Pansharpening by Combining Enhanced Spatial-Spectral Fidelity and Gradient-Domain Guided Filtering

Abstract

Pansharpening techniques fuse the complementary information from a high-resolution panchromatic (PAN) and low-resolution multispectral (LRMS) images to produce a high-resolution multispectral (HRMS) image. However, most of the existing pansharpening methods have been affected in the full-resolution domain due to both the absence of ground truths and unavoidable unknown noises. To address this problem, a new pansharpening method has been proposed that combines enhanced sparse models and gradient-domain guided image filtering. Specifically, a deep multiscale Laplacian pyramid super-resolution network improves the resolution of the original LRMS image instead of bicubic interpolation. Then, the accurate preservation of spatial-spectral characteristics is achieved in a variational framework with enhanced spatial-spectral fidelity in the image gradient domain. Meanwhile, the gradient-domain guided image filter is used to effectively improve the extraction accuracy of spatial characteristics from the PAN image. Finally, the enhanced sparse regularization on the latent HRMS image is designed to remove noise and artifacts while promoting piecewise-smooth solutions. The experimental results on public satellite datasets demonstrate the superiority of the proposed method against existing pansharpening methods in terms of both full-resolution performance indexes and visual quality.

Keywords

• Enhanced sparse model
• gradient-domain guided filtering
• multiscale super-resolution network
• pansharpening