Super-Resolution of Sentinel-2 Images Using a Geometry-Guided Back-Projection Network with Self-Attention

TL;DR

This paper introduces a novel super-resolution method for Sentinel-2 multispectral images that combines geometry-guided learning with self-attention mechanisms to enhance spatial and spectral detail fusion.

Contribution

It proposes a geometry-guided super-resolution model with a cluster-based learning procedure and a self-attention mechanism within an unfolded back-projection framework.

Findings

Outperforms classical super-resolution methods

Achieves superior fusion of spectral and spatial information

Effective across diverse landscape types

Abstract

The Sentinel-2 mission provides multispectral imagery with 13 bands at resolutions of 10m, 20m, and 60m. In particular, the 10m bands offer fine structural detail, while the 20m bands capture richer spectral information. In this paper, we propose a geometry-guided super-resolution model for fusing the 10m and 20m bands. Our approach introduces a cluster-based learning procedure to generate a geometry-rich guiding image from the 10m bands. This image is integrated into an unfolded back-projection architecture that leverages image self-similarities through a multi-head attention mechanism, which models nonlocal patch-based interactions across spatial and spectral dimensions. We also generate a dataset for evaluation, comprising three testing sets that include urban, rural, and coastal landscapes. Experimental results demonstrate that our method outperforms both classical and deep learning-based super-resolution and fusion techniques.