谱-空图嵌入的高光谱图像多核分类算法

Spectral-spatial Graph Embedding Based Hyperspectral Image Multi-kernel Classification

作为一种非常有效的预处理步骤,降维算法被广泛地应用于高光谱图像分类中.为了联合利用高光谱图像的光谱维和空间维信息,本文提出了一种基于谱-空图嵌入降维的多核融合分类算法,自适应融合降维后的空谱特征进行分类.该算法主要由三个步骤组成:首先,将训练集中的每个像素点作为顶点,每个顶点用对应像素的光谱特征描述,以此构造一个光谱图,利用图嵌入模型求得一个低维投影矩阵;其次,利用主成分分析模型提取高光谱图像的第一个主成分,并将其划分成不同大小的超像素块,以每个超像素块为顶点,每个顶点用超像素块中所有像素点的平均值来描述,从而构造一个空间图,再次使用图嵌入模型求得一个低维投影矩阵;最后,对于高光谱图像中的每个像素点,可用两个不同的投影矩阵分别求得其对应的低维特征表示,利用多核学习的方法对两者进行有效的融合,自适应学习融合权重,提升了后续SVM分类的准确性.在两个公开的高光谱图像数据库上进行了测试,验证了本文算法的有效性.

As a very useful process of preprocessing,dimensionality reduction（DR）algorithms have been widely used for hyperspectral image classification.To fully explore the spectral and spatial information of hyperspectral images,we propose a spectral-spatial graph embedding DR for multi-kernel fusion classification in this paper,which adaptively fuses the dimensionality reduced spatial and spectral feature for classification.Our algorithm mainly consists of three main steps.First of all,each pixel in the training set is considered as a vertex to construct a spectral graph.In the spectral graph,each vertex is described by the spectral features of its corresponding pixel.A graph embedding model is then used on the spectral graph to derive a low-dimensional projection matrix.Secondly,principal component analysis is used to extract the first principal component from the original hyperspectral image.The first component is then divided into superpixels with different sizes.We use each superpixel as a vertex,which is represented as the average values of pixels in the superpixel,to construct a spatial graph.Similarly,another low-dimensional projection matrix can be obtained using the graph embedding model.Finally,for each pixel in the hyperspectral image,we use these two projection matrix separately to get their corresponding low-dimensional representations,fused by a multi-kernel learning method.Fusion weight is adaptive learned,which improves the classification accuracy of SVM.Experiments on two public hyperspectral images can validate the effectiveness of our proposed algorithm.