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】             高连如                                                                                                                                                  

【导师】             童庆禧；张兵；郑兰芬

【 学位年度 】    2007

【论文级别】      博士    

【关键词】         高光谱遥感，目标探测，光谱降维，图像降噪，特征提取 

【Key words】  Hyperspectral Remote Sensing, Target Detection, Dimensionality Reduction, Noise Reduction, Feature extraction 

【中文摘要】

     高光谱遥感图像具有几何、光谱和辐射三重属性，这为目标的精确探测和识别带来了可能。高光谱遥感数据所具有的精细光谱信息为其目标探测应用提供了巨大的优势，但同时由于探测器件的限制，高光谱图像往往不具有较高的空间分辨率，感兴趣目标一般都以低概率出露的形式存在，这就使得传统的基于空间特征的目标探测方法在高光谱遥感目标探测中难以发挥效能。因此如何扬长避短和有效利用这些信息对目标探测效果和精度有着非常重要的影响，这就需要面向目标探测进行信息增强与特征提取技术研究。

     本论文在系统总结高光谱遥感目标探测技术发展现状的前提下，针对目标探测中的光谱不确定性、降维与降噪处理、混合像元与端元提取、低概率出露目标探测四个关键问题进行了探讨，并在此基础上进行了高光谱图像噪声评估、高光谱数据降维和特征提取技术研究，分析了光谱不确定性、遥感器系统特性和数据预处理对目标探测的影响。在综合考虑以上关键技术和影响因素的基础上，发展了更稳定且更具有普适性的面向低概率出露目标的探测算法。论文主要研究成果和结论如下：

1. 研究了成像光谱仪的通用噪声模型，对传统的LMLSD在复杂地物覆盖图像中不适用的问题进行了分析，提出了基于边缘块剔除、高斯波形提取和残差调整的改进方法，同时提出了基于均匀区域划分的高光谱图像噪声评估方法。论文所发展的方法能适用于更广泛的地物覆盖情况。
2. 研究了目前应用较为广泛的高光谱数据降维方法，并就最大噪声分离方法中的若干关键问题进行了分析，在此基础上以标准主成份分析为出发点，从尺度效应、噪声影响、信息量损失和数据结构变化四方面对各种降维方法进行了比较分析，确定各方法的适用范围，制定了面向数据和目标探测应用的高光谱数据降维方法选择策略。
3. 研究了高光谱遥感中主观人为因素造成的光谱不确定性问题。研究结果表明，虽然人为“伪装”造成的光谱变异发生在整个波段范围，但是利用短波红外窗口仍可以完成对目标的识别。研究了涂层覆盖下目标板材光谱的不确性，发现涂层覆盖造成的目标光谱不确定性主要对光谱曲线幅度产生影响，曲线的波形变化不大，因此可以将波形特征作为主要度量指标进行目标探测。
4. 研究了高光谱数据获取过程和高光谱数据处理通用流程，并从涂层覆盖、光谱分辨率、光谱仪成像方式、空间分辨率、图像噪声和数据降维六个方面分析了这些因素对高光谱遥感目标探测的影响，提出了进行目标探测精度整体控制的新思路。
5. 研究了低概率出露目标存在的主要形式，并就高光谱遥感目标探测中普遍存在的混合像元问题进行了探讨，提出了基于抗噪声和部分约束独立分量分析的混合像元分解方法，并借鉴了多算法融合思想，综合利用空间连续性分割、单形体体积法、独立分量分析和RX探测算子，实现了高光谱图像的端元快速提取和异常目标探测。 

【Abstract】

Hyperspectral remote sensing images incorporate properties of geometry, spectra and radiance. These features make facilitate detection and recognition of targets, and largely promote the development of target detection using hyperspectral imagery. Hyperspectral remote sensing provides detailed spectral data for target detection and recognition, the data provides our subject opportunity. However, due to detectors’ capability limitations, hyperspectral images always do not have high spatial response. Therefore, interested targets are exposed with low probability commonly. Traditional target detection methods based on spatial features are invalid for hyperspectral images. How to enhance the strong points and how to use this information is an important issue for target detection in hyperspectral imagery. Thereby it is quite necessary to make comprehensive researches on spectra enhancement and feature extraction.

      Based on integrated study of the actuality of target detection using hyperspectral imagery, this dissertation analyzes four key problems in the subject, namley spectral uncertainty, dimension reduction and image denoising, mixed pixel and endmember extraction, low probability exposed target detection. The research is focused on noise estimate, data dimension reduction and feature extraction, along with that, analyzes influences of spectral uncertainty, remote sensing system and data pretreatment on target detection are also discussed. At last, more robust detection algorithms for low low probability exposed target are developed. The main aspects of this dissertation are as follows:

1. The general noise model in imaging spectrometer is discussed. To reduce the pressure of image content to traditional local means and local standard deviations method, three improved methods are developed, which are based on edges eliminate, Gaussian curve extraction and residual adjustment. At the same time, a new noise estimate method based on homogeneous region division is developed. This method is more reliable and adaptable, and works well for hyperspectral images with diverse land cover types.

2. This dissertation study widely used dimension reduction methods in hyperspectral imagery. Some key questions in maximum noise fraction are also analyzed. Based on analysis of characters of these dimension reduction methods, such as sensitivity to magnitude change and noise, information loss and data structure change, a dimension reduction method selection strategy for target detection is developed.

3. Spectral uncertainty due to human activities is studied. Results show that man-made camouflage change the spectra of the target in the full band ranges, but spectral characteristics in short-wave infrared range can be used to distinguish different materials under paints. Paints can only affect the amplitude of spectra, and the shapes of the spectra are fairly consistent. Therefore, spectral shape should be the main measurement to detect targets.

4. Hyperspectral imagery measurements and data processing are studied. Six aspects contain paints, spectral response, imaging spectrometer mode, image noise and dimension reduction are discussed. The influences of these factors to target detection are analyzed. This research provides a strategy to precision control of target detection in hyperspectral imagery.

5. The form of low probability exposed target existed in hyperspectral imagery are studied. The phenomenon of mixed pixel in hyperspectral target detection is also discussed. A low noise sensitive and partial restricted independent component analysis method is developed for decomposition of mixed pixels. Based on the idea of multiple algorithm fusion, spatial continuity, convex volume, independent component analysis and RX detector are combined together to extract endmember quickly and  to detect target accurately.