Real-Time Wildfire Risk Intelligence
PyroSight
Click anywhere on the US map to get a next-day fire spread prediction powered by live satellite weather data, evidential deep learning, and Rothermel fire physics.
How It Works
From Click to Risk Assessment
When you click a location on the map, PyroSight fetches real environmental data and runs a physics-informed deep learning model to predict where fire will spread in the next 24 hours.
01
Fetch Live Data
12 environmental channels are assembled from GRIDMET (weather), NASA FIRMS (active fires), MODIS (vegetation), SRTM (elevation), and GPWv4 (population) — all for a 64×64 km area centered on your click.
02
Run Inference
A dual-branch U-Net processes fuel data (terrain, vegetation) and weather data (wind, temperature, humidity) through cross-attention modules, then fuses with Rothermel fire-spread physics equations.
03
Assess Risk
The model outputs a per-pixel fire probability map with calibrated uncertainty via Evidential Deep Learning. Risk is classified as Critical, High, Moderate, or Low based on peak probability and area coverage.
Real Data Sources
12 Environmental Channels
GRIDMET
Wind, Temp, Humidity, Precip, ERC
Daily, 4km
NASA FIRMS
Active Fire Detections
Hourly, 375m
SRTM
Terrain Elevation
Static, 30m
MODIS
Vegetation Index (NDVI)
8-day, 500m
GPWv4
Population Density
Static, 1km
GRIDMET
Drought Index (PDSI)
5-day, 4km
GRIDMET
Fire Energy (ERC)
Daily, 4km
MODIS
Prior Fire Mask
Daily, 1km
The Science
Model Architecture
Dual-Branch U-Net
The fuel branch (3×3 convolutions) processes terrain, vegetation, population, and prior fire data. The weather branch (5×5 depthwise-separable convolutions) handles wind, temperature, humidity, and precipitation. Cross-Attentive Feature Interaction Modules (CAFIM) fuse the branches at three encoder scales, learning where weather amplifies fuel risk.
Rothermel Physics
A deterministic physics branch computes fire spread rate from the Rothermel equations — incorporating slope gradients (Sobel filters on elevation), wind direction vectors, and fuel moisture proxies. These physics features are fused with the neural network output, ensuring predictions respect physical fire behavior: fire goes downwind, uphill, and through dry vegetation.
Evidential Deep Learning
Instead of a standard softmax, the model outputs Dirichlet distribution parameters (α) for each pixel. This provides calibrated epistemic uncertainty in a single forward pass — no ensembles or Monte Carlo dropout needed. The model can say “90% fire probability with high confidence” or “60% probability but I’m very uncertain” — a critical distinction for operational decision-making.
Risk Classification
CRITICALPeak >80%, >5% high-risk area.HIGHPeak >50%, >2% area.MODERATEPeak >30%.LOWMinimal risk. Confidence is derived from the model’s evidential uncertainty.
Try It Now
Click anywhere on the US map to assess wildfire risk using real-time satellite and weather data. No account needed.
Open the Risk Map →