ISOFire 🔥🛰️

This is the official implementation of

Real-time physics inversion for retrieval of sub-pixel wildfire temperatures from VSWIR imaging spectroscopy

by William Keely et al. from NASA Jet Propulsion Laboratory, California Institute of Technology

Introduction

ISOFire retrieves sub-pixel wildfire temperature from VSWIR imaging spectroscopy, applied to NASA's AVIRIS-3. A full-physics forward model resolves both solar and emitted radiance from a temperature distribution and fits the full spectral range.

The retrieval fits a two-component (hot + smoldering) Gaussian temperature mixture per pixel and reports the analytic moment-matched single Gaussian: mu = effective fire temperature E[T] in Kelvin, and sigma = its uncertainty. The solver is a batched two-stage Adam optimizer written in JAX, fast enough on a GPU to run within flight cadence.

Two entry points share the same retrieval:

	isofire.onboard	isofire.airborne
Use	flight operations	post-processing and analysis
Inputs	RDN, LOC, BANDMASK, OBS	RDN, LOC, GLT, BANDMASK, OBS
Output	unwarped ENVI pairs for a downstream ortho step	orthorectified GeoTIFF (COG)

Module	Role
isofire.rt	Physics: 6S templating and parsing, spectral resampling, emissive radiance.
isofire.retrieval	Algorithms: water vapor, endmembers, masks, the mixture solver, moment matching.
isofire.airborne	Scene watcher, ortho, GeoTIFF output. Shared I/O helpers.
isofire.onboard	Scene watcher for flight operations, pre-ortho ENVI output.

Installation

We use uv.

git clone https://github.com/isofit/isofire.git
cd isofire
uv sync                      # CPU JAX
uv sync --extra cuda12       # GPU JAX (recommended)

Build 6S

The retrieval calls a 6S Fortran binary as a subprocess. 6s is vendored under sixs/, so a fresh clone has everything it needs. Requires gfortran.

cd sixs/Sixs
make FFLAGS="-O -ffixed-line-length-132 -std=legacy -fno-pie -no-pie"
cd ../..
export ISOFIRE_SIXS_DIR=$PWD/sixs/Sixs

The -fno-pie -no-pie flags are required on modern Linux toolchains; without them the link step fails with relocation R_X86_64_32 against '.data' can not be used when making a PIE object. The resulting sixs/Sixs/sixsV2.1 is platform-specific and gitignored, so each clone builds its own.

Data files

The fire endmember library, Kurudz solar irradiance, and AVIRIS-3 wavelength/FWHM cache ship inside the package at src/isofire/data/. Resolution order is the CLI flag, then $ISOFIRE_DATA_DIR, then the packaged copy.

AVIRIS-3 L1B scenes are publicly available from the AVIRIS-3 data portal. A scene is processed once all of its ENVI pairs are present in the input directory, grouped by scene id (AV3<date><time>[_<chunk>]).

Usage

Onboard (flight operations; pre-ortho ENVI products):

uv run python -m isofire.onboard \
    --input_dir     /path/to/l1b \
    --to_ortho_dir  results/to_ortho

Airborne (orthorectified GeoTIFFs, used for the results in the paper):

uv run python -m isofire.airborne \
    --input_dir          /path/to/l1b \
    --fire_temp_dir      results/fireTemp \
    --fire_temp_unc_dir  results/fireTempUnc

Both are watchers: they poll the input directory and process each scene as its files land. run_isofire.sh is a SLURM example with the tuned defaults from the paper. See --help for the full flag list.

Outputs

Product	Units	Contents
<scene_id>_fireTemp	K	effective fire temperature, E[T]
<scene_id>_fireTempUnc	K	1-sigma uncertainty

In-scene pixels that were not retrieved carry -1; out-of-swath pixels in the orthorectified products carry 0. Retrieved values are strictly positive, so both sentinels are unambiguous.

Citation

If you use this code, please cite us:

@article{keely2026realtime,
  title={Real-time physics inversion for retrieval of sub-pixel wildfire temperatures from VSWIR imaging spectroscopy},
  author={Keely, William R.},
  journal={arXiv preprint arXiv:2608.07580},
  year={2026}
}

Acknowledgements

This research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration. A portion of this work was funded by NASA ROSES 80NM0018F0590 and by Realtime Active Fire Data from NASA's AVIRIS-3 Sensor for Fire Operations 80NSSC25K0225. AVIRIS-3 is sponsored by the NASA Earth Science Division.

We thank the 6S team for their work; see the official 6S site for documentation and the canonical source.

Licensed under Apache-2.0. The vendored 6S radiative transfer code in sixs/ is distributed under its own terms.