方法探索,得出:a)通过分析成像与非成像高光谱数据提取的红边特征曲线及红边位置,发现两仪器获取数据有很好的一致性,由此验证自主研发PIS光谱数据是可靠的。b)利用窄波段成像光谱曲线的峰谷特征构建新型特征参数,即“三边”的光谱变化速率、“三边”夹角及衍生变量在叶绿素含量反演中进行应用。结果表明:相比传统的“三边”光谱参数,峰谷特征参数能有效提高叶绿素含量的反演能力。c)利用PIS对活体玉米植株进行成像立方体采集,在影像中选取玉米倒1叶至倒4叶的叶片光谱,构建了叶片叶绿素含量反演模型;用小区玉米叶片样本进行验证,得出预测模型的决定系数R2=0.897;验证模型的决定系数R2=0.887,均方根误差RMSE=1.8;在不同植被覆盖度下(盆栽-小区)的玉米植株叶片都有很好效果,说明采用这种方法构建的模型具有较高的精度。d)利用PIS获取“图谱合一”的小麦叶片影像数据,采用偏最小二乘法研究同一叶片不同位点组合所建叶绿素含量预测模型的精度;研究不同层位叶片叶绿素含量估测模型的精度。研究发现,同一叶片2、4、6位点组合的模型精度高于1、3、5位点组合的精度;不同层位叶片模型精度为中层>上层>下层;所有叶片建立的综合模型精度最高。3)把VNIS在田间小尺度范围内进行应用,基于其“图谱合一”的优势,a)在高光影像中寻找植被、裸土、光照叶片和阴影叶片的光谱差异,构建归一化光谱分类指数,实现了不同类型地物的分割与光谱提纯。b)在不同地物光谱提纯的基础上发现,当植被与土壤混合存在时,对叶绿素密度敏感的波段基本上都在红与近红外波段区间;当植被光谱提纯后(剔除土壤光谱),对叶绿素密度敏感的波段范围增大,尤其是蓝、绿波段。由此说明,背景土壤对光学遥感反演植被叶绿素密度有较大影响。此外,阴影叶片会影响植被叶绿素密度敏感波段的选择,当构建新型植被指数时,要根据植被冠层叶片结构,尝试把阴影比例作为一个影响因子,在公式中加以体现,以便提高叶绿素密度定量反演的精度。4)对不同密度大豆冠层的多角度数据进行分析,研究同一视场内植被与土壤混合光谱信息提纯前后的冠层BRDF变化特征,发现:a)在主平面观测时,土壤光谱去除后,纯植被冠层反射率在前向观测时,随着天顶角的减小而增大,这和视场内同时存在植被和土壤时的研究结果不同;后向观测时,随着天顶角的增加而增大;后向反射率高于前向,这和混合植被的BRDF特征一致。b)在垂直主平面方向上,土壤光谱去除前后的不同密度大豆冠层反射率在垂直主平面都有一致的对称性,去除后的前后向反射率对称性更强。5)利用多角度成像数据对大豆冠层叶绿素密度的反演进行解析与评价,发现:a)(0°,20°,40°,60°)的天顶角组合有最高的R2=0.834(预测模型)和最小的RMSE=6.13;(20°,40°,60°)天顶角组合的决定系数值高于(0°,20°,40°)的组合。在混合植被、纯植被、光照植被三类数据中有一致的结果。b)在不同天顶角下,40°天顶角是反演叶绿素密度的最优角度。c)在不同方位角下,0°方位角(太阳主平面的后向观测)是反演叶绿素密度的最优角度。d)天顶角变化是影响大豆冠层叶绿素密度反演精度的主要因素,这归根结底是由观测视场中的背景土壤及阴影叶片面积比例发生变化而导致的。
【Abstract】 How to fast, nondestructively and precisely detect crop nutrition status using spectral technology is always a key problem for agricultural quantitative remote sensing. Currently, owing to restriction of ability and price for remote sensors, a larger number of researches were most focused on monitoring crop growth and probing remote sensing mechanism using non-imaging hyperspectra from ASD FieldSpec(?)3 field spectrometers. In recent years, various kinds of hyperspectral imaging spectrometers with low price and high ability were developed and applied in the field by domestic and overseas researchers, which has been an important development trend for quantitative remote sensing.In this work, field hyperspectral imaging spectrometer was used to collect imagery data of crops at both single leaf and multi-angle canopy scales. Those crops specifically included wheat, corn and soybean at different growth stages, and they were then systematically analyzed in the aspects of spectral reflectance data from imageries and correlations between the data and crop chlorophyll information. Three aspects were mainly researched:1) the reliability of spectral data from Pushbroom Imaging Spectrometer (PIS) was validated; 2) remote sensing detection mechanism and retrieval of crop chlorophyll information were explored using the advantage of hyperspectral data combing image with spectrum; 3) through analyzing the influences of multi-angle observation and image classification, the aim was to explain bidirectional reflectance distribution function (BRDF) varied features of crop canopy and estimate its chlorophyll density. Some important conclusions can be drawn as follows:A new method of field calibration was proposed. The processing was to build a transform relationship between two types of digital number (DN) values simultaneously acquired by visible and near-infrared imaging hyperspectral spectrometer (VNIS) and ASD, and then collected DN values of standard white panel by ASD was transformed to results of VNIS, and consequently the reflectance of whole image was calculated by corresponding calibration formulae. The result presented that reflectance curve calculated by the new method was similar with that from ASD, which proved that the proposed method in this study could satisfy the requirement of field calibration for hyperspectral imaging sensors.Through analyzing spectral characteristic curves of red edge and variation of red position extracted from imaging and non-imaging hyperspectral data, the result showed that they had high consistency between two types of data from PIS and ASD, so it could be proved that the performance of self-developed PIS was reliable. On that basis, new characteristic parameters were further extracted from peak-valley features of narrow-band imaging spectral curves, such as spectral change ratio, spectral angles and derivative variables of three edges, and then they were used to assess chlorophyll content of maize leaf. The results showed that new peak-valley characteristic parameters could more effectively improve the accuracy of prediction model of chlorophyll content in comparison with traditional feature parameters from three edges.VNIS was used to collect data combining image with spectrum in the field. After analyzing the spectral differences among vegetation, bare soil, illuminated leaves and shadowed leaves from crop images, normalized spectral classification index was built and used to classify different targets in the image. The results indicated that background soil affected the inversion accuracy of chlorophyll density using spectral remote sensing. In addition, shadowed leaves also influenced the assessment accuracy of chlorophyll density. When the new vegetation index was presented, the shadow percentage as influence factor should be considered and be reflected in the calculation formulae, so the aim of this study was to improve the retrieval accuracy of chlorophyll density.VNIS was also used to collect multi-angle data of soybean canopy in different densities. It could be found that BRDF features of soybean canopy existed in vegetation and soil and pure vegetation (masked bare soil) were specifically researched. Some results can be concluded that 1) For forward observation of principal plane, canopy reflectance of pure soybean vegetation gradually increased when zenith angle changed from 60 degree to 0 degree, which was different from the result of existed vegetation and soil; while for backward observation, canopy reflectance of pure soybean vegetation gradually increased when zenith angle changed from 0 degree to 60 degree, which was same as the result of existed vegetation and soil.2) For forward observation angle of perpendicular principle plane, there were consistent symmetry of BRDF features of soybean canopy in different densities between mixed targets (vegetation and soil) and pure vegetation (masked bare soil), and the symmetry of latter BRDF features was higher than the former.The paper analyzed the correlation between multi-angle spectral data of soybean canopy and chlorophyll density, some results could be concluded that the changes in zenith angles were the most key factor to affect assessment accuracy of chlorophyll density of soybean canopy, and the reason was that the percentage of background soil and shadowed leaves gradually changed in the field of view.
【Key words】 Hyperspectral Imaging; Field Calibration; Peak-valley Characteristic Parameters; Spectral Classification Index; Multi-angle; BRDF; Crop;