Authors: Ritik Shah (ritik.shah@aritheon.com), Marco Duarte (mduarte@ecs.umass.edu)

HyperBench is a benchmarking framework for fusion-based hyperspectral super-resolution (HSR) methods.

It is designed for the setting in which a high-resolution hyperspectral image (HR-HSI) is reconstructed by fusing:

• a low-resolution hyperspectral image (LR-HSI), and
• a high-resolution multispectral image (HR-MSI)

HyperBench provides a standardized and reproducible pipeline for generating synthetic degradations, evaluating reconstruction quality, and logging structured results.

HyperBench is currently released as a flexible benchmarking framework rather than a fixed leaderboard benchmark. It is intended to help researchers evaluate HSI-MSI fusion methods under controlled, reproducible degradation settings while allowing custom datasets, models, PSFs, SRFs, metrics, and experiment configurations. Official benchmark tracks and leaderboard protocols are under development.

HyperBench is intended to be a community-extensible framework. Its usefulness will grow rapidly as researchers contribute realistic PSFs, SRFs, degradation models, sensor configurations, metrics, and baseline integrations. We welcome and appreciate open-source contributions that help make hyperspectral super-resolution and HSI-MSI fusion evaluation more reproducible, diverse, and representative of real imaging conditions.

Hyperspectral super-resolution methods are often evaluated under inconsistent experimental setups, making direct comparison difficult. HyperBench addresses this by providing:

• a unified degradation pipeline
• a consistent evaluation protocol
• a model-agnostic integration interface
• structured experiment outputs

The framework is intended strictly for evaluation, not training.

HyperBench implements the full evaluation workflow:

1. Load a hyperspectral scene from a .mat file
2. Generate synthetic inputs:
   • LR-HSI via spatial degradation
   • HR-MSI via spectral degradation
3. Pass degraded inputs into a model pipeline
4. Normalize and validate model outputs
5. Optionally apply output clipping
6. Compute evaluation metrics
7. Record results to a structured CSV file

• Built-in PSFs for spatial degradation
• Built-in SRFs for MSI simulation
• Configurable downsampling ratios
• Configurable MSI band counts
• Configurable noise levels

• RMSE
• PSNR
• SSIM
• UIQI
• ERGAS
• SAM

Models must expose:

or optionally if you profile your model's FLOPs, number of parameters, GPU memory usage, etc. put the model statistics with their label into a dictionary called statistics for HyperBench to include these in the CSV files produced in addition to the quality metrics which will be included in this output CSV by default:

Supported backends include NumPy, TensorFlow, and PyTorch. This run_pipline function must be wrapped in a class:

This ExamplePipeline class can then be given to HyperBench:

This setup allows HyperBench to directly give the LR HSI, HR MSI, SRF, and PSF as inputs to your model. If your model does not need any of these inputs (for example some models do not require the SRF or the PSF as inputs), you can simply define your ExamplePipeline class to take them as None:

If your method requires scene specific or degradation specific hyperparameter tuning, you must take care of this within your run_pipeline function itself. HyperBench does not support any specific hyperparameter settings, since it expects the user to directly supply a model contract through the run_pipeline function defined by the user, allowing you to explicitly define the exact hyperparameters that your model needs when running HyperBench on a specific dataset and with specific degradations.

To install and seamlessly run hyperbench, create a fresh conda environment:

Once created and activated, proceed to install hyperbench:

This package installs all the hyperbench dependencies, however, any additional libraries that your models require such as tensorflow, pytorch, tqdm, etc. will have to be installed seperately.

This repository does not host any datasets. Popular hyperspectral datasets in the .mat format can be easily downloaded online. Users can also visit this link: https://github.com/ritikgshah/SpectraLift/tree/main/Datasets where the Washington DC Mall, Kennedy Space Center, Botswana, Pavia University, and Pavia Center datasets in .mat format are hosted.

We provide in depth documentation, allowing a user to seamlessly start generating experiments with hyperbench. It is recommended that as a new user, you first go through the documentation in detail. Documentation index is the first file recommended as it lays out the structure and sequence that the documentation files should be read in.

The notebooks/ directory provides a structured, tutorial-based walkthrough of HyperBench.

These notebooks are designed to guide users from basic usage to full benchmark orchestration in a clear and practical way. Each notebook focuses on a specific part of the framework, while collectively demonstrating how all components work together in a complete hyperspectral fusion benchmarking pipeline.

The sequence in which you should explore these notebooks is clearly explained in the notebooks summary. It is recommended to read through this file before exploring individual notebooks to have a seamless experience in understanding the tutorial.

The cli/ directory provides a clean, script-based pathway to run HyperBench end-to-end—without notebooks. It mirrors the full workflow demonstrated in Notebook 07 (End-to-End Benchmark Pipeline) and is intended for repeatable, production-style experiments. Command line interface help provides further details about how to structure python scripts, how to use yaml and json configuration templates to run your own experiments, and how to integrate a simple pipeline (the same one used in the notebooks tutorial) to get started with the HyperBench framework.

The examples/ directory demonstrates how real-world hyperspectral super-resolution models can be seamlessly integrated into the HyperBench framework.

Unlike the tutorial notebooks, which focus on understanding the framework using simplified pipelines, this section shows how HyperBench operates in practical research settings with full model implementations. Examples readme provides exact details about which methods are included, how to go through the examples directory, and enables users to seamless understand how to structure their models for easy integration with HyperBench.

HyperBench provides:

• synthetic benchmarking
• degradation generation
• evaluation metrics
• CSV logging
• model integration

HyperBench does not provide:

• model training
• dataset hosting
• pretrained models

MIT License