Multi-band image fusion using Gaussian process regression with sparse rational quadratic kernel

This paper proposes an approach for multi-band image fusion using a multiple output variable Gaussian Process (GP) model. The considered model uses a new covariance function, which is a product of an intrinsically sparse kernel and a Rational Quadratic Kernel (RQK) to model the pixel coordinates and intensity of the high spatial resolution image. The new kernel serves as a stochastic prior for each band of the estimated image. The developed approach allows the exchange of information between the different modalities enabling local structure of the high spatial resolution image on which the model is trained. The accuracy performance and image quality assessment show that the proposed approach achieves compelling enhancement when compared with other fusion methods.