About the data

Firefinder shows weekly wildfire risk along power-grid corridors for Portugal, Spain and France. The country is divided into hexagons of roughly 5 km², and every week each hexagon gets a probability that wildfire reaches that hexagon, computed from satellite imagery, weather, terrain and land cover.

That risk is then rolled up onto every power line corridor in the national grid and ranked, because vegetation near power lines is a leading cause of both wildfires and large-scale outages. Knowing which corridors to worry about this week is more useful than knowing which regions are generally fire-prone.

Everything is built from free public data, and the whole project is open source.

The forecasts come from gradient-boosted decision trees (XGBoost) trained on nearly a decade of fire history. For each hexagon and week, the model weighs 21 factors: how dry and stressed the vegetation looks from space, the week's fire weather, drought over the last 30 and 90 days, how steep the terrain is, what kinds of fuel cover the ground, distance to power lines, and the time of the fire season.

Some honest numbers. Wildfire touching any given cell in any given week is rare: between roughly 1 in 700 and 1 in 12,000 depending on the country. The model ranks risk well (ROC-AUC of 0.85 for Portugal, 0.86 for Spain and 0.90 for France, measured on held-out years it never saw during training), but raw model scores are not probabilities. We calibrate them against observed fire frequency, so the percentages you see are true weekly likelihoods (even severe cells rarely exceed a few percent), and the tiers (Severe, High, and so on) compare each cell against the rest of the country that same week.

The "why" bars you see when you click a corridor or cell are SHAP attributions: the model's own accounting of which factor pushed the score up or down for that place and week.

Limitations, stated plainly: the weather used is observed, not forecast. Small fires are missing from the training labels. And corridor risk means fire risk near the line, not a claim that the line causes fires; no public European dataset attributes wildfire occurrences to grid equipment, so that distinction is beyond what open data can support.

Live NASA FIRMS detections never change the forecast scores. Instead, cells containing a detection are shown as active fires on the map, and corridors within 1.5 km of a detection appear under the Active fires tab with their distance to the nearest one. The weekly forecast shown for those places was computed before the detection, so treat it as context, not a reaction to the fire.