WILDFIREAI / Dixie Fire 2021

Active · 960k acDemo

WildfireAI, Dixie Fire Spread Forecast

Spatiotemporal Wildfire Spread Forecasting

Dissertation research · geospatial deep learning

Multi-model deep learning ensemble · 2021 Dixie Fire, Northern California · ConvLSTM · 3D U-Net · Physics-informed hybrids · Optuna-tuned ensemble

0.3996

Best IoU

0.9666

Best ROC-AUC

102

Days

Channels

Architectures

System Pipeline

6 layers

GEE Ingestion

ERA5-Land · FIRMS · NDVI · GlobFire · Static 1km rasters

›

Feature Eng.

20-ch cube (T=3) · NDVI Δrate · FIRMS lag · Aspect sin/cos

›

Model Training

Focal+Dice · α=0.9, γ=3 · Optuna on ensemble

›

Registry

registry.json · Best: UNet3D · IoU=0.3996

›

Inference

burn_prob.tif · burn_mask.tif · threshold=0.50

›

Dashboard

FastAPI backend · Interactive map & charts

Next-Day Burn Probability

3D U-Net, best model

39.88°N 121.37°W · Feather River Canyon · Plumas/Butte Co., CA

High (>0.7) Med (0.4–0.7) Low (0.2–0.4) Unburned MTBS Perimeter Ignition

Date2021-08-13

Drag to pan · scroll or buttons to zoom · timeline scrubs simulated spread

Feature Importance

Permutation IoU

Static Fuel Model

static

+0.0234

Total Precipitation

weather

+0.0154

FIRMS Lagged Density

dynamic

+0.0116

Aspect Cosine

static

+0.0048

Elevation

static

+0.0047

NDVI Change Rate

dynamic

+0.0017

Aspect Sine

static

+0.0013

Mean Temperature

weather

+0.0002

Static Slope

static

+0.0000

Max Wind Speed

weather

-0.0005

■ Static fuel model is the strongest predictor.
■ Wind speed showed negative importance (−0.0005), Dixie was fuel-moisture driven, not wind-driven.

Model Comparison

5 architectures

Model	IoU↑	F1↑	Prec↑	Rec↑	AUC↑	Thresh
ConvLSTM-Only	0.3773	0.5478	0.6322	0.4833	0.9342	@0.90
ConvLSTM+Physics	0.3967	0.5680	0.6055	0.5350	0.9594	@0.90
3D U-Net ★	0.3996	0.5710	0.5367	0.6100	0.9507	@0.50
3D U-Net+Physics	0.3925	0.5637	0.6145	0.5207	0.9666	@0.75
Hybrid Ensemble	N/A	N/A	N/A	N/A	N/A	N/A

Metric Radar

Select model

IoU & F1 Comparison

Best IoU

0.3996

3D U-Net @0.50

Best ROC-AUC

0.9666

UNet3D+Physics

Best F1

0.5710

3D U-Net @0.50

Threshold Sensitivity

3D U-Net

Early warning

0.35–0.45

High recall

Incident command

0.90–0.95

High precision

Environmental Clustering

K=2 · silhouette=0.6711

CLUSTER 0

New ignition / leading edge

Fire label: 1.0 (burning)
FIRMS lag density: ≈ 0
Max wind: 3.11 m/s
Use case: Early detection

CLUSTER 1

Established / active zone

Fire label: ~0.45
FIRMS lag density: Very high
Terrain: Higher elev, steep
Use case: Siege monitoring

0.6711

Silhouette (K=2). K-means on fire-relevant pixels only; filtering unburned pixels separated regimes clearly.