A bit of context.

The 2025 wildfire season concluded with unprecedented burnt-area extents across the northwestern Iberian Peninsula.

The EFFIS Burnt Area product estimated the burnt perimeters at approximately 1230 km², which triggered several misinterpretations in regional government communications and local media coverage.

First, an incorrect processing of the EFFIS dataset by a local newspaper led to an overestimation of this area by 40.5%, and no organization was able to provide any methodological correction neither regarding this (incorrect) figure nor regarding the limitations of EFFIS itself.

The EFFIS Burnt Area product does not provide burnt-area values, but rather fire perimeters. These are derived from the assessment of thermal anomaly detections by the MODIS instruments onboard the Terra and Aqua (EOS) satellites, and by VIIRS onboard the NOAA-20, NOAA-21, and Suomi-NPP platforms. They are subsequently refined through the use of multispectral Sentinel-2 imagery and various national databases.

Its purpose is to provide burnt-area estimates at the European scale, in a stable and reproducible manner across a wide range of biomes and geographic regions, rather than high-resolution cartography, not being able to detect small patches, low severity fires, and absolutely not equivalent to forest loss.

Scroll to transition from EFFIS burn scars to sen2fire burn scars.

Deep segmentation of burn scars.

What is shown below is a neural network—specifically a U-Net—designed and trained to identify the spectral signature of vegetation that has undergone combustion or intense radiation damage in recent wildfires at 10 m resolution. It uses a single multispectral Sentinel-2 L2A image as its sole input and outputs a three-class map:

• Vegetation that has sustained substantial damage.
• Non-combustible surfaces (such as water bodies, bare soil, and some artificial surfaces).
• Healthy vegetation.

Scroll to transition from Image A to Image B.

The development of a high-resolution cartography could hence arguably be important from an ecological standpoint, while also enabling continuous mapping, severity and impact, post-fire assessment, identification of true recurrence and extremely granular parameter extraction to feed a new generation of propagation models.

Scroll to transition from Amber-tinted to Color render.

Scroll to transition from Amber-tinted to Color render.

Scroll to transition from Amber-tinted to Color render.

Scroll to transition from Amber-tinted to Color render.

Despite relying on a training set of good yet imperfect quality, built through a supervised classification, restricted in scope and belonging to a single season, the results exhibit consistency and generalization across multiple seasons and diverse regions. Its confusion matrix is presented, showing an accuracy for the burnt-surface class of approximately 97%. It is also noteworthy, perhaps somewhat counterintuitively, that the classification displays accuracy higher than that and higher than the training set itself, a fact that is evident through visual inspection, but not yet formally validated as it depends on the annotation of another independent dataset.

Scroll to transition from Image A to Image B.

Scroll to transition from Image A to Image B.

Lessons from a toy model.

A forest-fire model is a class of automata (see Wikipedia's entry) that resembles in some limited aspects the propagation of a wildfire. It consists on a grid whose cells can acquire four different states:

• Combustible.
• Non-combustible.
• Burning.
• Burnt.

The amount of fuel is controlled by the parameter p, which represents the probability of a single cell to be randomly assigned a combustible state during the initial conditions setup. When the simulation starts, each cell will change from combustible to burning if there is an adjacent cell on fire.

Different probabilities lead to a variety of structural patterns:

When plotting several observable quantities as a function of the occupation probability p, an interesting behavior emerges.

These quantities converge into a phase transition in the thermodynamic limit, located approximately at p = 0.59. Below this value, the ‘fire’ self-extinguishes. Above this value, it propagates through wide fronts. Any intermediate situation collapses into a very narrow band of probabilities.

The analogy with a real wildfire, assuming its propagation is dominated by fuel availability (although in reality it is far more complex) is evident.

It should be noted that the model is not reversible in any other than the trivial case. Information is lost and multiple paths lead to the same state.

Instead of using a simple occupancy probability, random pseudo-flammability fields (or the probability of ignition provided an adjacent cell is already on fire) were generated. Another deep learning architecture was trained on the generated dataset to infer the latent ignition field that generated and is compatible with the observed burn scar pattern:

Scroll to transition from Generated to Inferred.

In fact, by combining this prior with the real-world data, the model was able to predict the ignition probability latent field from satellite imagery acquired before the wildfires themselves.

Scroll to transition from Image A to Image B.

This latent field can be compared compared against the actual burn scars of real events that took place later in the season:

... or directly ignited at the approximate ignition point of a real fire:

Keep in mind this system contains a single parameter (ignition probability). It lacks orography, human activity, meteorology, atmospheric coupling, fuel structure, temporal dynamics, energy flow, things that, on the other hand could be generated in a more complex simulation or implemented as a physical layer within our deep architecture but, as for now, it only contains cells switching on and off and yet it accurately models fire spreading in critical and supercritical regimes.

A real wildfire is just one of many realizations that take place in a simulated universe.