高光谱图像分类是高光谱研究的重要内容, 也是许多其他应用的前提。 针对传统高光谱图像分类仅考虑光谱信息而忽略空间信息的问题, 对距离依赖的中餐馆模型(distance dependent Chinese restaurant process, ddCRP)进行改进, 提出一种基于空-谱约束的中餐馆过程混合模型(spatial-spectral Chinese restaurnt process, ssCRP)用于高光谱图像聚类。 该模型充分考虑像素邻域的空间和光谱信息, 并将其统一纳入模型的建模及求解过程中, 得到一般基于像素的聚类方法无法实现的效果, 可在一定程度上满足高光谱图像聚类分析的需求。 首先, 为利用高光谱图像的空间和光谱信息, 定义基于像素空间距离和光谱角的指数衰变函数作为像素间相似性的度量。 然后, 在考虑像素相似性的基础上利用基于餐桌的构造形式为每个像素确定所在的餐桌。 最后, 对每张餐桌分配一道菜作为聚类类别, 从而达到聚类的目的。 利用航空可见-近红外成像光谱仪AVIRIS高光谱影像评估该模型性能, 实验结果表明: ssCRP模型可较好地实现高光谱图像的自动聚类, 与传统的K-means和ISODATA方法相比, 该模型结果斑块规整, “椒盐效应”得到抑制, 具有较高的空间一致性, 分类精度高, 其总体精度达到63.57%, Kappa系数为0.632 3, 能很好反映真实地物分布。 同时, 分类结果的地物间边界清晰, 能很好保持图像边缘。

The classification of hyperspectral images is one of the most important study fields. The spectral information is used in traditional classification of hyperspectral images, while the spatial correlativity information is ignored. To solve this problem, a novel model called spatial-spectral Chinese restaurant process (ssCRP) is proposed to cluster the hyperspectral images, which is an extension of Chinese restaurant process. Both the spatial and spectral information are considered in the modeling and inference of the method. The proposed model clusters the hyperspectral images better than tradional methods and satisfies the requirement of hyperspectral image clustering. Firstly, in order to consider both spatial and spectral information, a new similarity measurement is defined withthe exponential decay function based on the spatial distance and spectral angle among pixels. Then, each pixel is associated with a table based on the table construction by considering the similarity. Finally, each table is allocated with a dish which corresponds to a cluster. Thus, each pixel of the hyperspectral image is allocated with a clustering label. The true hyperspectral image collected by airborne visible infrared imaging spectrometer (AVIRIS) is used to evaluate the performance of our model. Experimental results indicate that the proposed model outperforms traditional K-means and ISODATA. Compared with those of the two methods, the result of the proposed model is more regular with lower salt-and-pepper effect with higher spatial consistency. The classification accuracy of the proposed model reaches to 63.57% and the Kappa coefficient is 0.632 3, much higher than those of K-means and ISODATA. Meanwhile, the edges of the result of our model are well preserved.