中国科学院遥感与数字地球研究所-高光谱遥感应用技术研究室

【作者】             刘建贵                                                                                                                                             

【导师】             童庆禧；郑兰芬

【 学位年度 】    1999

【论文级别】      博士    

【关键词】          高光谱遥感，城市地物，人工目标，数据定标，反射率转换，CA变换，正交变换，匹配滤波，目标检识 

【Key words】   the vegetation index time series of remotely sensed data, Harmonic Analysis, HANTS, number of frequency, outlier detection 

【中文摘要】

利用高光谱数据进行城市遥感是对传统遥感数据源如航片、多光谱遥感数据等的有力补充。可以利用高光谱遥感数据丰富的光谱信息，从光谱分析与光谱匹配技术出发对城市地物和人工目标进行精细分类，提供城市规划、环境监测与评估、城市变迁以及相关的社会经济等方面的信息。研究高光谱遥感数据处理和分类方法与模型以及利用高光谱遥感数据对城市地物及人工目标分类是本文的主要目的。

        本文从以下几个方面对高光谱城市遥感进行了探讨：

1. 研究了高光谱仪器系统的成象机理和模型、高光谱数据处理模型与算法。从高光谱遥感机理出发，通过分析光学成象原理和仪器模型，深入探讨了传感器定标、辐射和大气校正等问题。讨论了各种反射率转换方法，将它们归结为三种方法：辐射传输方法、归一化方法和经验方法。介绍了光谱匹配的概念和光谱数据库以及它们在地物识别中的重要性，并采用经验线性法对城市高光谱图象进行了高精度反射率转换。

2. 研究了城市地物及人工目标的光谱特征在高光谱图象上的表征。分析了各种城市地物的实验室光谱特征以及它们在高光谱图象中的表达，比较研究了高光谱数据与TM重采样数据的差异，认为高光谱数据对于反映城市地物详细的光谱特征明显优于传统的多光谱遥感数据，城市遥感中可以利用高光谱数据区分复杂多样的地物覆盖类别。

3. 发展了光谱特征提取方法。通过对高光谱遥感数据特点的分析，说明特征提取的必要性，指出对于地物分类来说，不但类别样本的一阶统计量，而且二阶统计量以及波段之间的协变化中都包含着重要的分类信息。在考虑了类别样本分布的一阶及二阶统计量的基础上，提出采用类内与类间相对距离作为特征提取的衡量标准，通过改进CA方法，对城市地物的高光谱图象数据进行了光谱维特征提取和波段选择，并通过巴氏距、散点图来说明特征提取的效果，并对几种城市地物特征提取进行了实验研究。

4. 发展了一个适用于高光谱图象的目标检测与分类算法。将高光谱图象象元光谱看为信号序列，引入信号检测理论进行高光谱图象中地物检识与分类；提出采用“空间不变、线性可加图象序列”来描述高光谱图象，并采用SD滤波器、正交子空间投影等方法进行城市地物在全波段的匹配滤波检测与分类。采用该方法对模拟数据进行了系统的分析研究，并对城市高光谱遥感数据进行了分类处理。讨论了投影算子作用于图象数据的归一化输出值与象元中各种物质成分含量之间的关系，以及它们可能受到的影响。从一个全新的观点出发进行图象分类和分类结果评价，对于特征不突出的城市地物分类是非常有效的，同时它可以估计出微弱目标的存在，表明采用信号分析技术进行图象分类是可行的。

【Abstract】

In urban remote sensing,  hyperspectral data is a potent supplement to traditional data sources ,such as aerial photograph and  multispectral data. By using the detailed spectral information contained in hyperspectral data, fine classification of urban and man-made objects can be achieved through spectral matching techniques. Moreover, information for urban planning, urban change, environment monitoring and assessment,  and corresponding social and economical information can be extracted. The  aim of this dissertation is to develop methods and models for  hyperspectral data processing and classification, and  to use hyperspectral data for the recognition and classification of urban and man-made objects. This dissertation concentrate on the following aspects,

1. The basic principles and models for hyperspectral imaging systems, the methods and algorithms for hyperspetral data processing were studied.   First, base upon the sensor and optical models the problems on sensor calibration, radiometric and atmospheric correction were studied. The various methods for reflectance conversion methods were discussed and further reduced into three kinds, the methods based upon radiative transfer theory ,  the normalization methods  and empirical  methods. The spectral matching concept were introduced. Several spectral library and tools for spectral analysis were presented. By using the empirical line method, reflectance conversion of urban hyperspectral data were made with high precision.

2. The spectral representation of urban and man-made objects  in  hyperspectral data were analyzed.  First, the optical spectral characteristics of  urban and man-made objects were studied by using spectra measured on the  ground. Second,  the spectral information represented in hyperspectral data were studied. Third , the comparison between  high spectral resolution data and multispectral data were made.  These studies indicate that the diversity and complex urban land cover types can be distinguished by using hyperspectral data.

3. A method for spectral feature extraction of urban and man-made objects was developed. Feature extraction is necessary for hyperspectral data processing because of the correlation and dispersion of discriminating information between and among the numerous spectral channels. The importance of the  first and second order spectral statistical characteristics for classification were discussed. Improved CA transformation, a new feature extraction method, is developed, which is based upon the relative distance between and within class pairs. By using Bhattacharyya distance and scatter plot, the effectiveness of the feature extraction method was evaluated.

4. A detection and classification method, subspace projection and matched filtering were developed for hyperspectral data analysis. Pixel spectral can be considered as signal series, therefore, theory from signal detection can be introduced into hyperspectral data processing and classification.  A “Spatially invariant, linearly additive” data model was introduced to discribe hyperspectral data. By using SD filter, orthogonal subspace projection technique and matched filters, the desired objects can be detected and further classified. Urban and man-made objects were classified using this method. The results make it clear that this method is rather efficient for hyperspectral data classification.

5. A strategy for detection and classification of urban and man-made objects  were developed based on the spectral characteristics. This method classify the various land cover types hierarchically.  As a testimony of the methods provided, a PHI and a HyMap hyperspectral images were experimented.