Dimensionality Reduction and Classification of Hyperspectral Images by Spatial-Spectral Similarity Measure Using Distance Correlation Coefficient

In the analysis of hyperspectral (HS) images, dimensionality reduction and classification are two important issues. In this paper, we propose a new dimensionality reduction and classification algorithm for HS images. Based on the basic fact that pixels in a small range of HS images are highly similar, meaningful features are extracted from spectral and spatial perspectives, and then distance correlation coefficient is used to measure the similarity of pixels in the space. This method can measure the degree of correlation between vectors of different dimensions and improve the existing K-nearest neighbor (KNN) classification method. At the same time, we also introduce the new similarity measure to ISOMAP dimensionality reduction algorithm, so that it can effectively deal with the relationship between spectral information and spatial information. We have carried out experiments on the classification performance of five real HS image data sets under different measures. The experimental results show that the new algorithm greatly improves the accuracy of HS image classification and has a good application prospect.