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

Imagine trying to guess the temperature of a campfire just by looking at a single, blurry photo of it from a satellite high above. That's the challenge scientists face when studying wildfires from space. The problem is that a single "pixel" (the smallest dot on a map) often contains a chaotic mix of things: a tiny, super-hot flame, a smoldering pile of ash, and the cool, green grass surrounding it. It's like trying to guess the average temperature of a room by sticking a thermometer through a tiny hole, only to find the hole is sometimes over a hot stove and sometimes over an ice cube. Traditional methods often guess a single "average" temperature, but that can be misleading because fire isn't uniform; it's a wild mix of burning and smoldering. This matters because knowing the exact heat of a fire helps us predict how much smoke and pollution it's spewing into the air, how fast it might spread, and how dangerous it is for people and animals.

Now, enter a new team of scientists who have built a "digital detective" to solve this puzzle. Instead of guessing a single number, they created a system that looks at the light coming from a fire and figures out the entire range of temperatures hidden inside that tiny dot. Think of it like listening to a band play a song. A simple listener might just say, "It sounds loud." But this new detective can hear the individual instruments—the high-pitched violins (the hot flames) and the deep, rumbling bass (the smoldering embers)—and tell you exactly how loud each one is playing. They tested their detective on data from NASA's high-tech airplane camera, which can see light we can't even imagine, and found that it can accurately reconstruct the fire's heat profile, even when the fire is so hot it "blows out" the camera's sensors.

The Paper's Story: A Fire Detective with a Superpower

In this work, the authors present a new way to figure out how hot wildfires are using data from a special camera called AVIRIS-3, which flies on an airplane. This camera doesn't just take pictures; it acts like a super-sensitive prism, splitting light into hundreds of tiny colors (bands) from the visible spectrum all the way into the short-wave infrared. This is crucial because hot fires glow in these invisible colors, and the shape of that glow tells a story about the temperature.

The Old Way vs. The New Way
Previously, scientists tried to solve this by assuming a fire pixel was just a simple mix of two things: a single "hot" temperature and a single "cool" background. It was like assuming a smoothie was just a mix of strawberries and milk. But in reality, a fire is more like a chaotic smoothie with chunks of ice, hot fruit, and warm syrup all swirling together. The old methods also struggled when the fire was so hot that it saturated the camera (like a microphone blowing out when someone screams too loud), often forcing scientists to throw away that data.

The new method, however, treats the temperature inside a single pixel as a distribution. Instead of asking, "How hot is this pixel?", it asks, "What is the spread of temperatures inside this pixel?" They model this as a mix of two main groups: a "flaming" group (the hot, active fire) and a "smoldering" group (the cooler, dying embers). By using a clever mathematical trick called a "Gaussian mixture," they can estimate the average temperature and how much the temperature varies within that single dot.

How They Did It
The team built a "forward model," which is essentially a physics simulator. They programmed it to predict what the light should look like if a fire had a certain temperature and a certain background. Then, they used a super-fast computer brain (running on a GPU inside the airplane) to work backward. They compared the real light the camera saw with the light their simulator predicted, tweaking the temperature numbers until the two matched perfectly.

To make this work in real-time, they used a smart optimization tool (called Adam) that converges quickly. They also solved a tricky problem: atmospheric water vapor. Just like fog can blur your vision, water vapor in the air can distort the light from the fire. The team pre-calculated how much water vapor was in the air for every spot on the ground and held that fixed, so the computer could focus entirely on solving for the fire's temperature.

The Results: A Hot Success
The team tested their system in three main ways:

1. The "Fake Fire" Test: They took pictures of normal ground (no fire) and secretly injected a known, fake fire temperature into the data. They then ran their system to see if it could find the temperature they hid. The result? It was incredibly accurate. Across a range of simulated temperatures from 500 K to 1700 K, the system had an average error of only 34.1 Kelvin (K) and a root mean square error (RMSE) of 41.8 K. The correlation was extremely high at 0.987, meaning the predicted temperatures matched the fake ones almost perfectly.

2. The Real Fire Test: They applied the system to 168 flight lines from the 2025 FireSense campaign, covering over 4 million potential fire pixels. They checked how well their model fit the actual light data. In the most important infrared bands, the difference between their model and the real observation was tiny—less than 10% residual error. The average difference in brightness was less than 1 W m⁻² sr⁻¹ nm⁻¹, showing the physics model is very solid.

3. The "Zoom Out" Test: This was the most exciting part for the future. They wanted to see if this method would work for satellites, which have much "coarser" (blurrier) pixels than the airplane. They took their sharp, 5-meter resolution data and averaged it into 60-meter blocks to mimic a satellite view. They then compared the "satellite" result to the "average" of all the sharp pixels underneath it. The results showed that the method is robust: the difference between the coarse estimate and the fine-grained truth had a mean absolute error of only 27.16 K. This suggests the method can be used on future space missions like EMIT without losing its accuracy.

What They Found About the Fires
When they looked at the data from the entire campaign, they discovered something interesting about the nature of wildfires. Most of the pixels they analyzed were smoldering (cooler, under 900 K), making up about 95.3% of the fire pixels. Only about 4.7% were truly flaming (hotter than 900 K). This means that for every single pixel showing a raging flame, there are more than twenty pixels showing a smoldering fire.

They also tracked a prescribed burn at Fort Stewart over several hours. As the fire burned, they could watch the "flaming" tail of the temperature distribution shrink and the "smoldering" population grow, showing how the fire transitioned from active burning to a slow, smoky decay.

The Bottom Line
This paper doesn't just give us a new number; it gives us a new way of seeing. By moving from a single temperature guess to a full temperature distribution, the authors have created a tool that is robust, fast, and ready for space. They showed that even when a fire is so hot it saturates the camera, or when we look at it from a blurry satellite view, we can still accurately estimate the heat. This is a big step toward better fire monitoring, helping us understand not just where the fire is, but exactly how hot it is burning and how much energy it is releasing into our atmosphere.