基于谱域-空域结合特征和图割原理的高光谱图像分类

Hyperspectral image classification based on spectral-spatial combination features and graph cut

高光谱图像中存在着特征维度高而训练集小的问题。为解决该问题,提出了一种2步走的分类方法:1)通过支持向量机对图像进行初步分类,根据分类结果计算出每个类别的均值特征;2)使用1)计算出来的均值特征作为能量函数的数据项,然后利用图割原理对图像做二次分类。实验中发现:空间上相近的像素点往往具有相似的特征,且属于同一个类别。针对这种现象,提取一个将谱域特征和空域特征相结合的新特征。该特征既包含了光谱信息也包含了空间信息,具有较好的分类性能和鲁棒性。在Indian Pine数据集和Pavia University数据集进行实验,实验结果表明了本文提出方法的有效性。

The high-dimension of the feature vs. small-size of training set is an unsolved problem in the hyperspectral image classification task. To solve this problem a two-step classification method is proposed. Firstly,a preliminary classification is performed by the support vector machine（ SVM） and the classification results are used to calculate the mean feature（ MF） of each class. Secondly,a classification based on the graph cut theory is applied with the MFs as an input of the energy function. The experimental results showed that spatially nearby pixels have large possibilities of having the same label and similar features. Therefore,a new feature called spectral-spatial combination（ SSC） is extracted that combines the spectral-based feature and spatial-based feature. The SSC feature contains the related spectral and spatial information of each pixel and provides better classification performance and robustness. Experiment results on the Indian Pine dataset and the Pavia University dataset demonstrated the effectiveness of the proposed method.