Deriving of light information and scale effect analysis of the desert riparian forest in the arid region-Take the middle and lower reaches area of the Tarim

干旱区荒漠河岸林植被是干旱内陆河流域生态环境的核心，在抑制荒漠化过程和保护生物多样性等方面有着重要的生态意义。因此，实时监测荒漠植被生长状态信息尤为重要。与传统点尺度上耗时耗力的人工量测相比，遥感为获得不同尺度植被参量信息提供了一个便捷的多元化工具。位于干旱地区荒漠河岸林群落，其植被生长稀疏、类群结构简单，使得遥感影像上获取的荒漠河岸林光谱信息极其微弱，甚至于难以检测。另一方面，由于地表空间格局与过程在不同尺度上表现出明显的特征差异，作为描述地表特征的遥感影像数据同样具有尺度特征，表现为同一种格局和过程随遥感影像的覆盖范围和观测尺度而变化，而这种变化过程对于原本就难以提取的荒漠植被弱信息增加了难度。本文针对上述问题，以塔里木河干流中下游地区荒漠河岸林稀疏植被群落为切入点，做了以下研究：构建了干旱区荒漠河岸林信息提取几何结构概念模型；基于几何结构概念模型的理念，以高空间分辨率影像为信息源，采用分类决策树模型、几何光学物理模型以及光谱角匹配技术，探讨了荒漠河岸林的类别识别，进一步以正演手段实现了 MODIS亚像元结构模式的获取；建立了荒漠河岸林稀疏植被覆盖度遥感模型，对覆盖度信息尺度转换效应进行了分析。从荒漠河岸林植被外貌特征、生理生化参数变化特点出发，认为干旱区荒漠河岸林植被弱信息提取需提高类别识别的精度，实现生理、生化参量信息反演的高时效性。提出以高空间分辨率影像解决混合像元分解问题，进而采用线性模型正演的方式，获取MODIS亚像元结构模式；以高时间分辨率的MODIS影像提取生物物理、生化参数的几何结构概念模型。采用几何光学模型与光谱角匹配结合，解决混合像元信息分解，进行干旱区荒漠河岸林类别识别。首先从遥感视角的角度，将地物分解为目标和背景，提出塔里木河干流荒漠河岸林植被分类系统；其次用像元信息分解和多变量决策树法将非荒漠植被信息剔除，采用几何光学模型模拟各类荒漠植被的像元光谱，采用光谱角匹配的方法将荒漠植被进一步进行分解，得到塔里木河干流中下游地区典型研究区的植被分类专题图，分类精度结果表明：基于混合像元分解与几何光学模型的分类方法总精度达到了77.66%，Kappa系数为0.704，表明分类质量很好。 MODIS亚像元结构模式分析及影像模拟研究表明：随着像元尺度的扩展，图像对地物的分辨能力减弱，地物边界变得模糊。统计特征值最大值、最小值、均值及标准变差变化幅度不大，模拟影像包含的信息量没有太大的损失，这种基于线性模型基础，通过正演的方式所得到MODIS 250m、500m、1000m多尺度亚像元地类组成结构模式是有意义的。进一步分析模拟影像与真实影像光谱尺度效应，认为光谱特征的尺度扩展是有限度的，尤其是从500m扩展到1000m尺度，模拟影像的反射率变化范围与MODIS原始影像差异显著，说明信息量不一，这样的尺度扩展对于分析光谱反射率特征没有实际意义。干旱区稀疏荒漠河岸林植被覆盖度信息提取从遥感模型及模型的尺度效应两个方面进行了分析：首先以地面实测值与植被指数NDVI、RVI、DVI、RDVI建立经验模型，以NDVI指数盖度模型、亚像元结构模型、三波段最大梯度差法提取研究区植被盖度信息，通过模型检验分析，研究认为本文所提出的调整三波段最大梯度差法模型误差小，算法简单，是最适合于干旱区稀疏植被覆盖度提取的遥感模型；其次，研究将调整的三波段最大梯度差法模型带入真实影像反演得到覆盖度信息，通过简单平均尺度扩展方法，以TM覆盖度影像模拟不同尺度（MODIS 250m、500m、1000m）的覆盖度模拟影像作为验证信息源来检验模型在不同尺度上的反演效应及普适性，调整后的三波段最大梯度差法在不同尺度上反演覆盖度信息均得到了较好的结果。因此，在小尺度上获得验证的覆盖度信息通过该扩展方法可以作为验证其它尺度模型反演结果的有效信息源Ecosystems of desert riparian forest at the arid areas is the core of ecological environment in arid continental river basins, and play an important role in restricting land desertification and conserving biodiversity, etc. so it is very important to monitor the growing status of the desert vegetation. Comparing with the ground manual measurement at the site scale，the satellite remote sensing provides the technically consistent and temporally regular means of deriving the vegetation parameters at multi-scale. The spectrum of desert riparian forest from the remote sensed image is very light, so much as to detect difficultly. On the other hand, spatial pattern and procedure of the surface represent obvious feature variance on different scale, and the remote sensed data describing the surface feature have the same scale character, the same pattern and process will change with the covered space and observed scale，it enhance the difficulty to deriving the light information of the desert riparian forest. Taking the desert riparian forest belts along both riversides of the middle and lower reaches at the Tarim River Basin as the research object, this paper do some researches: a geometric-structure conception model is put forward to derive the light information of the desert sparse vegetation; Based on this geometric-structure conception model, the identify of the desert riparian forest type was discussed with classification trees、Geometric Optical model and Spectral Angle Mapper, and the sub pixel structure pattern are developed based on the MODIS pixels, the spectral values of the end member of ground objects are derived from the EO-1 Hyperion hyper spectrum image, the virtual MODIS data are redeveloped using the geometric-structure models of pixels, and they are validated with the MODIS data in the same period. Finally, the Percent vegetation cover remote sensed models of the sparse vegetation were built, and the scale transfer effect of the percent vegetation cover information was analyzed. Based on the appearance and physiological-biochemical characteristic, the study regards that it is very important to increase the accuracy of the desert vegetation’s classification and raise the temporal rate of the physiological-biochemical characteristic inversion. So, it is put forward to decompound the mix pixel using high spatial resolution image, obtaining the sub pixel structure pattern based on the MODIS pixel with linear model, and deriving the physiological-biochemical characteristic with the high temporal MODIS data, and it is the theoretical basis of the geometric-structure conception model. The research on the classification of the desert riparian forest include the classify system and pixel unmixing. Firstly, a new sparse vegetation classification was bring forward based on the target was divided to be object and background from the observation of the remote sensing. Secondly, non-desert vegetation information was masked off using the pixel unmixing and classification tree, the spectrum of the pixel was simulated with the Geometric Optical-Radiative Transfer model based on ground observation, mapping the vegetation of the study with unmixing the desert vegetation based on the pixel spectrum simulated. The results indicate that the classify quality are well with the accuracy coefficient 77.66% and Kappa coefficient 0.704. The research on the subpixel structure and image simulation based on MODIS pixel show that the resolving power to the object from image was weaken and the borderline of the object was thicken with the pixel scale extending. Statistic eigenvalues, such as max, min, mean and standard deviation are rarely affected by the scale extend. So, the multi-scale subpixel structure based on MODIS 250m, 500m and 1000m pixel with liner model make sense. And the study on the comparison between the virtual image and the MODIS at the same period indicated that there is a limit for spectrum scale extends form 500 meter to 1000 meter, significant difference from reflectivity show that the 500 meter resolution image can not be effectively down-scale to the 1000 meter resolution image. The deriving of the desert riparian forest coverage was analyzed form remote sensed model and scale effect. Firstly, empirical models were built using ground measurement and vegetation indexes, such as NDVI, RVI, DVI and RDVI. And then, four kinds of models, normalized difference vegetation index (NDVI) model, subpixel structure model, maximal gradient difference model and adjusted maximal gradient difference model, were adopted to derive the coverage form TM data. The result from the precision test to models mentioned above show the adjusted modified maximal gradient difference model prediction is match well the field measurement. Secondly, the virtual multiscale coverage image, such as 500 meter and 1000 meter resolution data, are generated using the simple mean scale extend method to verify the inversion information from MODIS data, and result indicated that the adjusted maximal gradient difference model seems to be a viable method at multiscale for the sparse vegetation