植被覆盖度的照相法测算及其与植被指数关系研究

【作者】 顾祝军；

【导师】 曾志远；

【作者基本信息】 南京师范大学 ， 地图学与地理信息系统， 2005， 硕士

【摘要】 植被覆盖度(Vegetation Coverage,简称VC)是衡量地表植被状况的一个最重要的指标,同时,它又是影响土壤侵蚀的主要因子。植被覆盖及其变化是全球和区域生态环境变化的重要指示。而植被覆盖度测量方法的改进以及测量精度的提高是各领域发展的需要。无论地表实测方法还是遥感相关模型,都因为技术的进步和应用的需求而对相关领域的研究提出了新的要求。 本文在国家自然基金项目“SPOT图像信息提取与土地利用/土地覆盖识别与监测研究”(编号:40371053)的资助下,以南京市主城区及其周边农田为研究区,从该区SPOT和ETM+两幅卫星图像提取的各种植被指数中,选取使用最为广泛的归一化植被指数(NDVI)作为遥感植被指标;野外调查与遥感成像时间准同步,累计20余天、行程数百公里,以亚米级精度的差分GPS获取几何控制点和所有实测样方的经纬度数据;用植被覆盖度地表实测方法中最为精确的数码照相法,结合传统的目估法,获取样方植被覆盖度;最后建立了不同卫星图像、不同植被类型、不同植被密度以及面向全区(图)的植被覆盖度遥感反演模型,并对模型进行了应用和精度检验,其精度都在80%以上。 本研究的主要成果概括如下: 1 数据方面 通过艰苦细致的野外实测,获取了针对两种卫星图像的数百个样方实测数据,包括每一样方的植被覆盖度、植被覆盖度相片、中心点经纬度和相对位置、植被类型、土地利用类型等,以及用于几何精校正的几何控制点数据; 2 实测方法方面 自行设计并成功构建了以数码照相机和DGPS为核心的野外植被覆盖度信息采集系统(VCCS);拍摄采样方法则根据不同卫星图像和植被状况,分别成功采用SPOT单层植被“5点法”、SPOT多层植被“5对法”和ETM+样方植被“5区法”;对多层植被,建立了由“上、”“下”两种植被覆盖度推算样方植被覆盖度的测算模型;每张植被覆盖度相片均采用垂直照相法拍摄; 3 数字图像处理方面 根据数码相片中可见光波段的光谱特性,对垂直向上和垂直向下照相法获取的相片设计了不同的决策树分类模型,实现相片植被覆盖度信息快速、准确的提取,克服了传统分类方法的不足;根据样方中心点在像元中的位置差异,将样方分为A型、B型和C型,进行不同的“邻域”处理,有效提高了样方—像元空间对应的准确性; 4 模型方面 成功构建了SPOT图像基于植被类型的VC反演模型(CM模型)、SPOT图像基于植被密度的VC反演模型(DM模型)、SPOT图像基于全研究区的VC反演模型(AM模型)、ETM+图像基于全研究区的VC反演模型(AM模型)各若干种,其中SPOT图像CM模型,将植被按地域和组成结构分为矮草、灌草、森林和农作物四种类型,建立了基于植被类型的非线性和线性相关模型各4个,下述几个非线性模型的相关性优于线性模型(y=VC,x=NDVI,下同): 矮草 y=4.8253x~2+2.0624x+0.3579 (R~2=0.6636) 灌草 y=4.562x~2+0.6899x+0.5546 (R~2=0.6356) 森林 y=-2.0552x~2+0.4631x+0.8612 (R~2=0.7145) 农作物 y=1.4706x~2+1.3582x+0.4734 (R~2=0.7391) 对以下几种模型进行了应用和精度检验: 1) SPOT图像DM模型 将SPOT样方数据按照植被覆盖度(密度)等级不同,建立线性和非线性回归模型,对以下三种非线性模型进行了应用检验,其总体精度为86.7045%: ①稀疏植被 y=-0.649x~2+0.5616x+0.2762 (R~2=0.7120)更多还原

【Abstract】 Vegetation coverage (VC) is the most important index to measure the vegetation status of the earth’s surface and the major factor to affect the soil erosion. The change of vegetation coverage indicates the changes of global and regional entironment. Moreover, it is necessary to improve the measurement and its precision of vegetation coverage for the development of all relative fields. Since the improvement of techniques and the need of applications, both the measurement of the earth’s surface and the retrieval models of remote sensing all put forward new demands for the researches on relative fields.Based on the project of national natural finance "research on information extraction, monitoring and identification of land use/land cover for SPOT images"(number: 40371053), this paper choose NanJing City and the farmland around it as the research area.We also choose NDVI, the most widely used index as the index of vegetation information of remote sensing from all kinds of vegetation indexes extracted from the SPOT and the ETM+ images of this area. With the field survey and the remote sensing imaging synchronously, we worked for more than 20 days and went hundreds of kilometers to get the longitude/latitude data of GCPs and all the samples by means of GPS with sub-meter precision.Meanwhile, we combined the digital photography method with traditional ocular method to get VC of the samples.Eventually, we set up relative models of this area based on different satellite images, different kinds of vegetation and different density of vegetation with the precisions above 80%.The achievements of the paper include:1 DataDuring the hard field survey, we got the data of hundreds of samples for the two satellite images including VC, VC photos, longitude and latitude of each sample center, types of vegetation and types of land use, as well as GCP data which used as geometrical correction references.2 Field measurement methodsWe successfully designed and founded the vegetation coverage collection system (VCCS) based on the digital camera and DGPS; According to different satellite images and vegetation, we adopted"5 points method"for SPOT monolayer vegetation, "5 pairs method"for SPOT multilayer vegetation and"5 sections method"for ETM+ data respectively. For multilayer vegetation, we set up calculating models to calculate VC of the samples from"upward and downward"VC; All photos of VC were taken vertically.3 Processing of the digital imagesWe designed different models of decision tree for photos taken from upwards and downwards according to the spectrum characters of visible light. By this way, we could extract VC information from photo data quickly and exactly and overcome the shortage of traditional classify method. Based on the differences of the location of the center point in each pixel on remote sensing images, we classified the samples into A, B and C types and processed in "adjacent field". The precision of correspondence to pixel space of samples was improved notably.4 Relative modelsWe founded different VC-VI relative models successfully including CM model based on vegetation types for SPOT image, DM model based on vegetation density for SPOT image and AM model based on the whole study area for SPOT image and ETM+ image. In the CM model for SPOT image, we divided the vegetation into four kinds including grass, shrub, forest and crop according to vegetation region and its structure. We also build four relative models of linearity and nonlinearity respectively, while the relativity of nonlinear models is better than the linear models (y=VC, x=NDVI):Grass y=4.8253x2+2.0624x+0.3579 (R2=0.6636)Shrub y=-4.562x2-K).899x+0.5546 (R2=0.6356)Forest y=-2.0552x2+0.4631x-K).8612 (R2=0.7145)Crop y=1.4706x2+1.3582x+0.4734 (R2=0.7391)We tested the models below from application and precision:1) The DM model for SPOT imageWe set up linear and nonlinear regressive models according to different grade of vegetation density of SPOT samples and tested three kinds of nonlinear models as follows. The whole precision is 86.7045%.?the sparse vegetation y=-0.649x2+0.5616x+0.2762 (R2=0.7120)?the normal density vegetation y=-0.5031 x2+0.3788x+0.5578 (R2=0.7437)?the dense vegetation y=-0.1644x2+0.3397x+0.8661 (R2=0.6851)2) The AM model for SPOT imageIt is a VC abstraction model for the whole image without classification for SPOT image. We take NDVI from the SPOT image as the independent variable and AM model as the band operation equation to abstract the VC data.We founded one simple equation, one quadratic equation, one cubic equation and one biquadrate equation respectively and tested the cubic equation which performed better with the precision of 86.7447%. The model is:y=-6.4515x3-0.5789x2+2.2346x+0.5697 (R2=0.7128 )3) The AM model for ETM+ imageIt is a VC abstraction model for the whole image without classification for ETM+ image. We take NDVI from the ETM+ image as the independent variable and AM model as the band operation equation to abstract the VC data. We founded one simple equation, one quadratic equation, one cubic equation and one biquadrate equation respectively and tested the cubic equation which performed better with the precision of 88.5053%. The model is:y=-6.4515x3-0.5789x2+2.2346x+0.5697 (R2=0.8804)更多还原