A pansharpening scheme using spectral graph wavelet transforms and convolutional neural networks

The objective of the multispectral pansharpening scheme is to obtain high spatial-spectral resolution multispectral (MS) images using high spectral resolution MS and high spatial resolution panchromatic (Pan) images. Some distortions are found in the multiresolution analysis (MRA) based on pansharpening. It can be minimized by correct matching of the lowpass filter image. This paper illustrates the pansharpening approach that is based on multistage multichannel spectral graph wavelet transform and convolutional neural network (SGWT-PNN). In this scheme, the Pan image is decomposed by a multistage multichannel SGWT, and then a weighted combination of SGWT decomposition produces a lowpass filter component. Using the convolutional neural network model, this lowpass filter image is converted to a better-matched filter image which is completely fit to the MRA-based pansharpening methods. Simulation results in the context of qualitative and quantitative analysis demonstrates the effectiveness of the proposed scheme applied on datasets collected by different satellites.