基于多尺度空谱鉴别特征的高光谱图像分类

Hyperspectral Image Classification Based on Multi-scale Discriminative Spatial-spectral Features

为了应对高光谱图像同质区域面积分布不均的问题,同时更充分地挖掘空间和光谱信息之间的内在联系,提出了一种基于多尺度空谱鉴别特征的高光谱图像分类方法。该算法首先对图像进行不同尺度的滤波操作,接着分别从得到的多幅图像中提取鉴别的空谱特征,并使用支持向量机(SVM)进行分类。最后,该算法采取"决策级融合"的策略,来综合不同滤波尺度图像的分类结果。在Indian Pines,Kennedy Space Center和University of Pavia数据集上的实验表明,该算法能够提取较为有效的空间信息,当随机选取10%的像素作为训练样本时,该算法的总体分类准确率均能达到96%以上,其分类精度和Kappa系数均优于其他分类算法。

In order to cope with the unevenness of homogenous regions’area in hyperspectral images,an algorithm based on multi-scale discriminative spatial-spectral features was proposed.First,the image is processed with multi-scale filters.Then discriminative spatial-spectral information is extracted from the filtered images before put into SVM classifiers.At last,classification results of the filtered image are combined with decision fusion strategy.The experimental results on Indian Pines,Kennedy Space Center and University of Pavia indicate the effectiveness of the extracted spatial information.The overall accuracy of this algorithm can reach up to 96%when 10 percent of samples are randomly selected for training.What’s more,the classification accuracy and Kappa coefficient are higher than the comparative algorithms.