融合相关系数和超像素联合稀疏表示的高光谱图像分类

Hyperspectral Image Classification Based on the Combination of Correlation Coefficient and Super-Pixel Sparse Representation

摘要: 高光谱图像含有丰富的光谱和空间特征，在传统方法中多使用高光谱图像的光谱特征而忽略其空间特征。联合稀疏表示解决了基于表示方法中未使用空间信息的问题，但存在对噪声点敏感和空间信息利用不足的问题。为了解决这个问题，本文中提出了考虑像素间相关性和空间领域信息的高光谱图像分类方法，该方法首先使用熵率分割将图像划分成不同大小的超像素区域。然后计算测试样本与训练样本间的相关系数，并基于超像素区域进行联合稀疏表示，获得稀疏系数。最后根据最小重构残差和相关系数对测试样本进行决策融合分类。为了验证所提方法的有效性，本文使用基准高光谱数据Indian Pines进行实验。实验结果表明本文提出的方法有效提高了高光谱图像的分类精度。

Abstract: Hyperspectral images are rich in spectral and spatial features. In traditional methods, the spectral features of hyperspectral images are often used while the spatial features are ignored. Joint sparse representation solves the problem of not using spatial information in the representation method, but it is sensitive to noise points and insufficient use of spatial information. In order to solve this problem, this paper proposes a hyperspectral image classification method which considers the correlation between pixels and spatial domain information. The method firstly uses entropy rate segmentation to divide the image into different sizes of super-pixel regions. Then, the correlation coefficient between the test sample and the training sample was calculated, and the sparse coeffi-cient was obtained by joint sparse representation based on the super-pixel region. Finally, ac-cording to the minimum reconstruction residual and correlation coefficient, the test samples are classified by decision fusion. To verify the effectiveness of the proposed method, Indian Pines, benchmark hyperspectral data, was used in the experiment. Experimental results show that the proposed method can effectively improve the classification accuracy of hyperspectral images.