Unsupervised hyperspectral image segmentation: Merging spectral and spatial information in boundary adjustment

Abstract

Hyperspectral Image (HSI) segmentation is one of the key prepossessing steps for subsequent applications and analysis. Unsupervised Segmentation without dimension reduction is a pervasive problem in Computer vision and a more challenging problem in HSI. A novel unsupervised spectral-spatial adaptive boundary adjustment/movement-based model and a framework is devised for HSI segmentation, which is initialized via weighted clustering. This model uses two key features of HSI data: spectral correlation and band spatial/contextual preference. Band preference is defined by the discriminative ability and informative amount contained in each band, and an adaptive mechanism is proposed to retain the spectral correlations in the spectral dimension and to match the actual structure in spatial dimensions. The proposed framework allows the image to be explored at different segmentation levels. Use of local structural regularity and self-similarity information from channel-group adaptive boundary adjustment based process, and final segmentation result based on the proposed four merge criteria's has a significant effect on the accuracy of the final segmentation. Experiments on real diverse hyperspectral data sets with different contexts and resolutions demonstrate the accuracy of the proposed framework over several well-known existing approaches and even proves its accuracy over the ground truth.