【作者】 张立福；

【导师】 张良培；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2005， 博士

【摘要】 多时相、多传感器卫星数据为我们在区域乃至全球尺度的环境变化监测提供了丰富的信息。Landsat/TM(ETM+)、Terra(Aqua)/MODIS、ADEOS-Ⅱ/GLI以及其他一些传感器为我们提供了大量多/高光谱数据。由于不同传感器在波段数、波长范围以及中心波长位置等方面存在差异,数据分析结果依赖于传感器,尤其是受波段数和波长的影响。因此,不同传感器得到的分析结果很难进行比较。 本文发展了一种通用光谱模式分解算法(UPDM)。UPDM是一种与传感器无关的多/高光谱遥感数据分析算法。“与传感器无关”是指对同一样本的数据分析结果不受传感器的限制。许多分析方法是依赖于传感器的。例如,主成分分析方法是一种多元统计方法,通过剔除冗余数据来压缩多光谱数据。虽然PCA可以应用于各种传感器,但其结果依赖于传感器。模式分解方法(PDM)是一种线性光谱混合分析方法,它把原始遥感数据每个像素的光谱分解为固定的标准模式(水体、植被、土壤)。虽然PDM方法可以应用于不同的传感器,但PDM系数随传感器而变,即使分析的是同一个目标。 本文第三章介绍了作者发展的通用光谱模式分解算法(UPDM)。卫星传感器记录的每像素的光谱数据,可以通过标准UPDM转换矩阵,转换成用三个(或四个)UPDM特征分量表示。标准模式定义的光谱范围为350～2500nm,即太阳辐射能量波长范围。当应用于不同的传感器时,从标准模式中选择出与传感器波段对应的数值组成转换矩阵即可。一般情况下,使用三种标准模式时,地表反射信息的95.5%可以用转换后的三个UPDM成分表示。每自由度的光谱重构误差为4.2%。但是,对于某些目标,如黄叶,光谱重构误差较大。根据具体的研究目的,可以选择一种附加模式,标准模式数量增加到4个。如果研究目的是植被变化,需要增加黄叶作为附加模式。本研究选择它作为附加模式。 实验选择了652种地面样本的光谱测量数据来研究UPDM。样本包括植被绿叶、黄叶、枯叶,各种土壤,各种水体,建筑材料,等等。利用这些地面测量光谱数据模拟了不同的传感器数据并比较他们的UPDM分析结果,卫星传感器包括Landsat/MSS、ALOS/AVNIR-2、Landsat/ETM+、Terra/MODIS和ADEOS-Ⅱ/GLI等。实验结果证明,UPDM特征转换产生的新特征具有“与传感器无关”的特性。即不同传感器得到的UPDM系数几乎相同。 针对不同的研究目的,人们设计出了各种植被指数。NDVI是一种被普遍采用的植被指数。但NDVI只利用了近红外和红色两个波段。EVI不仅利用了近红外和红色两个波段,还增加了蓝色波段用于纠正气溶胶对红色波段的影响,以及其他几个纠正因子来抵抗气溶胶的影响。基于PDM的植被指数VIPD在反映植被植土覆盖百分比、植被垂直密度和植被类型方面比NDVI和EVI有优势。但PDM含有依赖于传感器的参数。同样不能直接比较不同传感器得到的结果。更多还原

【Abstract】 Multi-temporal and multi-sensor satellite data supply a wealth of information for monitoring environmental changes at regional, continental, and global scales. Larger volumes of multi-spectral data have become available from Landsat/TM (ETM+), Terra (Aqua)/MODIS, ADEOS-II/GLI and other sensors. The characteristics of each sensor differ, as the number of bands, the band wavelengths and the central wavelength of each band vary by satellite. Thus, analysis results depend on sensor performance and are especially affected by the number of bands and wavelengths observed. Consequently, it is difficult to compare analysis results obtained using data from different satellite sensors.This paper developed a universal pattern decomposition method (UPDM). The UPDM is a sensor-independent method that is tailored for satellite data analysis. Sensor independence means that analysis results for the same sample should be the same or nearly the same, regardless of the sensor used. Most analysis methods are sensor dependent. For example, the principal component analysis (PCA) is a multivariate statistical method used to compress multispectral datasets by removing redundancy in such a way that each successive PC has a smaller variance. Although PCA method can be applied to data obtained from any type of optical sensor, the results differ depending on the sensor type, so PCA is a sensor-dependent method. The pattern decomposition method (PDM) is a type of spectral mixing analysis, which expresses the spectrum of each pixel as the linear sum of three fixed, standard spectral patterns (i.e., the patterns of water, vegetation, and soil). The PDM can be applied to data obtained from any satellite sensor. However, the resulting pattern decomposition coefficients may differ by sensor, even for the same sample object.In Chapter 3, we developed a universal pattern decomposition method (UPDM). Sets of spectral reflectance measured by a sensor are transformed by the UPDM into three or four coefficients. The "universal standard spectral patterns" are determined in the spectral region between 350 nm and 2500 nm (the solar reflected wavelength region). Sensor wavelength values are selected from the universal standard spectral patterns to construct the transform matrix of each sensor. On average, 95.5% of land-cover spectral reflectance information can be transformed into the three decomposition coefficients and decomposed into the three standard patterns with about 4.2% error per degree of freedom. However, some objects, such as yellow leaves, have slightly larger decomposition errors. Depending on the research purpose, supplementary standard patterns can be applied. For example, to study vegetation changes in more detail, an additional supplementary spectral pattern can be added to reproduce the spectral reflectance. In this study, a yellow-leaf pattern was used as a supplementary spectral pattern.To study sensor independence using the UPDM, we analyzed about 652 ground-measured samples, including green-leaf, yellow-leaf, dead-leaf, soil, water, and concrete samples etc.. We made simulated data of Landsat/MSS, ALOS/AVNIR-2, Landsat/ETM+, TerraMODIS, and ADEOS-II/GLI sensors using ground-measured data, and compared the analyses results of these data. The results demonstrated that the pattern decomposition coefficients obtained using the UPDM are nearly sensor independent.Various vegetation indices have been developed for specific research objectives. A commonly used index is the normalized difference vegetation index (NDVI). However, it uses only red and near infrared reflectance data. The enhanced vegetation index (EVI) uses the red and near infrared bands, and also includes blue-band reflectance data to correct for aerosol influences in the red band, and some other aerosol resistance coefficients. The vegetation index based on the PDM (VIPD) is more sensitive than the NDVI for determining the vegetation cover ratio, vertical vegetation thickness, and vegetation type. However, the PDM has sensor-dependent parameters. It is difficult to directly compare results obtained using data from different sensors.In Chapter 4, we proposed a new vegetation index based on the universal pattern decomposition method (VIUPD). The VIUPD is defined as a linear sum of the pattern decomposition coefficients but is sensor independent. The VIUPD has many benefits over the conventional VIPD. We compared how our new vegetation index (VIUPD), the NDVI, the EVI, and the VIPD represent the relationships between photosynthesis, the vegetation area ratio, and the number of overlapping leaves. Results demonstrated that the VIUPD reflected vegetation concentrations, the amount of CO2 absorption, and the degree of terrestrial vegetation vigor more sensitively than did the NDVI and EVI. The NDVI and EVI became more rapidly saturated as a function of PAR. The VIUPD is more suitable for multi-spectral analysis than the EVI, NDVI, and VIPD.For validation of the UPDM, in Chapter 5, four UPDM coefficients were computed using Landsat/ETM+ and Terra/MODIS data observed over the Three Gorges region. Vegetation indices were computed in the same multi-dimensional space. UPDM coefficients computed with 6-band ETM+ data, with wavelengths between 350 and 2500 nm were compared to coefficients computed with MODIS data in bands 1 to 7. Both datasets were re-sampled to a spatial resolution of 484.5 m. The DN value was converted to a reflectance value by considering radiometric calibration and atmospheric correction. Reflectance values are the input vector for calculating UPDM coefficients.The four UPDM coefficients derived from the satellite data are independent of the sensor. The independence of Cs and Cv is better than Cw and C4, because both Cw and C4 have values near zero. Consequently, any small bias will move them far from a linear line. UPDM coefficients and vegetation indices (VIUPD, NDVI, and EVI) were computed using 3x3 pixel averages to evaluate the effect of pixel spatial location errors. Coefficients and vegetation indices computed this way both showed smaller root mean square (rms) values. Results also suggest that the VIUPD is sensor-independent, especially in areas with little topographic influence.In Chapter 6, we proposed a new classification method based on the UPDM. Sets of spectral reflectance measured by a sensor are transformed by the UPDM into three coefficients with three fixed spectral reflectance patterns. This paper considered ETM+ data for a classification study. The satellite digital signal number (DN) was first converted to reflectance value after adjusting for the influence of Rayleigh scattering. Application of the UPDM to satellite reflectance data reduces the number of UPDM features from the original hyper-multi dimensional data. Classification results are compared to classification accuracy from PCT using MDC (using minimum Euclidean distance and minimum Mahalanobis distance) and MLC algorithms. The classification used PCT and UPDM were similar. Unlike PCT components, UPDM components have physical meanings. Classification results using UPDM are sensor-independent, which are very significant for comparison of results derived from different data.更多还原