【作者】 何同弟；

【导师】 李见为；

【作者基本信息】 重庆大学 ， 仪器科学与技术， 2014， 博士

【摘要】 高光谱遥感是当前遥感技术发展的一个前沿领域，它利用很多很窄的电磁波波段从感兴趣的物体获得有用信息。高光谱图像作为遥感领域的一项重大突破，在保留较高空间分辨率同时，其光谱分辨率有极大的提高，达到了纳米的数量级，可以用来探测和识别传统全色和多光谱遥感中不可探测的地物类别。与传统的多光谱遥感图像相比，高光谱遥感图像有着信息量大、光谱分辨率高等特点，这使得在描述与区分地物类别方面的能力有了大幅提高，进而为地物光谱信息的精确处理与分析提供了可能。高光谱遥感系统已在全球许多国家的先进对地观察遥感系统中占有重要的位置，己成为地球陆地、海洋、大气观察的生力军。但是由于高光谱图像具有较高的数据维数，常规的图像分类方法在处理高光谱图像时有较大的限制，如何从大量的高光谱数据中快速而准确地挖掘出所需要的信息，实现高精度的分类，仍是一个亟待解决的问题。本文从高光谱图像数据的特点入手，在对现有算法进行分析的基础上，针对高光谱遥感图像分类算法进行深入研究。主要的研究工作如下：①在对高光谱遥感影像进行预处理之后，对所用高光谱图像做了大气校正。几何校正选取为二次多项式模型，重采样采用的是最近邻插值法，精度方面的要求得到了充分保证，为下一步的正确分类打下了坚实的基础。②提出了一种基于自适应粒子群优化算法的RBF神经网络高光谱遥感图像分类方法。由于人工神经网络具有并行处理、模糊识别和非线性映射等优点，很适合高光谱图像分类，但是其参数难选。采用自适应粒子群优化算法对RBF神经网络的参数进行了优化，建立了基于粒子群优化算法的的RBF神经网络模型，分类实验结果表明了基于粒子群优化的RBF神经网络模型具有很高的分类精度。③提出了一种基于自适应粒子群优化算法的SVR高光谱遥感图像分类方法。首先分析了支持向量回归的核函数的构造和模型参数的优选问题。由于本文数据样本较少，模型参数优选的比较复杂，本文采用了CV估计模型推广误差，并使用自适应粒子群优化算法来优选SVR模型参数，构建了基于粒子群优化算法的SVR高光谱遥感图像分类模型，在一定程度上解决了高光谱数据标记样本不足的问题。④从稀疏表示的基本理论出发提出了一种基于自适应稀疏表示的高光谱分类方法。利用训练样本构建字典，聚类每一步迭代所产生的余项，将聚类中心作为新的字典原子，然后将测试样本看成冗余字典中训练样本的线性组合，令字典能够更适应于样本的稀疏表示。通过对高光谱图像的分类实验，验证了自适应稀疏表示算法的有效性。更多还原

【Abstract】 Hyperspectral remote sensing is a frontier field in the development of remotesensing technology, and it obtains useful information from interesting objects by usingnarrow bands of electromagnetic waves. Hyperspectral remote sensing image is a majorbreakthrough in the field of remote sensing, and spectral resolution was greatlyimproved while retaining a high spatial resolution. As a result, it can be used to detectthe land-cover types which could not be detected with traditional panchromatic andmulti-spectral images. Compared with the traditional multi-spectral remote sensingimage, hyperspectral image (HSI) provides a large amount of information and highspectral resolution, and the ability in description and analysis of land-cover types hasbeen improved greatly, which makes it possible to process and analyze the spectralinformation of land-cover types more precisely. In many countries, hyperspectral remotesensing play an important role in the earth observation system, and it has become a newforce to the observation of earth’s land, ocean and atmosphere. Due to the highdimension of HSI, traditional classification methods have some limitations whendealing with HSI. How to accurately exploit the information from a large amount ofhyperspectral data for achieving high-precision classification, is still a serious problem.Based on the characteristics of HSI, this paper focuses on the research of HSIclassification. The main research contents are as follows:①After the preprocessing of HSI, atmospheric correction was carried out on HSI.The quadratic polynomial mode was selected for geometric correction, and the nearestneighbor interpolation method was used for resampling. Then the requirements foraccuracy are fully guaranteed, which lays a solid foundation for classification.②A RBF neural network method based on adaptive particle swarm optimization(PSO) was proposed for HSI classification. The neural network has the merits ofparallel processing, fuzzy recognition, and nonlinear mopping, which gives the benefitto the classification of HSI. However, it is very difficult to choose proper parameters.An adaptive PSO algorithm is used to optimize the parameters, and the RBF neuralnetwork model based on PSO is constructed for classification. Experimental resultsshow that the RBF neural network based on PSO has a better classification accuracy forHSI classification.③The support vector regression (SVR) with adaptive PSO is proposed for HSI classification. At first, the construction of kernel functions and the optimization ofmodel parameters are introduced. For the conditions of small number of data samplesand the complexity of model parameters, the CV estimation model is used. The adaptivePSO method is used to choose the parameters for the model of SVR in hyperspectralremote sensing classification, which is useful to solve the problem of insufficienttraining samples in HSI.④According to the sparse representation theory, a new method based on adaptivesparse representation (ASP) is proposed for HSI classification. In this method, thedictionary is constructed with training samples. The remainder of each iteration isclustered, and the center of clustering is considered as the atom of the new dictionary.Furthermore, the testing samples are regarded as the linear combination of trainingsamples in the redundant dictionary, which makes the dictionary adapt to sparserepresentation of samples more effectively. The validity of ASP is verified by theexperiments for HSI classification. The effectiveness of the proposed ASP method isverified on hyperspectral image data sets.更多还原