ResSR: A Computationally Efficient Residual Approach to Super-Resolving Multispectral Images

TL;DR

The paper tackles the high computational cost of multispectral image super-resolution (MSI-SR) by introducing ResSR, a model-based framework that decouples spectral and spatial processing into parallel branches and combines them with a lightweight residual correction. The spectral branch uses a low-rank subspace representation via SVD, while the spatial branch relies on bicubic upsampling; a residual correction merges these to recover accurate spectral and spatial features. A key contribution is a closed-form, pixel-linear solution with complexity $O(N_p K^3)$, enabled by a spatially-decoupled forward model and a small spectral subspace dimension $K$ (typical $K=2$). Experimental results on simulated and real Sentinel-2 data show that ResSR matches or exceeds the quality of existing MSI-SR methods while being 2×–10× faster and capable of handling much larger images. This makes high-fidelity MSI-SR practical for large-scale or time-critical remote sensing applications, addressing both accuracy and efficiency barriers in current approaches.

Abstract

Multispectral imaging (MSI) plays a critical role in material classification, environmental monitoring, and remote sensing. However, MSI sensors typically have wavelength-dependent resolution, which limits downstream analysis. MSI super-resolution (MSI-SR) methods address this limitation by reconstructing all bands at a common high spatial resolution. Existing methods can achieve high reconstruction quality but often rely on spatially-coupled optimization or large learning-based models, leading to significant computational cost and limiting their use in large-scale or time-critical settings. In this paper, we introduce ResSR, a computationally efficient, model-based MSI-SR method that achieves high-quality reconstruction without supervised training or spatially-coupled optimization. Notably, ResSR decouples spectral and spatial processing into separate branches, which are then combined in a residual correction step. The spectral branch uses singular value decomposition plus a spatially-decoupled approximate forward model to upsample the MSI, while the spatial branch uses bicubic upsampling. The residual correction step combines these branches to recover accurate spectral and spatial MSI features. ResSR achieves comparable or improved reconstruction quality relative to existing MSI-SR methods while being 2$\times$ to 10$\times$ faster. Code is available at https://github.com/hdsullivan/ResSR.

Figures (5)

• Figure 1: Comparison of the proposed ResSR pipeline with the standard pipeline for deep learning-based and iterative SVD model-based MSI-SR methods. ResSR decouples spatial and spectral processing, enabling a non-iterative, pixel-linear solution that significantly reduces computation while preserving accurate spatial detail.
• Figure 2: Overview of the ResSR pipeline. The spectral branch of ResSR uses singular value decomposition plus a spatially-decoupled approximate forward model to upsample the MSI, while the spatial branch uses bicubic upsampling. The residual correction process combines these two branches, extracting accurate high-frequency details from the spectral branch and low-frequency details from the spatial branch. By decoupling the spatial and spectral processing, ResSR achieves computational efficiency throughout the entire method.
• Figure 3: Comparison of $2\times$ and $6\times$ super-resolved bands to 2m GSD APEX ground truth and 4m / 12m GSD original resolutions, shown using a false-color composite. In panels (d)-(h), the left half shows the $2\times$ super-resolution case (4m $\rightarrow$ 2m) using a false-color composite of bands B7, B11, B12, while the right half shows the $6\times$ super-resolution case (12m $\rightarrow$ 2m) using bands B1, B9, B9. Since LRTA is not defined for multiple lower spatial resolutions, we exclude this method from our 6$\times$ super-resolution comparison. Bicubic interpolation produces blurred images, LRTA exhibits blocking artifacts, DSen2 introduces pixel-intensity distortions, and SupReME introduces high-frequency artifacts (shown in red). In contrast, ResSR produces sharp spatial detail with minimal artifacts, though it appears slightly sharper than the ground truth. Additionally, ResSR is roughly $2 \times$ faster than DSen2 and $100 \times$ faster than SupReME.
• Figure 4: Comparison of $2\times$ and $6\times$ super-resolved bands to 20m GSD Sentinel-2 ground truth and 40m / 120m GSD original resolutions, shown using a false-color composite. LRTA exhibits blocky artifacts, DSen2 loses high-frequency structure and introduces pixel-intensity distortions, and SupReME produces grid-like artifacts (shown in red). In contrast, ResSR matches the ground-truth pixel intensities and preserves fine spatial detail.
• Figure 5: Comparison of $2\times$ and $6\times$ super-resolved bands to 20m / 60m GSD Sentinel-2 original resolutions, shown using a false-color composite. LRTA introduces blocking artifacts, while DSen2 and SupReME lose high-frequency detail or introduce grid-like distortions (shown in red). In contrast, ResSR maintains accurate pixel intensities and restores spatial detail without visible artifacts.