An improved Bayesian nonparametric mixture model to fusing both panchromatic and multispectral images for classification

Abstract

In this paper, we present an improved nonparametric Bayesian model based on a generalized metaphor of Chinese restaurant franchise (gCRF), which can take advantage of both panchromatic and multispectral images to obtain a classification map. There are two drawbacks in the gCRF when it is used to fuse panchromatic and multispectral image for classification, first, since superpixels which are obtained using other segmentation algorithm are considered as basic analysis units instead of pixels in the gCRF, the quality of final classification result depends on the calibre of over-segmentation map. Second, when classify PAN and MS image using the gCRF, semantic segments extracted from PAN image are sharing with MS image and then they are allocated clustering labels using MS image which is richer in spectral information. All the local semantic segments extracted from panchromatic image are supposed to be suitable for representing the local spatial structures in multispectral image, which is not objective in practice. In this paper we propose an improved gCRF, focusing on overcoming the two shortcoming of the gCRF. First of all, the formation of superpixels are integrated into the nonparametric Bayesian framework of the improved gCRF to obtain qualified superpixels. Second, the quality of the semantic segments is checked before sharing with MS image and corresponding measure is taken to dealing with the situation that the semantic segments are not suitable to represent the local structure of MS image. We evaluate the efficiency of the improved model and show it obtains the state-of-art results.