Spectral Clustering Ensemble and Unsupervised Clustering for Land cover Identification in High Spatial Resolution Satellite Images

Unsupervised clustering plays a dominant role in detailed landcover identification specifically in agricultural and environmental monitoring of high spatial resolution remote sensing images. A method called Approximate Spectral Clustering enables spectral partitioning for big datasets to extract clusters with different characteristic without a parametric model. Various information types are used through advanced similarity criteria. Selection of similarity criterion optimal for the corresponding application is required. To solve this issue a Spectral Clustering Method is proposed which fuses partitioning obtained by distinct similarity representations. This Spectral Clustering Ensemble adopts neural Quantization in the place of Random Sampling, and advanced similarity criterion in the place of Gaussian kernel distance with distinct decaying parameters, and a two level ensemble. The built up areas in the high resolution images can be detected using unsupervised detection. In this process first, a large set of corners from each of the input images are extracted by an improved Harris corner detector. Then, the extracted corners are incorporated into a likelihood function to discover candidate regions in each input image. Given a set of candidate build-up regions, in the second stage, the problem of build-up area detection is concised as an unsupervised grouping problem. The performance of these algorithms is evaluated by Accuracy, Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI). Experimental results show a significant betterment of the resulting partitioning obtained by the proposed ensemble, with respect to the evaluation measures in the applications.