Case study · Public safety

Seeing wildfire risk before smoke rises.

Francisca is a geospatial deep-learning platform that predicts wildfire ignition risk and spread trajectories — giving incident commanders the early-staging window that saves lives, homes, and ecosystems.

Deep learning Geospatial Remote sensing Public safety
Conv-LSTM
Temporal model core
Sentinel-2
Satellite data fusion
Cross-agency
Response enablement

The problem

Wildfire response is a race between risk and readiness. Traditional early-warning systems rely on weather alerts and historical fire frequency — useful, but reactive. By the time an ignition is reported, the incident team is already behind. What commanders need is forward-looking situational intelligence: where a fire is likely to start in the next days, how fast it will move if it does, and which direction to pre-stage resources.

Our approach

Francisca treats wildfire risk as a spatiotemporal prediction problem and builds a deep-learning platform around three layers:

  • Ignition risk modeling. Conv-LSTM networks combine recent meteorological telemetry, vegetation-moisture indices from satellite imagery, and terrain rasters to predict ignition probability across a region at grid resolution.
  • Spread simulation. Once a risk zone lights up, the model projects likely spread trajectories based on wind, slope, fuel load, and historical fire-behavior signatures.
  • Decision surface. Predictions render as interactive GIS risk maps for incident command, with confidence intervals and data provenance so decisions are defensible.

Data fusion

  1. Satellite. Sentinel-2 imagery for vegetation-moisture indices and burn-scar mapping.
  2. Meteorological telemetry. Temperature, humidity, wind speed and direction, recent precipitation.
  3. Terrain. Elevation, slope, aspect, fuel-type classification.
  4. Historical. Prior ignition locations and spread patterns to anchor the model in actual fire behavior.

Tech stack

Conv-LSTM temporal modeling Sentinel-2 satellite fusion GIS risk mapping Remote-sensing analytics Meteorological telemetry integration Predictive environmental simulation

Why it mattered

Incident command got a decision surface — not just a data dump. Resources pre-staged the day before ignition instead of chasing smoke plumes. Cross-agency coordination improved because everyone was working from the same risk map, not separate interpretations of the same weather report.

Explore a similar system →