SAE-LWIR: A MODTRAN-Generated Dataset for Atmospheric Compensation in Standoff LWIR Hyperspectral Imaging

Dataset of the paper "Set-Based Transformer for Atmospheric Compensation in Standoff LWIR Hyperspectral Imaging" (IGARSS 2026, Oral).

Fabian Perez, Nicolas Quintero, Jeferson Acevedo, Hoover Rueda-Chacón — Universidad Industrial de Santander, Bucaramanga, Colombia

Project Page · Code · Paper

Description

To the best of our knowledge, this is the first publicly available dataset tailored to atmospheric compensation (AC) in standoff long-wave infrared (LWIR) hyperspectral imaging. It was generated with MODTRAN5 from the clear-sky atmospheric profile database (CSP, derived from ECMWF ERA5), filtered to 36,547 clear-sky profiles (cloud coverage ≤ 10%, relative humidity ≤ 90%, subset of ocean-surface profiles removed).

At-sensor radiance follows the standoff radiative transfer equation, simulated at:

• 7 standoff ranges: R = {30, 90, 150, 210, 270, 330, 390} m
• 7 target temperatures: T = {280, 285, 290, 295, 300, 305, 310} K (gray-body emissivity ε = 0.95)
• 256 spectral bands over the 8–13 µm LWIR window (Gaussian ISRF, FWHM = 40 nm)

yielding 255,829 samples (36,547 profiles × 7 temperatures). Radiance is reported in microflicks (µW·sr⁻¹·cm⁻²·µm⁻¹). Downwelling radiance was computed at a fixed 45° viewing angle; transmittance and path radiance (upwelling) were computed along the line of sight for each range using gas concentrations specified independently for the first 126 atmospheric layers.

Files

File	Shape	Content
forward.npy	(36547, 7, 7, 256)	At-sensor radiance L — [profiles, temperatures, ranges, bands]
transmittance.npy	(36547, 7, 256)	Transmittance τ — [profiles, ranges, bands]
upwelling.npy	(36547, 7, 256)	Atmospheric path radiance Lₐ — [profiles, ranges, bands]
downwelling.npy	(36547, 256)	Shared downwelling spectrum L_d — [profiles, bands]
processed_data/	—	Raw atmospheric state variables per profile: temperature (t.npy), water vapor (h2o.npy), ozone (o3.npy), pressure (p.npy), skin temperature (skt.npy), surface pressure (sp.npy), total column water vapor (tcwv.npy), cloud coverage (tcc.npy), latitude/longitude, classification and clear-sky indices

Splits used in the paper: 70% train / 10% validation / 20% test (random).

Usage

hf download SemilleroCV/SAE-LWIR --repo-type dataset --local-dir data

import numpy as np

forward = np.load("data/forward.npy", mmap_mode="r")        # (36547, 7, 7, 256)
transmittance = np.load("data/transmittance.npy", mmap_mode="r")
upwelling = np.load("data/upwelling.npy", mmap_mode="r")
downwelling = np.load("data/downwelling.npy", mmap_mode="r")

Training/evaluation code: https://github.com/Factral/SAE-LWIR

Citation

If you find this dataset useful, please cite our paper (arXiv:2606.08324):

@inproceedings{perez2026setbased,
  title={Set-Based Transformer for Atmospheric Compensation in Standoff LWIR Hyperspectral Imaging},
  author={Perez, Fabian and Quintero, Nicolas and Acevedo, Jeferson and Rueda-Chac{\'o}n, Hoover},
  booktitle={IGARSS 2026 - IEEE International Geoscience and Remote Sensing Symposium},
  year={2026}
}

Acknowledgments

This work was supported by the Air Force Office of Scientific Research (AFOSR) through the Southern Office of Aerospace Research and Development (SOARD) under grant number FA8655-25-1-8010.