An unsupervised classification approach for hyperspectral images based adaptive spatial and spectral neighborhood selection and graph clustering

In remote sensing image processing, the classification is an interesting step to distinguish the image scene composition and can be of interesting role in different applications such as environmental monitoring and geological studies. Unlike the clustering, the classification needs labeled data for the training; however gaining these labeled data was always an expensive and hard task. For that, in this paper we propose an unsupervised classification approach that gains its labeled data from the proposed spatial and spectral graph clustering approach. The proposed adaptive spatial and spectral neighborhood selection approach is an extension of the k nearest neighborhood that assigns an adaptive number of neighbors to each pixel depending in its spatial and spectral relationship to its neighboring pixels. Then, this neighborhood will be clustered, to provide the first labeled training set, based on a hierarchical graph clustering algorithm. Finally, an SVM classification with a recursive kernel will be performed on the selected first labeled data at a first step and then the classification results are improved with the classification iterations of the recursive kernel. Experimental results on real hyperspectral images proved that with few iterations of the recursive kernel the proposed approach results are similar and even better then the supervised classification