基于广义双线性模型的高光谱解混

发布时间：2018-05-25 03:18

本文选题：高光谱图像解混 + 非负矩阵分解　；参考：《西安电子科技大学》2015年硕士论文

【摘要】：由于遥感数据集的空间分辨率,遥感成像仪在自然环境中的收集的光谱信号必然是各种物质的混合物。因此,准确估计需要光谱解混。混合像素分解方法按照所采用的分解模型,大致可以分为基于线型光谱混合模型的分解方法和基于非线性光谱混合模型的分解方法。本文考虑了丰度的稀疏性、空间信息和建模的多样性,改进了现有的非线性解混方法,具体如下:1.高光谱数据的相关性会导致数据的稀疏性,而且每个像素并非包含所有端元。而大多数现有的非线性解混算法没有考虑数据的稀疏信息。针对非线性解混算法没有考虑数据的稀疏性,提出了稀疏约束的广义双线性模型解混。广义双线性模型(GBM)已被广泛用于非线性高光谱图像解混。高光谱数据的高度相关性导致了丰度的稀疏性。目前正则化方法通常用来约束丰度的稀疏性,目的通过将丰度矩阵的稀疏性约束添加到GBM模型中,拓展semi-NMF,得到L_（1/2）约束半非负矩阵分解(L_（1/2）-semi-NMF)算法来估计丰度和非线性系数。将GBM分成的线性部分和二阶部分,并分别使用迭代优化算法优化。克服了半非负矩阵分解算法容易陷入局部最小点的缺点,收敛速度加快且不易限于局部最优解。在高光谱合成数据和真实数据上的实验结果表明:该方法提高了解混的稳定性和结果的正确性。2.在双线性场景中,植被和土壤之间通常发生多重散射,而包含植被和土壤等物质的高光谱图像在边界区域才可能发生双线性混合,考虑了图像区域差异性,提出基于区域自适应分割的高光谱图像解混方法。先用K均值聚类方法对高光谱数据聚类,将图像分割为匀质区域和细节区域。匀质区域采用线性模型,用稀疏约束的非负矩阵分解方法解混,细节区域采用广义双线性模型,用稀疏约束的半非负矩阵分解方法解混,很好的保持了双线性丰度的边缘信息。对比实验表明:所提出的方法有效提高了高光谱遥感图像的解混准确率。3.大多数现有的稀疏NMF算法对于高光谱解混只考虑欧几里得结构的高光谱数据空间。事实上,高光谱数据更可能位于一条嵌入高维空间的低维流形。针对非线性解混算法没有考虑高光谱数据内在的流形结构,提出了图约束的广义双线性模型解混。添加的图正则可以保持原始图像和丰度图之间的密切联系,改进的方法能改善解混性能。
[Abstract]:Because of the spatial resolution of remote sensing data sets, the spectral signals collected by remote sensing imagers in the natural environment must be mixtures of various substances. Therefore, accurate estimation requires spectral unmixing. According to the decomposition model, the mixed pixel decomposition method can be divided into linear spectral mixed model decomposition method and nonlinear spectral mixed model decomposition method. In this paper, the sparsity of abundance, the diversity of spatial information and modeling are considered, and the existing nonlinear demultiplexing methods are improved as follows: 1. The correlation of hyperspectral data leads to data sparsity, and not every pixel contains all endpoints. However, most of the existing nonlinear unmixing algorithms do not consider the sparse information of the data. In this paper, a generalized bilinear model with sparse constraints is proposed to solve the problem that the data sparsity is not considered in the nonlinear de-mixing algorithm. The generalized bilinear model (GBM) has been widely used in nonlinear hyperspectral image demultiplexing. The high correlation of hyperspectral data leads to the sparsity of abundance. At present, regularization methods are usually used to constrain the sparsity of abundance. Aim to estimate the abundance and nonlinear coefficients by adding the sparse constraint of abundance matrix to the GBM model and extending the semi-NMFs to obtain the LSP 1 / 2) constrained semi-negative matrix decomposition algorithm. The GBM is divided into linear part and second order part, and the iterative optimization algorithm is used respectively. It overcomes the shortcoming that semi-nonnegative matrix factorization algorithm is easy to fall into the local minimum point, and the convergence speed is accelerated and is not easy to be limited to the local optimal solution. The experimental results on hyperspectral synthetic data and real data show that the proposed method improves the stability of the mixture and the correctness of the results. In bilinear scenarios, multiple scattering usually occurs between vegetation and soil, while hyperspectral images containing vegetation and soil are likely to be bilinear mixed in the boundary region, taking into account the regional differences of the images. A method of hyperspectral image de-mixing based on region adaptive segmentation is proposed. The K-means clustering method is used to cluster the hyperspectral data, and the image is divided into homogeneous region and detail region. Linear model is used in homogeneous region, non-negative matrix decomposition method with sparse constraint is used to solve the problem, generalized bilinear model is used in detail region and semi-non-negative matrix decomposition method with sparse constraint is used to solve the problem. The edge information of bilinear abundance is well preserved. The experimental results show that the proposed method can effectively improve the accuracy of hyperspectral remote sensing images. Most existing sparse NMF algorithms only consider the hyperspectral data space of Euclidean structure for hyperspectral demultiplexing. In fact, hyperspectral data are more likely to be located in a low dimensional manifold embedded in a high dimensional space. In this paper, a graph constrained generalized bilinear model is proposed to solve the problem that the nonlinear unmixing algorithm does not take into account the intrinsic manifold structure of hyperspectral data. The added graph regularization can maintain the close relationship between the original image and the abundance graph, and the improved method can improve the demultiplexing performance.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP751

【参考文献】