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

arXiv:2608.07580 · cs.CV, astro-ph.IM, cs.LG · Submitted 2026-08-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Real-time physics inversion for retrieval of sub-pixel wildfire temperatures from VSWIR imaging spectroscopy".

Jane: The paper was written by William R. Keely, Philip G. Brodrick, Katherine Mistick, Adam Chlus, Robert O. Green et al. from Jet Propulsion Laboratory, California Institute of Technology and University of Utah.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Title: Tom: Welcome back to the show, everyone. Today we're digging into a paper that's got a mouthful of a title: "Real-time physics inversion for retrieval of sub-pixel wildfire temperatures from VSWIR imaging spectroscopy." Jane, I'm going to be honest, I had to read that title three times before I could even pronounce it.

Jane: Tom, I'm right there with you. But once you break it down, it's actually pretty straightforward. VSWIR just means visible to shortwave infrared light, and these researchers are using a sensor called AVIRIS-three that flies on an airplane and measures that light. The big idea is they can look at a wildfire from above and figure out exactly how hot different parts of that fire are.

Tom: And not just "hot" in a general sense, but they're getting down to individual pixels. We're talking about five-meter resolution, which means they can see the structure of the fire itself, the flaming parts versus the smoldering parts.

Jane: Exactly. And the "sub-pixel" part of the title is key here. A single pixel might contain both a really hot flame and a cooler smoldering area mixed together. This paper figures out how to separate those contributions mathematically.

Lu: What excites me is the "real-time" aspect. They're running this on the aircraft's GPU while the plane is still in the air. That's a massive shift from the old workflow where you'd collect data, land, process it for days, and then maybe get answers.

Tom: So Lu, you're saying this changes the game from "we'll tell you about the fire next week" to "we can tell you about the fire while we're still flying over it"?

Lu: Exactly. And that matters for firefighting operations. If you can see where the hottest, most active parts of the fire are in real time, you can direct resources there immediately.

Jane: The authors tested this on the two thousand twenty-five FireSense campaign, which covered fires from California to Georgia. Over four million pixels of fire data, and they got temperature estimates with really impressive accuracy.

Meng: I'm curious about the computational side though. Running a physics-based inversion on a GPU while the plane is moving, that's not trivial. What kind of optimization did they use?

Tom: Great question, Meng. They used something called Adam optimization, which is a standard machine learning technique, but they applied it to a physics model rather than a neural network. Two stages, actually, first fitting the surface background, then fitting the fire temperature distribution.

Jane: And the results are solid. They verified their method by injecting known temperatures into real background spectra, and they recovered those temperatures with a mean error of about thirty-four Kelvin. That's really good for this kind of remote sensing work.

Tom: So we've got real-time, physics-based, sub-pixel temperature retrieval for wildfires. I think we're just scratching the surface here. Let's keep going and talk about what this actually means for how we understand fires.

Summary: Jane: So we've established what the title means, but let's talk about what this paper actually accomplishes. The summary in the abstract lays out three major results, and I want to unpack those with everyone.

Tom: Go for it, Jane. I think the first big result is that they verified their forward model on simulated data. They took real background spectra, injected a known fire temperature, and then ran their retrieval to see if they could recover it.

Lu: And they got an RMSE of forty-one point eight Kelvin across a range from five hundred to one thousand seven hundred Kelvin. That's a huge temperature range, from smoldering embers all the way up to intense flaming combustion.

Meng: The RMSE being under fifty Kelvin across that entire range tells me the physics model is well-posed. If there were major errors in the radiative transfer assumptions, you'd see those errors blow up at the extremes.

Jane: The second result is the campaign-scale application. They ran this on one hundred sixty-eight overflights from the two thousand twenty-five FireSense campaign, covering fires in California, Alabama, Texas, Georgia, and Florida. That's about four million fire pixels.

Tom: And the residual fit, meaning how well their model matches the actual observed radiance, was within ten percent across the shortwave infrared bands. That's a strong validation that the model is capturing the real physics.

Lu: The third result is probably the most forward-looking. They tested whether this method can work from space by coarsening the airborne data to simulate what a satellite would see. The absolute error between fine and coarse resolution retrievals was about twenty-seven Kelvin.

Meng: That's the EMIT connection, right? The paper mentions EMIT, which is a spaceborne imaging spectrometer on the International Space Station. If this method works at coarser resolution, it could be applied to EMIT data.

Jane: Exactly, Meng. And that's huge because EMIT covers the whole Earth. You could get wildfire temperatures globally, not just where an airplane happens to fly.

Tom: So we've got validation on simulations, validation on real campaign data, and a path to spaceborne deployment. That's a pretty complete package for one paper.

Lu: I'd add that the fact they're using a full physics forward model, rather than a statistical or empirical approach, means the method should generalize better to conditions they haven't seen yet.

Jane: And that's important because fires are getting more extreme and more unpredictable. We need tools that can handle novel conditions, not just ones that work on historical data.

Tom: Alright, so we've covered the summary. But I want to dig into the methodology more, because the way they handle the temperature distribution inside each pixel is really clever.

Improvements: Tom: So Jane, we've talked about what the paper does, but let's talk about what it improves on. Previous methods for fire temperature retrieval had some real limitations.

Jane: Right, and the paper is pretty explicit about this. Older methods assumed a single temperature for the fire inside each pixel. You'd get one number, like "this pixel is eight hundred Kelvin," and that was it.

Lu: But real fires aren't like that. At millimeter scales, you have flames at one thousand two hundred Kelvin right next to smoldering fuel at six hundred Kelvin. A single temperature is a gross oversimplification.

Meng: So what did they do differently? They used a Gaussian mixture model, right?

Jane: Exactly. They model the temperature distribution inside each pixel as a mixture of two components, a hot flaming component and a cooler smoldering component. Each has its own mean and standard deviation.

Tom: And then they collapse that into a single moment-matched Gaussian, giving you an effective fire temperature and a measure of how spread out the temperatures are within that pixel.

Lu: That's a real improvement. You're not just getting a point estimate anymore. You're getting a distribution, which means you get uncertainty information for free.

Meng: The other improvement is how they handle atmospheric effects. Previous methods simplified the atmosphere, but this paper uses a full radiative transfer model with per-pixel water vapor estimation.

Jane: That's important because water vapor absorbs specific wavelengths of light. If you don't account for it correctly, you'll get errors in your temperature retrieval. They pre-solve water vapor and hold it fixed during the temperature inversion.

Tom: And saturation handling. When a fire is really hot, the sensor's detectors can saturate, meaning they max out and can't measure anymore. The paper handles this by masking saturated bands on a per-pixel basis.

Lu: Which is smart because saturation isn't uniform. One pixel might saturate in the shortwave infrared while another doesn't. By masking adaptively, they retain as much information as possible.

Meng: So the improvements are: temperature distributions instead of point estimates, proper atmospheric correction, and adaptive saturation handling. That's a solid list.

Jane: And the result is better accuracy, as we saw with the thirty-four Kelvin mean absolute error on simulated data. But more importantly, it's more physically realistic.

Tom: I also like that they're being honest about limitations. They acknowledge that the Gaussian assumption might not perfectly represent the true temperature distribution, but it's a calibrated summary of what the spectrum constrains.

Lu: That's the right attitude. You model what you can, you're honest about what you can't, and you provide uncertainty so users know how much to trust the numbers.

Jane: Alright, so we've covered the improvements. But I want to go back to the actual first page of the paper and look at the introduction more carefully, because there's some context there that really matters.

First Page: Jane: So Tom, we've been talking about the methodology and the results, but let's go back to the first page of the paper, "Real-time physics inversion for retrieval of sub-pixel wildfire temperatures from VSWIR imaging spectroscopy." The introduction lays out why this work matters.

Tom: And the key point there is that fire temperature is not just an academic curiosity. It governs what gets emitted into the atmosphere, how deep the soil gets heated, and how the ecosystem recovers afterward.

Lu: Exactly. The combustion temperature determines the trace gases and particulates that are released. If you get the temperature wrong, you get the emissions inventory wrong.

Meng: And that feeds into climate models, air quality forecasts, and public health warnings. It's not just about the fire itself, it's about everything downstream.

Jane: The paper also traces the history of fire temperature retrieval back to Dozier in one thousand nine hundred eighty-one who used a bispectral method with AVHRR data. That was mid-infrared and thermal infrared, and it had limitations.

Tom: Those limitations being sensitivity to background temperature estimates and the assumption of a single fire temperature across hundreds of meters.

Lu: The shift to shortwave infrared was a game-changer. SWIR bands have strong emitted radiance from hot fires, negligible background thermal radiance, and weak solar reflected radiance. That makes the thermal signal much cleaner.

Meng: And imaging spectrometers like AVIRIS-three give you hundreds of contiguous narrow bands across the VSWIR range. That's way more information than the few broad bands of AVHRR or MODIS.

Jane: The paper also mentions that previous spectral mixture models used one or two temperature endmembers. But in situ measurements show that combustion temperatures vary at millimeter scales.

Tom: So even the two-endmember models were oversimplifying. The Gaussian mixture approach in this paper is a step toward representing that real complexity.

Lu: And the atmospheric treatment. The paper criticizes previous approaches for oversimplifying radiative transfer. By using a full 6S-based model with per-pixel water vapor, they remove residual atmospheric structure that could bias temperature retrieval.

Meng: I appreciate that they're building on prior work rather than dismissing it. They cite Dennison and Roberts, Matheson and Dennison, Waigl, Amici. They're clearly standing on shoulders here.

Jane: And they're addressing a specific problem Matheson and Dennison identified in two thousand twelve which is that fire temperature retrievals are sensitive to spatial resolution. As you aggregate fine pixels to coarser footprints, temperatures get biased cooler.

Tom: Which is exactly what they test in this paper with the fine-to-coarse aggregation experiment. And they show their method is more robust to that effect.

Lu: That's the key for spaceborne deployment. If the method only works at five-meter resolution, it's useless for EMIT at sixty meters. But they showed it holds up.

Jane: So the first page sets up the problem beautifully, and the rest of the paper delivers on solving it. I think we've got a complete picture here.

Conclusion: Tom: Alright, let's wrap this up. We've spent the show talking about "Real-time physics inversion for retrieval of sub-pixel wildfire temperatures from VSWIR imaging spectroscopy," and I think we've covered a lot of ground.

Jane: We have. Let me try to pull it together. This paper presents a method that uses a full physics forward model to retrieve wildfire temperatures from imaging spectroscopy data, and it does it in real time on the aircraft's GPU.

Lu: The key innovations are the Gaussian mixture representation of sub-pixel temperatures, the proper atmospheric correction with per-pixel water vapor, and the adaptive saturation handling. Together, those give you better accuracy and uncertainty information.

Meng: And the validation is solid. They tested on simulated data with known temperatures, they applied it to the entire two thousand twenty-five FireSense campaign with over four million fire pixels, and they showed it works at coarser resolutions suitable for spaceborne sensors.

Tom: The campaign results were striking too. Over ninety-five percent of the retrieved pixels were in the smoldering regime, below nine hundred Kelvin, with only about four point seven percent in the flaming regime. That really highlights how much of a fire's lifetime is spent smoldering.

Jane: And the Fort Stewart time series showed the method can track a fire over time, watching the flaming population rise and fall as the fire develops. That temporal dimension is something we haven't had before.

Lu: The implications are broad. This could improve emissions inventories, fire behavior modeling, and operational firefighting. And the path to spaceborne deployment through EMIT and future missions like EAGLE-VSWIR means global coverage is within reach.

Meng: From an engineering standpoint, the fact that it runs in real time on the aircraft is the most impressive part. That's not a research prototype, that's an operational tool.

Tom: So we've got a paper that combines solid physics, clever optimization, rigorous validation, and practical deployment. That's a rare combination.

Jane: And it's going to change how we study and respond to wildfires. We're saying goodbye to this paper, but I have a feeling we'll be seeing its influence for years to come.

Tom: Thanks for joining us, everyone. Next time, we'll be looking at another exciting paper from the arXiv. Until then, stay curious.

Jane: And stay safe out there. Goodnight, everyone.

William R. Keely, Philip G. Brodrick, Katherine Mistick, Adam Chlus, Robert O. Green, Philip E. Dennison

Jet Propulsion Laboratory, California Institute of Technology · University of Utah

cs.CV, astro-ph.IM, cs.LG

Submitted: 2026-08-04

Comments: In Review in Remote Sensing of Environment

Code: https://github.com/isofit/isofire

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 69/100

Terminology

Summary

Summary

This paper presents a real-time, physics-based inversion framework for retrieving sub-pixel wildfire temperatures from VSWIR (visible-to-shortwave infrared) imaging spectroscopy data, specifically applied to NASA's Airborne Visible Infrared Imaging Spectrometer (AVIRIS-3). The retrieval uses a full-physics approach in which a forward model is employed to resolve both solar and emitted radiance derived from a temperature distribution and utilizes the full spectral range in the residual fit. The optimization employs state-of-the-art nonlinear least squares methods implemented for fast convergence on the on-board GPU, allowing for estimation of effective fire temperature within flight cadence.

The forward model represents observed radiance as the sum of a solar-reflected term and a fire-emitted thermal term. The solar term uses a 6S-based radiative transfer parameterization with atmospheric terms evaluated from a per-scene 6S look-up table generated for a grid of column water-vapor values. The surface reflectance is modeled as a mixture of five fire-relevant endmembers (green vegetation, non-photosynthetic vegetation, two soils, and a combined char/ash class) constrained to the unit simplex. The emitted radiance term integrates Planck radiance over a temperature distribution, parameterized as a two-component Gaussian mixture with a flaming and a smoldering component, with each component fit over its own SWIR window (hot component: [900, 1800] nm; smoldering component: [1500, 2500] nm). The mixture is collapsed to a single moment-matched Gaussian per pixel, reporting an effective fire temperature E[T] and sub-pixel temperature spread σ[T].

Column water vapor is pre-solved per pixel using the 940 nm absorption feature, smoothed over a 12×12 super-pixel neighborhood, and held fixed during temperature inversion. The inversion uses a two-stage Adam optimizer: Stage 1 fits surface mixture coefficients, a single scalar temperature, and initial thermal scaling over the full [350, 2500] nm window for 30 steps at learning rate 0.03; Stage 2 fixes the surface mixture and fits the two-component Gaussian temperature mixture with thermal scaling over the SWIR windows for 100 steps at learning rate 0.015. All state variables are mapped to normalized coordinates to equalize learning rates, and saturated bands are masked per-pixel.

Fire pixels are identified using the Hyperspectral Fire Detection Index (HFDI), defined as HFDI = (L(2430 nm) − L(2061 nm)) / (L(2430 nm) + L(2061 nm)), with a uniform threshold of HFDI ≥ −0.1. The paper notes that this threshold's minimum detectable temperature depends on solar zenith angle and gray-body emissivity scaling, as characterized in an appendix.

The retrieval was applied to all 168 unique AVIRIS-3 flight passes acquired during the 2025 FireSense campaign, spanning the Eaton Fire (Los Angeles, 11 January 2025), Geneva State Forest prescribed burn (Alabama, 27 March 2025), Targets of Collaboration wildfires and prescribed fires (Alabama, Florida, Mississippi, March 2025), Crabapple Fire (Texas, 18 March 2025), and Fort Stewart prescribed burns (Georgia, April 2025), totaling 4 million plausible-fire pixels with an average spatial resolution of 5 m.

Verification on simulated spectra with injected single-temperature thermal emissions (500–1700 K range) yielded RMSE = 41.8 K, MAE = 34.1 K, and R2 = 0.987 for retrieved effective fire temperature versus injected temperature. Forward model residuals across 250,000 randomly sampled campaign pixels at three diagnostic wavelengths (1097 nm, 1596 nm, 2249 nm) showed R2 = 0.98 at each wavelength, with mean residuals of +0.35, −0.67, and −0.55 W m−2 sr−1 nm−1 respectively.

Spatial resolution robustness was tested by aggregating fine 5 m AVIRIS-3 radiance to coarse 60 m grids (12×12 block-mean) for fifty scenes. The coarse-pixel moment-matched Gaussian posterior was compared to the empirical distribution of underlying fine-resolution E[T] values. The mean difference between fine-resolution average E[T] within a footprint and coarse-resolution E[T] exhibited an RMSE of 27.16 K. The Quantile Error Curve showed the coarse posterior is 50–75 K colder than the empirical fine distribution at the 5th–10th quantiles and 25–40 K hotter at the 90th–95th quantiles, with departures interpreted as the posterior correctly absorbing uncertainty about unobservable per-pixel temperatures.

Campaign-scale temperature distribution across the 4 million pixels showed that "smoldering effective temperatures dominate the campaign: 4,110,810 pixels (95.3%) fall in the smoldering regime (E[T] ≤ 900 K) against 201,043 (4.7%) in the flaming regime (E[T] > 900 K)." The paper adopts E[T] = 900 K as the split between smoldering-dominated and flaming-dominated pixels, referencing reported smoldering upper bound (850 K) and flaming onset (950 K). A time series from Fort Stewart on 14 April 2025 with 7-minute cadence over twelve scenes showed the active-fire area expanding and contracting, with flaming temperatures most pronounced in mid-sequence scenes of highest pixel count.

The paper discusses limitations including the Gaussian temperature distribution assumption (noting SWIR radiance constrains mean and width but not shape), the absorption of gray-body emissivity into the thermal scaling term α with potential α-temperature degeneracy, and the assumption that surface reflectance can be expressed as a sparse mixture of fire-relevant endmembers. Future work could extend the retrieval to orbital imaging spectrometers like EMIT, collocate with VIIRS active-fire detections and ECOSTRESS land surface temperature data, and use repeat overflights to link retrieved temperatures to pre-fire fuel maps and weather conditions.

Improvements for AI systems

Based on the paper, I can identify several concrete improvements to AI systems for wildfire monitoring and remote sensing:

Improvement: Build an AI system that inverts a full radiative transfer forward model (6S-based) using two-stage Adam optimization on GPU, rather than relying on spectral mixture models or empirical band ratios.

Capabilities:

  • Retrieve per-pixel effective fire temperature E[T] with RMSE of 41.8 K against known injected temperatures

  • Process 4 million pixels per campaign within flight cadence (real-time onboard processing)

  • Handle saturated bands adaptively by masking them per-pixel without collapsing the fit

  • Account for atmospheric water vapor, surface reflectance mixtures, and gray-body emissivity simultaneously

Abstract

In this work, we present a wildfire temperature retrieval framework for VSWIR imaging spectroscopy data, employed on data from NASA's Airborne Visible Infrared Imaging Spectrometer (AVIRIS-3). The retrieval framework utilizes a full-physics approach in which a forward model is employed to resolve both solar and emitted radiance derived from a temperature distribution and utilizes the full spectral range in the residual fit. To optimize the forward model retrieval, we use state-of-the-art nonlinear least squares methods implemented for fast convergence on the on-board GPU, allowing for estimation of effective fire temperature within flight cadence. We verify the forward model assumptions on simulated spectra with an injected thermal signature and find good agreement with an RMSE of 41.8 Kelvin (K). We apply the retrieval over the full 2025 FireSense AVIRIS-3 campaign, totaling 168 overflights with probable active fire spectra, and demonstrate a residual radiance fit of at most 10% across bands in the short-wave infrared (SWIR). Lastly, we verify the applicability of the retrieved posterior fire temperature parameters to generalize to space-borne imaging spectrometers such as EMIT, by retrieving at coarsened spatial resolution. We find that the posterior distribution exhibits good coverage of the underlying sub-pixel temperature range with an absolute error of 30 K across quantiles and a mean absolute error of 27.16 K between spatial resolutions.

Sources

Related papers