UNICORNproject/high-resolution_fire_danger

UNICORN Project — High-resolution Fire Danger Model Outputs Dataset

Partner: MITIGA Project: UNICORN

Overview

This dataset contains high-resolution fire danger outputs produced by two complementary modelling systems developed by MITIGA:

1. A short-term forecast model, which produces daily fire danger predictions for the next 1–5 days.
2. A long-term stochastic model, which produces a long-run annual wildfire-risk baseline.

Both systems combine machine-learning-based ignition probability estimation with a physics-based fire-spread model, but they differ in time horizon, and the weather/ignition data driving them. Their outputs are organized into two separate folders, described below.

Note: the outputs included in this dataset are illustrative samples intended to showcase the system's capabilities for the NW Iberian Peninsula, and do not represent final production outputs.

---

Folder structure

.
├── short-term_daily_forecasts/
│   ├── YYYY-MM-DD/
│   │   ├── burn_probability.tif
│   │   └── flame_length_max.tif
│   ├── YYYY-MM-DD/
│   │   ├── burn_probability.tif
│   │   └── flame_length_max.tif
│   └── ... (one sub-directory per forecast day, example days included)
│
└── long-term_annual_probabilities/
    ├── annual_probability_bootstrap_p05.tif
    ├── annual_probability_bootstrap_p50.tif
    ├── annual_probability_bootstrap_p95.tif
    ├── annual_probability_bootstrap_std.tif
    └── flame_length_max.tif

---

1. short-term_daily_forecasts/

Daily, high-resolution fire danger forecasts for the next 1–5 days. Each sub-directory is named after the forecast date (YYYY-MM-DD) and contains the outputs for that single day. A few example days are included in this dataset.

Methodology summary

The short-term forecast system has two computational modules:

• Ignition probability estimation: a ML model trained on historical wildfire ignition records from public agencies, using anthropogenic predictors, landscape predictors (orography, vegetation, and fuel categories), and meteorological predictors (Fire Weather Index and its subcomponents). This produces a daily, high-resolution ignition probability map.
• Fire spread and behaviour simulation: stochastic ignition points are sampled from the ignition probability map, and each is propagated using a physics-based fire-spread model, driven by meteorological forecasts retrieved from ECMWF and high-resolution fuel, terrain, and vegetation data. Running an ensemble of thousands of stochastic simulations per day yields burn probability and fire behaviour outputs.

Files (per day)

File	Description	Units
burn_probability.tif	Probability that each pixel burns on that forecast day, derived from the ensemble of stochastic fire simulations	fraction (0–1)
flame_length_max.tif	Maximum simulated flame length at each pixel, a proxy for fire intensity	feet

---

2. long-term_annual_probabilities/

A long-run wildfire-risk baseline: the probability that each pixel burns in a typical year, estimated from a large ensemble of synthetic fire years rather than a forecast tied to specific dates.

Methodology summary

The long-term model:

• Analyzes real historical fires in the region of interest to characterize ignition frequency, fire size, duration, and the fire-weather conditions typically associated with ignitions.
• Uses the ignition probability ML model to generate a spatial ignition-probability map, so that simulated fires are concentrated where ignitions are realistically plausible rather than distributed randomly.
• Builds a large ensemble of synthetic fire-weather years by resampling real historical weather (ERA5) on days whose fire-danger index (FWI) statistically resembles past ignition conditions.
• Physically simulates each synthetic fire with a fire-spread model over real terrain and vegetation data.
• Aggregates results across the full ensemble, converting per-pixel burn counts into an annual burn probability.

Uncertainty in the annual burn probability is quantified via a bootstrap resampling technique, yielding a median estimate along with lower/upper bounds and a standard deviation.

Files

File	Description	Units
annual_probability_bootstrap_p50.tif	Median (50th percentile) annual burn probability across bootstrap samples	fraction (0–1)
annual_probability_bootstrap_p05.tif	Lower bound (5th percentile) of the bootstrapped annual burn probability	fraction (0–1)
annual_probability_bootstrap_p95.tif	Upper bound (95th percentile) of the bootstrapped annual burn probability	fraction (0–1)
annual_probability_bootstrap_std.tif	Standard deviation of the bootstrapped annual burn probability	fraction (0–1)
flame_length_max.tif	Maximum simulated flame length at each pixel across the synthetic fire ensemble, a proxy for fire intensity	feet

---

General notes

• All outputs are raster files in GeoTIFF (.tif) format, at high spatial resolution.
• Both systems rely on the same physics-based fire-spread engine run over real terrain, fuel, and vegetation data, but differ in their weather inputs (forecast vs. resampled historical/reanalysis data), ignition modelling approach, and time horizon (specific future days vs. a synthetic long-run climatology).
• Burn probability values in both datasets are expressed as a fraction between 0 and 1 (not a percentage).
• Flame length is expressed in feet in both datasets and can be used as an indicator of potential fire intensity and suppression difficulty.