Controllable Reference Guided Diffusion with Local Global Fusion for Real World Remote Sensing Image Super Resolution

TL;DR

This paper introduces CRefDiff, a controllable diffusion-based method for remote sensing image super-resolution that effectively fuses local and global reference information, addressing real-world challenges and accelerating inference.

Contribution

The paper proposes CRefDiff, a novel controllable diffusion model with dual-branch fusion and a new dataset, advancing real-world remote sensing image super-resolution.

Findings

CRefDiff achieves state-of-the-art performance on RealRefRSSRD.

The dual-branch fusion mechanism improves adaptation to land cover changes.

The Better Start strategy accelerates inference significantly.

Abstract

Super resolution techniques can enhance the spatial resolution of remote sensing images, enabling more efficient large scale earth observation applications. While single image SR methods enhance low resolution images, they neglect valuable complementary information from auxiliary data. Reference based SR can be interpreted as an information fusion task, where historical high resolution reference images are combined with current LR observations. However, existing RefSR methods struggle with real world complexities, such as cross sensor resolution gap and significant land cover changes, often leading to under generation or over reliance on reference image. To address these challenges, we propose CRefDiff, a novel controllable reference guided diffusion model for real world remote sensing image SR. To address the under generation problem, CRefDiff leverages a powerful generative prior to produce accurate structures and textures. To mitigate over reliance on the reference, we introduce a dual branch fusion mechanism that adaptively fuse both local and global information from the reference image. Moreover, the dual branch design enables reference strength control during inference, enhancing the models interactivity and flexibility. Finally, the Better Start strategy is proposed to significantly reduce the number of denoising steps, thereby accelerating the inference process. To support further research, we introduce RealRefRSSRD, a new real world RefSR dataset for remote sensing images, consisting of HR NAIP and LR Sentinel2 image pairs with diverse land cover changes and significant temporal gaps. Extensive experiments on RealRefRSSRD show that CRefDiff achieves SOTA performance and improves downstream tasks.