Hyperspectral-Image-Super-Resolution-Benchmark: a reading list, not a codebase

What the field is, and the four classes the list is built on

The repository is described as a list of hyperspectral image super-resolution resources collected by Junjun Jiang, with an invitation to contact the author if important resources are missing, which is the maintainer signalling that the list is a living document rather than a finished one.

The framing paragraph at the top is the part worth reading, because it defines the field. Hyperspectral image super-resolution is a kind of technique that can generate a high spatial and high spectral resolution image from one of several kinds of observed data. There are four: a low-resolution multispectral image, for example an LR RGB image, a high-resolution multispectral image such as an HR RGB image or another 2D measurement, a low-resolution hyperspectral image, or a combination of high-resolution multispectral and low-resolution hyperspectral data.

From that, the list divides the field into four classes: spatiospectral super-resolution, spectral super-resolution, single hyperspectral image super-resolution, and multispectral and hyperspectral image fusion. The author adds a note that hyperspectral image reconstruction from 2D measurement is counted as a class of spectral super-resolution, which is the kind of editorial decision that makes a taxonomy usable rather than merely a diagram.

The consequence of organising by input type is that the taxonomy is the navigation. A reader who knows what data they have can find the section that matches, and that is a different and more useful organisation than a chronological paper list, because the input type is what you know when you start a project.

The distinction between the four classes, put plainly, is whether the method is adding spatial detail, spectral detail, or both, and whether it is working from one image or fusing two. Almost every design decision in the field follows from that question.

The survey section, and why the 1999 papers are still there

The first section of the list is headed Pioneer Work and Technique Review, and it is the section that tells you how the field thinks.

It opens with unmixing based multisensor multiresolution image fusion from TGRS 1999 by B. Zhukov et al., then the application of the stochastic mixing model to hyperspectral resolution enhancement from TGRS 2004, then Hardie's MAP estimation for hyperspectral image resolution enhancement using an auxiliary sensor from TIP 2004, and then Eismann's work on arbitrary response functions in TGRS 2005. The unifying idea across those is the stochastic mixing model and maximum a posteriori estimation, which is the classical way to pose the problem: you have a low resolution spectral measurement and a high resolution spatial measurement, you write down a forward model of how they relate, and you solve for the high resolution spectral image.

After the pioneers come the reviews. There is the 2015 hyperspectral pansharpening review by Loncan, the 2017 comparative review of hyperspectral and multispectral data fusion by Yokoya, a 2021 IGARSS survey by Zhang, a 2021 Information Fusion paper by Dian, a 2023 survey of spectral super-resolution meeting deep learning by Jiang He, a 2023 Remote Sensing review by Chi Chen, and a 2023 IEEE/CAA Journal of Automatica Sinica survey and perspective by Xinya Wang. Then the more recent surveys that show where the field has moved: diffusion models for hyperspectral image processing and analysis, diffusion models in low-level vision, Vision Mamba, transformers meeting hyperspectral imaging, and HyperSIGMA, a hyperspectral intelligence comprehension foundation model from TPAMI 2025.

Two of the older entries carry code links, which is useful. The 2015 pansharpening review links a PDF and a zip of results from the author's own site, and the 2017 fusion review links a PDF and the HSMSFusionToolbox from the open remote sensing site.

The reason the section starts in 1999 is that the list is organised by technique, and the techniques in current use are the ones the pioneers set up. A reader who starts with the 2023 surveys and works backwards will arrive at the same papers, but a reader who starts with the survey and wants to understand what a model is actually optimising has a path.

The last entry in the section is HyperBench, described as a benchmarking framework for fusion-based hyperspectral super-resolution methods, with links to a project repository and a PDF.

The method sections, and what the challenges contributed

The rest of the list is organised by class, and the entries carry PDF and code links wherever the authors made them available, which is uneven but a useful signal in itself.

The spatiospectral section opens with a 2020 TGRS paper on joint spatial and spectral super-resolution using a convolutional neural network by Mei et al. It then marks the maintainer's own work with a label, listing a multi-task interaction learning paper submitted to IEEE TIP and in peer review, and a deep unfolding network paper published in IEEE TCI 2022 with a link to the US3RN code repository. That deep unfolding connection is the interesting thread, because it is the classical MAP formulation of the previous section rendered as a network, which is why the pioneers and the recent work belong in the same list.

The spectral super-resolution section is dominated by the NTIRE challenges, the 2018 challenge on spectral reconstruction from RGB images, the 2020 challenge on spectral reconstruction from an RGB image, and the 2022 challenge on stereo image super-resolution, each with the challenge report as the citation. The pattern that emerges is that the challenges did for this field what they did for image restoration, which is to give it a common evaluation set and a common set of metrics, and the challenge winners then become the standard methods.

Several of those winners are listed with code. MST++, the multi-stage spectral-wise transformer, is noted as the winner of the NTIRE 2022 challenge on spectral reconstruction and links both a PDF and a repository. HDNet, a high-resolution dual-domain learning paper for spectral compressive imaging, links a PDF and code, as does the mask-guided spectral-wise transformer, and both are from the same group.

The later entries in the section move towards transformer architectures for remote sensing, a semisupervised degradation constrained network, a jointed spectral super-resolution applied to satellite imagery, and a dense residual channel re-calibration network for spectral super resolution.

The structure of the section is a history of the field told through its benchmarks. A reader who wants to know which method to start from will find that the challenge winners are the safest default, and a reader who wants the current state will find the transformer entries at the end.

What a paper list cannot do for you

It is worth being clear about what this repository is and is not, because the name suggests more machinery than it contains.

The top level of the repository contains exactly one file, the README. There is no dataset, no evaluation script, no leaderboard and no model. The HyperBench entry links out to a separate benchmarking framework project, and the method entries link out to the authors' own code repositories, which is where anything executable lives.

So the value is entirely in the curation, and there are three things the curation does that a search engine would not.

First, the classification. Every entry sits under a heading determined by the input data it consumes, and the author's note about counting 2D reconstruction as a class of spectral super-resolution is a small piece of editorial judgement that makes the sections mutually exclusive in a way that keyword search never would.

Second, the ordering within the sections, which in the spectral section runs from the challenges through the challenge winners to the transformer work, so a reader can see what the community converged on and when it converged.

Third, the links. The PDF and code links are scattered and inconsistent, so their presence is meaningful, and the code for the challenge winners and for the maintainer's own deep unfolding work is the fastest route to a working implementation.

The limitation is the obvious one. A reading list reflects one person's judgement about what matters, and the most recent entries are preprints and recent surveys, which by definition are the least settled part of the list. The list also stops where the sources it was built from stop, so anything published after the last push on 2026-09-29 is not here, and there are no GitHub releases, so there is no version history to consult.

What the list does not do is tell you which method to use. That question is answered by the benchmarks, and HyperBench is the pointer to where that work lives.

Editorial conclusion

This repository suits a graduate student or engineer entering hyperspectral image super-resolution who needs the taxonomy and the canonical papers in one place rather than reconstructing the field from citation chasing, since the four classes are defined by input type and the reading order follows that. It does not suit you if you want a benchmark you can run, because the repository contains a README and nothing else, no datasets, no evaluation code and no leaderboard. Before you rely on it, note that the list is the maintainer's own selection and the newest entries are still preprints, so check the peer-reviewed versions and the linked implementations yourself, and treat the HyperBench reference as the place to look for a comparable evaluation framework.

Frequently asked questions

Official sources

1. Issues
2. junjun-jiang/Hyperspectral-Image-Super-Resolution-Benchmark on GitHub
3. README