加权空-谱自适应近邻聚类的高光谱图像分类

Hyperspectral image classification using weighted spatial and spectral clustering with adaptive neighbors

高光谱图像聚类算法可以对海量的高光谱图像数据进行信息提取,完成地物类别的初步分类。自适应近邻聚类(clustering with adaptive neighbors,CAN)作为一种新型的聚类算法,利用样本间的局部连通性实现聚类,聚类效果较好,但是该算法的性能受样本间相关性的影响较大。基于此,文章提出了一种新的融合高光谱图像的空间信息和光谱信息的分类方法,即加权空-谱自适应近邻聚类(weighted spatial and spectral clustering with adaptive neighbors,WSS-CAN)法,该方法通过引入样本点的近邻窗口尺度和光谱因子2个参数对高光谱图像进行重构,增强了样本间的相关性,对重构后的图像进行CAN聚类,有效提高了分类精度。在Indian Pines和Salinas-A数据库上的实验结果表明,由WSS-CAN得到的总体精度分别为56.33%、77.90%,分别比其他聚类算法提升了11.52%~18.47%、10.1%~14.79%,聚类效果较好。

The algorithm of hyperspectral images（HSI） clustering can extract the valuable infor-mation, which can be used to classify the ground truth？ from vast HSI data. Clustering with adaptive neighbors （CAN）,a new clustering algorithm, performs well by using the local connectivity effective-ly, but the result is susceptible to the correlation of samples. Inclassification method,weighted spatial and spectral clustering with adaptive neighbors（WSS-CAN） is proposed. This algorithm combines both spatial window and spectral factor to rewhich strengthens the correlation of pixels. CAN is used to cluster the reconstructed data, thus in-creasing the classification accuracy. The benchmark tests on Indian Pines and Salinas-A demonstrate that the performance ofWSS-CAN is better than that of other algorithms. The beaccuracy（OA） obtained by WSS-CAN on the HSI datasets is 56. 33 % and 77. 90 % receeding that by other algorithms 11. 52%-18. 47% and 10. 1%-14. 79%.