基于邻域分割的空谱联合稀疏表示高光谱图像分类技术研究

传统的高光谱遥感影像分类算法侧重于光谱信息的应用。随着高光谱遥感影像的空间分辨率的增加，高光谱影像中相同类别的地物在空间分布上呈现聚类特性，将空间特性有效地应用于高光谱遥感影像分类算法对分类精度的提升非常关键。但是，高光谱影像的高分辨率提供空间聚类特性的同时，在不同地物边缘处表现出的差异性更加明显，若不对空间邻域像素进行甄选，直接将邻域光谱信息引入，设计空谱联合稀疏表示进行图像分割，则分类误差较大，收敛速度大大降低。将光谱角引入空谱联合稀疏表示图像分类理论中，提出了一种基于邻域分割的空谱联合稀疏表示分类算法。该算法利用光谱角计算相邻像素的空间相似度，剥离相似度较低的邻域像素，将相似度高的邻域像素定义为同类地物，引入空谱联合稀疏表示模型中，采用子联合空间追踪算子和联合正交匹配追踪算子对其优化求解，以最小重构误差为准则进行分类。选取AVIRIS及ROSIS典型光谱影像数据进行实验仿真，从中可以看出，随着光谱角分割阈值的提高，复杂的高光谱影像分类精度和平滑区域的高光谱影像分类精度均逐步提高，表明邻域分割在空谱联合稀疏表示分类中的必要性。

A Novel Spatial-Spectral Sparse Representation for Hyperspectral Image Classification Based on Neighborhood Segmentation

Traditional hyperspectral image classification algorithms focus on spectral information application,however,with the increase of spatial resolution of hyperspectral remote sensing images,hyperspectral imaging presents clustering properties on spatial domain for the same category.It is critical for hyperspectral image classification algorithms to use spatial information in order to improve the classification accuracy.However,the marginal differences of different categories display more obviously.If it is introduced directly into the spatial-spectral sparse representation for image classification without the selection of neighbor-hood pixels,the classification error and the computation time will increase.This paper presents a spatial-spectral joint sparse representation classification algorithm based on neighborhood segmentation.The algorithm calculates the similarity with spectral angel in order to choose proper neighborhood pixel into spatial-spectral joint sparse representation model.With simultaneous subspace pursuit and simultaneous orthogonal matching pursuit to solve the model,the classification is determined by computing the minimum reconstruction error between testing samples and training pixels.Two typical hyperspectral images from AVIRIS and ROSIS are chosen for simulation experiment and results display that the classification accuracy of two images both improves
as neighborhood segmentation threshold increasing.It concludes that neighborhood segmentation is necessary for joint sparse representation classification.