HyperFM Dataset

Overview

This dataset contains **4,250 hyperspectral samples** from NASA PACE mission, curated for machine learning and computer vision research. It is designed to support tasks such as regression, segmentation, and representation learning using hyperspectral imaging (HSI) data.

A larger version of this dataset (**3TB+**) can be generated using the provided codebase (link below).

Directory Structure

.
├── hsi/
├── target/
├── train_list_2k.csv
├── valid_list_250.csv
├── test_list_2k.csv

Description

– **hsi/**  
  Contains hyperspectral data samples. Each file corresponds to a single observation. Dim: (96,96,291)

– **target/**  
  Contains labels or target values associated with each HSI sample. Four (4) Regression Tasks: *cld_xxx.npy [Dim: (96,96,4)]. One (1) Segmentation Task: *cldmask_xxx.npy [Dim: (96,96)]

– **train_list_2k.csv**  
  Training split with ~2,000 samples.

– **valid_list_250.csv**  
  Validation split with ~250 samples.

– **test_list_2k.csv**  
  Test split with ~2,000 samples.

Each CSV file defines the mapping between input samples and their corresponding targets.

Dataset Statistics

– Total samples: **4,250**
– Training set: ~2,000 samples  
– Validation set: ~250 samples  
– Test set: ~2,000 samples  

Full Dataset Generation

A significantly larger dataset (**3TB+**) can be generated using the official codebase:

– GitHub repository: *https://github.com/umbc-sanjaylab/HyperFM*

The repository includes scripts for:
– Data preprocessing
– Hyperspectral cube generation
– Label construction
– Dataset scaling

Associated Publication

– Paper: Zahid Hassan Tushar, and Sanjay Purushotham, “HyperFM: A Efficient Hyperspectral Foundation Model with Spectral Grouping”, CVPR 2026 (findings).

Please cite the paper if you use this dataset in your research.

Usage Notes

– Ensure consistent preprocessing across splits when training models.
– Refer to the CSV files for correct pairing of inputs and targets.
– Large-scale experiments are recommended using the full generated dataset.

License

CC BY 4.0

Contact

For questions, issues, or collaboration:

– **Zahid Hassan Tushar**  
  Email: ztushar1@umbc.edu  

– **Sanjay Purushotham**  
  Email: psanjay@umbc.edu  

Acknowledgments

If you use this dataset, please acknowledge the authors and associated publication.