【作者】 王晓飞；

【导师】 张晔；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2011， 博士

【摘要】 高光谱遥感是在测谱学基础上逐渐发展起来的新型遥感技术,除了空间图像信息外,其所具有的精细光谱信息,克服了宽波段遥感探测的局限,被广泛应用于多种领域,成为对地观测最重要的信息源之一。但由于成像原理与制造技术等因素的限制,高光谱图像的空间分辨率相对较低,给进一步应用,如特定目标的检测识别带来一系列的问题。为此,论文分别从信息融合和混合像素分解角度研究了高光谱图像的分辨率增强方法,旨在提高基于图谱结合的高光谱图像目标检测的性能。论文首先对遥感成像中涉及到的电磁波理论进行简单的介绍,分析了遥感图像的空间分辨率与光谱分辨率间的关系,即随着光谱分辨率的增加,在CCD等性能参数不变下,遥感图像的空间分辨率下降的原理。并在介绍了高光谱图像特性的基础上,对PCA、MNF及LDA变换的降维算法的原理进行了分析,研究其各种算法的特点及应用范围。降维算法是重要的高光谱图像预处理技术,这一部分的工作为后文的开展打下一个基础。然后对常用的高光谱图像目标检测算法进行了介绍。通过对支持向量数据描述的研究,分析并验证了其单类分类的性能及其适用范围;针对传统纯像素目标检测算法大部分无法解决目标与背景样本数量不平衡的问题,论文提出了基于SVDD的高光谱图像目标检测算法,把目标检测问题转化为单类分类问题。实验结果表明,与经典的光谱角度制图和有约束能量最小化算法相比,该算法仅需要较少的目标类训练样本就可以得到与前两者相近的检测结果,当增加背景样本时,本文方法可以将目标更容易的从背景中分离出来,为利用空间信息进一步检测提高了便利,使最终的检测结果优于上述两种算法。针对空间分辨率的不足,论文借助于空间信息补偿的思想,提出了基于相关向量机的增强高光谱图像分辨率的数据融合算法。由于该方法需要利用其它图像的信息,因此首先研究了多图像间的配准技术,并在对现有配准算法进行改进的基础上,提出了基于高斯拟合的配准控制点提取算法,从而获得了高精度的配准结果。在此基础上,研究了辅助信息补偿分辨率的方法,提出了基于RVM的融合算法,在增强高光谱图像空间信息的同时,较好地保持了原光谱特征。将增强后图像应用于纯像素小目标检测的实验表明,论文算法可以解决由于高光谱图像空间分辨率不足而导致的检测效果不佳的问题,分辨率增强后图像的检测精度明显优于融合前各图像的检测结果。最后,在缺少辅助信息的情况下,论文研究了通过光谱解混来改善高光谱图像空间分辨能力、解决光谱混叠的问题;利用子像素制图技术来实现高光谱图像空间分辨率的增强。对光谱解混的两个主要步骤:端元提取和混合像素分解分别进行了系统的研究。针对原始N-FINDR算法提取光谱端元时对噪声影响敏感的问题,提出了基于无监督聚类的端元提取算法。该算法利用K均值聚类方法从高光谱数据中选出光谱曲线代表集,再从代表集中找到光谱端元,实验表明该算法具有较强的抗噪性。针对传统算法在混合像素分解时,在含未知地物的像素处解混结果易出现较大偏差的问题,提出了一种基于SVDD的高光谱图像混合像素分解算法。该算法首先利用SVDD将高光谱数据分成完全由已知地物数据混合的像素和包含未知地物的像素两类,两类交界处为已知地物和未知地物混合的数据,然后对这些像素点进行混合像素分解,实验结果表明该算法可以有效地解决因存在未知端元对解混精度的影响,而且能给出未知端元的解混分量。在得到高精度的解混分量图的基本上,提出了基于感兴趣目标的子像素制图技术来改善高光谱图像的空间分辨率,经实验验证,该方法对检测出的目标形状保持较好,处理后的图像可以更容易地被利用光谱-空间信息联合的方法检测出目标。更多还原

【Abstract】 Hyperspectral remote sensing is a technique based on the spectroscopy, which contains abundant spectral information besides the spatial information of the images, and overcomes the limitations of the wide-band remote sensing detection. As a result, it has been widely used in many areas and becomes one of the most important earth observation information source. However, with the limitations such as imaging spectrometer manufacturing techniques and imaging principles, the spatial resolution of the hyperspectral images is relatively low. This brings a series of problems to the further applications, such as detection of the specific targets. Therefore, in this thesis, the methods of information fusion and mixed pixel decomposition have been researched to enhancement the image spatial resolution, and finally to improve the capability of the hyperspectral image target detection combining the spatial and spectral information.Firstly, this thesis introduces the principle of electromagnetic wave, analyses the relationship between spatial resolution and spectral resolution. For the same CCD, the spatial resolution is decreasing with the increase of spectral resolution. This thesis also introduces the characteristic of hyperspectral image, analyses the characteristic and applicable range of several dimension reduction algorithms, such as PCA, MNF and LDA. Dimension reduction is an important image preprocessing technique which lays the foundation for the subsequent work.And then, this thesis introduces the widely used target detection algorithms, analyses the support vector data description(SVDD) and its classification capability and applicable range. Aiming at the unbalanced problem between the target and background samples in most traditional pure-pixel target detection algorithms, a hyperspectral image target detection algorithm based on SVDD has been proposed in this thesis, which converts target detection to one-class classification problem. Experimental results show that compared with the traditional algorithms such as spectral angle mapping and constrained energy minimization, this algorithm needs less target training samples and achieves approximately the same results as these two methods, when increasing the background samples, this method is superior to those two ones.Aiming at low spatial resolution of hyperspectral image, by means of spatial information compensation, a data fusion algorithm of enhancing hyperspectral image resolution based on relevance vector machine(RVM) is proposed in this thesis. This algorithm needs the information from other source, so the registration techniques between multi-images is firstly studied. An algorithm of registration control point extraction is proposed, which based on the Gaussian fitting and a RVM geometric model, and then a more precise registration result can be obtained. In terms of supplementary information compensation method, a fusion algorithm based on RVM is proposed, which aims at enhancing the spatial information of the hyperspectral images, and at the same time maintaining the original spectral characteristics. When applying the enhanced images to a pure-pixel small target detection, the result showed that the proposed method can solve the problem of poor detection results caused by the low resolution of the hyperspectral images. The detection precision of the enhanced resolution image is much better than the original one.In the condition of lack of assistant information, spectral unmixing techniques has been researched in order to enhance the spatial resolution of the hyperspectral images and solve the spectral mixing problems. Mixing pixel decomposition includes two key steps: endmember extraction and abundance solving. Due to the noise sensitivity for the spectral endmember extraction in the original N-FINDR algorithm, a new endmember extraction algorithm is proposed on the basis of unsupervised clustering algorithm. This algorithm utilizes K-means clustering to select a representative spectrum from hyperspectral data, and then select the endmember from these spectra. Experimental results showed that the algorithm has a strong anti-noise performance. On the conditions that the existence of some unknown features contained in many pixels, unmixing results on these points will have a large deviation. An unmixing method for hyperspectral image based on SVDD is proposed. It firstly divides the hyperspectral data into two categories by SVDD, one is the mixed pixels with known materials, and the other contains unknown materials. The boundaries between them are considered to be a mixed pixel of the known and unknown ground covers. And these pixels are then decomposed. Experimental results show that the algorithm can effectively solve the low unmixing precision problem caused by the unknown pixels, and can give the mixing factor for the unknown pixels. Based on the high accurate unmixing fraction map, a sub-pixel mapping technique for interested targets is adopted to improve the spatial resolution. The experimental results show that the method can preserve the target shape better, and the targets in the processed image are prone to be detected by combining the spatial and spectral information.更多还原