基于稀疏性的高光谱图像亚像元目标检测研究

发布时间：2018-05-22 17:07

本文选题：高光谱图像 + 亚像元　；参考：《南京理工大学》2014年硕士论文

【摘要】：高光谱图像光谱分辨率高,具有图谱合一的特性,能够提供区分不同物质的诊断性光谱信息,结合该光谱信息可提高对目标和背景进行定量分析的能力,因此高光谱目标检测技术在目标检测领域具有独特的优势。由于地物分布情况复杂和成像光谱仪空间分辨率的较低等原因,待检测的目标通常与其他地物共同组成混合像元,此时目标以亚像元形式存在。高光谱亚像元目标检测是目标检测研究的前沿和难点,本文着眼于如何利用高光谱数据的稀疏性提高检测效果,对高光谱图像亚像元目标检测技术进行了研究,主要工作和成果如下： 1.研究了高光谱遥感图像的光谱混合模型,详细的介绍了多元信号估计和信号检测理论,推导了四种经典的亚像元目标检测算法——约束能量最小化方法(CEM)、基于加权样本自相关矩阵的CEM、正交子空间投影算法(OSP)、适应匹配子空间检测算法(AMSD),并在第四章用实验证明了这些算法的可行性。 2.给出了基于稀疏约束的线性混合光谱分解模型,推导了四种经典的稀疏性解混算法——正交匹配追踪算法(OMP)、迭代光谱混合分析算法(ISMA)、变量分离的增广拉格朗日算法(SUNSAL)、基于加权L1正则化的SUNSAL算法,给出了这四种算法的具体实现步骤,并提出了基于L1/2正则化的稀疏性解混方法。实验证明基于L1/2正则化的稀疏性解混方法在图像信噪比较小的情况下性能比其他四种算法好而且更加稳定。 3.利用高光谱数据的稀疏性进行目标检测。本文将稀疏性分解算法和自适应匹配子空间检测算法相结合提出了SU-AMSD算法,并介绍了该算法的基本流程与具体实现。为了提高目标检测的效率与精度,本文对该算法进行了改进,提出了一种基于Lib-IEA的SU-AMSD算法。实验证明,与上面介绍的四种经典亚像元目标检测算法相比基于光谱库和IEA算法的SU-AMSD算法的效果更好,而基于Lib-IEA的SU-AMSD算法则没有达到预期的目标。
[Abstract]:The spectral resolution of hyperspectral image is high, and it has the characteristic of unifying the spectrum, which can provide diagnostic spectral information that distinguishes different substances. Combined with this spectral information, the ability of quantitative analysis of target and background can be improved. Therefore, hyperspectral target detection technology has a unique advantage in the field of target detection. Due to the complex distribution of ground objects and the low spatial resolution of the imaging spectrometer, the target to be detected usually forms a mixed pixel with other ground objects, and the target exists in the form of sub-pixel. Hyperspectral sub-pixel target detection is the frontier and difficulty of target detection. This paper focuses on how to use the sparsity of hyperspectral data to improve the detection effect, and studies the sub-pixel target detection technology of hyperspectral image. The main work and results are as follows: 1. The spectral mixing model of hyperspectral remote sensing image is studied, and the theory of multivariate signal estimation and signal detection is introduced in detail. In this paper, four classical subpixel target detection algorithms, constrained energy minimization method, CEM based on weighted sample autocorrelation matrix, orthogonal subspace projection algorithm, adaptive matching subspace detection algorithm are derived. In chapter 4, we use the Experiments show that these algorithms are feasible. 2. A linear mixed spectral decomposition model based on sparse constraints is presented. Four classical sparse demultiplexing algorithms, orthogonal matching tracking algorithm, iterative spectral mixed analysis algorithm, extended Lagrangian algorithm with variable separation, SUNSAL algorithm based on weighted L1 regularization, are derived. The implementation steps of the four algorithms are given, and a sparse demultiplexing method based on L 1 / 2 regularization is proposed. Experiments show that the sparse demultiplexing method based on L1 / 2 regularization is better and more stable than the other four algorithms in the case of low SNR. 3. Target detection is carried out by using the sparsity of hyperspectral data. In this paper, SU-AMSD algorithm is proposed by combining sparse decomposition algorithm with adaptive matching subspace detection algorithm, and its basic flow and implementation are introduced. In order to improve the efficiency and accuracy of target detection, this paper improves the algorithm and proposes a SU-AMSD algorithm based on Lib-IEA. Experimental results show that the SU-AMSD algorithm based on spectral library and IEA algorithm is more effective than the four classical subpixel target detection algorithms mentioned above, but the SU-AMSD algorithm based on Lib-IEA does not achieve the desired goal.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】