Pixel classification of large-size hyperspectral images by affinity propagation

Abstract

Affinity propagation (AP) is now among the most used methods for unsupervised classification. However, it has two major drawbacks: (1) the number of classes (NCs) is over-estimated when the preference parameter value is initialized as the median value of the similarity matrix; and (2) the partitioning of large-size hyperspectral images is hampered by its quadratic computational complexity. To overcome these two drawbacks, we propose an approach which consists of reducing the number of pixels to be classified before the application of AP. To reduce the number of pixels, the hyperspectral image is divided into blocks, and the reduction step is then independently applied within each block. This step requires less memory storage, since the calculation of the full similarity matrix is no longer required. AP is applied on the new set of pixels, which is then set up from the representatives of each previously formed cluster and nonaggregated pixels. To correctly estimate the NCs, we introduced a bisection method which aims to assess intermediate classification results using a criterion based on pixel interclass variance. The application of this approach on hyperspectral images shows that our results are efficient and independent of the block size.