Bayesian Data Fusion for Adaptable Image Pansharpening

Currently, most optical Earth observation satellites carry both a panchromatic sensor and a set of lower spatialresolution multispectral sensors. In order to benefit from both sources of information, several pansharpening methods have been developed to produce a multispectral image at the spatial resolution of the panchromatic band. The aim of this paper is to suggest a novel approach to the pansharpening problem within a Bayesian framework. This Bayesian data fusion (BDF) method relies on statistical relationships between the various spectral bands and the panchromatic band without suffering from restricting modeling hypotheses. Furthermore, it allows the user to weight the spectral and panchromatic information with respect to either visual or quantitative criteria, which leads to adaptable results according to users' needs and study areas. The performance of this approach was compared to existing methods based on markedly different subset images from very high spatial resolution IKONOS images. Results showed that BDF yielded the highest spectral consistency. Furthermore, small details were adequately added to the pansharpened images with little artifact as compared to those created using wavelet-based methods. Finally, the method was fast and easy to implement owing to its straightforward formulation. As it does not have any intrinsic limitations on the type of data to be processed or the number of bands to be merged, it also appears to be very promising for optical/SAR or hyperspectral image fusion.