In July 2026 a fire started near La Mierla, a village of about forty people in the Sierra Norte de Guadalajara. Eighteen days later it had burned through some 35,000 hectares, the largest Spanish wildfire of the year. I spent a weekend rebuilding its story from open satellite data, and it taught me more about a field I knew nothing about than a month of reading would have.

The result is an interactive page and a repository with the full pipeline. This post is about what I actually learned building it.

Why thermal

A normal satellite image is reflected sunlight. No sun, no image. The thermal infrared band is different: it measures the radiation objects emit because of their own temperature, which means it works at midnight exactly as well as at noon.

There is a subtler advantage, and it is the one that hooked me. In the optical image, the burn area of La Mierla is a dark stain you could honestly mistake for terrain shadow. You need to know where to look. In the thermal band the same pixels sit at over 60 °C while everything around them is twenty degrees cooler. One image requires an expert. The other requires eyes.

The pipeline

Landsat 8 and 9 carry thermal sensors and their entire archive is free. The pipeline is four small Python scripts:

  1. Search and fetch. Scenes come from Microsoft Planetary Computer through STAC, the catalog standard the whole industry converged on. The files are cloud-optimized GeoTIFFs, so instead of downloading 500 MB scenes I read only my area of interest through HTTP range requests. Twelve scenes, 41 MB.
  2. Temperature. The raw numbers become degrees Celsius, clouds get masked out using the quality band, and everything lands on one common grid.
  3. Detection. Every pixel gets a z-score against its own June and July baseline. Separately, an index called dNBR compares vegetation before and after to map burn severity.
  4. Export. The results become small static PNGs that a Next.js page renders with MapLibre. No backend, nothing to fall over.

The hottest fire pixel came out at 17 standard deviations above its own baseline. Statistically speaking, that pixel is screaming.

The two bugs that looked perfectly fine

Here is the thing about remote sensing that I did not appreciate until I was in it: wrong results look normal. A temperature map that is off by two degrees looks exactly like a correct one. Nobody crashes. No exception is thrown.

So I made the pipeline validate physical invariants at every step, and that decision paid for itself twice in one afternoon.

First, the burned-area check failed: I was detecting 6,000 hectares against an official figure of 35,000. The code was right and my map was wrong. The fire had run sixty kilometres northeast, far beyond the box I had drawn around the ignition point. Without that assertion I would have published a beautiful, confident analysis of one sixth of the fire.

Second, after widening the box, two reservoirs showed up classified as high-severity burn. Water reflects infrared in ways that break the dNBR formula, a classic false positive I only knew to look for because the "burned lake" looked suspicious. One quality-band bit masks it out.

What the data refused to say

I went in expecting to show that thermal sees through smoke. It does, partially, in physics. But in all three mid-fire scenes I have, there is no dense plume over the front at the satellite's morning pass, so my data cannot demonstrate the claim, and the page does not make it.

What the data does show is arguably better: four days after the fire was declared controlled, the scar still radiates at around 60 °C. If you wanted to monitor a fire zone for reignition, the thermal band would keep telling you things long after the optical image went quiet.

What I would do next

More scenes and more fires, obviously. But the interesting direction is the one this project only gestures at: revisit time. Landsat passes every few days; a fire moves every hour. The gap between those two numbers is an entire industry, and it is why companies are launching dedicated thermal constellations right now. Building a toy version of the ground segment gave me enormous respect for the real ones.

The viewer is here, the code is here, and the fire, thankfully, is out.