Residual Dense Network for Pan-Sharpening Satellite Data

Abstract

Pan-sharpening is a multi-sensor fusion task that aims to enhance the spatial resolution of spectral data using panchromatic data of the same scene. This work proposes a deep Residual Dense Model (RDM) for Pan-Sharpening (PS) of satellite data which learns hierarchical features that can efficiently represent the local complex structures from panchromatic data. This work addresses the two general problems emphasized in pan-sharpening application viz., spectral and spatial preservation. The proposed Residual Dense Model for Pan-Sharpening network (RDMPSnet), preserves the spectral information by spectral mapping of Low-Resolution Multi-Spectral data (LRMS) while the spatial preservation is achieved by learning the hierarchical structural features from High-Resolution Panchromatic data (HRP). To extract this structural feature RDMPSnet is trained end to end with Low Resolution (LR) panchromatic patches and High Resolution (HR) residue patches to learn a non-linear mapping. The trained non-linear mapping network is capable to generate structural feature for any LRMS data which is injected into the mapped spectral data. The network is experimentally evaluated with Worldview2 and IKONOS2 satellite data and shows that the proposed RDMPS achieves favorable performance both visually and quantitatively against state-of-the-art methods.