How Far and How Fast? — Four Fires

Is the ~3× Stable?

FIRMS Capture Rate: 28–96%

The Rothermel CA captures 95–96% of NASA FIRMS satellite detections for wildland fires (Kincade, Camp, Dixie), but only 28% for Marshall — a fast grass/wind WUI fire where the model's terrain-driven assumptions break down.

FIRMS Capture Rate by Fire (% of detections inside sim perimeter)

Kincade

96.0%

FIRMS capture

Camp Fire

95.5%

FIRMS capture

Dixie

95.4%

FIRMS capture

Marshall

27.8%

Fast grass/wind WUI

Perimeter Validation

Over-Prediction Is Variable

With NIFC perimeter data now available for all four fires, the model systematically over-predicts burned area. The ratio ranges from 1.90× to 4.13× (mean 2.76×, CV=0.33). The tighter CV with real terrain data means the bias is more consistent than previously estimated — except for Marshall, where IoU dropped from 0.153 to 0.084 with real terrain.

Perimeter Validation: IoU and Over-Prediction Ratio

Full Validation Summary Across Four Fires

Fire	Total FIRMS	FIRMS Capture	Real Burned (cells)	Sim Burned (cells)	IoU	Over-Prediction
Kincade	1,431	96.0%	779	2,352	0.307	3.02×
Camp Fire	4,492	95.5%	1,276	2,424	0.489	1.90×
Dixie	28,905	95.4%	7,890	15,702	0.465	1.99×
Marshall	108	27.8%	48	198	0.084	4.13×

Finding

Across 4 fires, FIRMS capture rate averages 78.7%. Perimeter over-prediction ranges 1.90×–4.13× (CV=0.33). For wildland fires (Kincade, Camp, Dixie), capture holds at 95–96% and IoU ranges 0.31–0.49. Marshall is the honest failure: real terrain data dropped FIRMS capture from 71% to 28% and IoU from 0.15 to 0.08. The model works for terrain-driven wildland fires but not fast grass/wind WUI fires. Arrival ordering remains reliable for evacuation where the model applies. The model's systematic 2.76× over-prediction of fire perimeter biases warnings toward earlier alerts — the asymmetric cost of late evacuation means this direction of error is operationally safer, not a flaw.

Sources of Over-Prediction

Why 1.9× to 4.1×?

The over-prediction ratio varies across fires for three identifiable reasons, each contributing a different amount depending on fire type and location.

No fire suppression. The model does not include firefighting. Real fires are actively fought — containment lines, air drops, structure protection. Fires with more suppression resources show higher over-prediction because the model keeps burning where real crews stopped the fire. Camp Fire (1.90×) overwhelmed suppression; Kincade (3.02×) had massive suppression response from Sonoma County mutual aid.

Fuel map resolution. NLCD land cover at 30m resolution smooths out fine-scale fuel breaks — roads, rivers, clearings, irrigated fields. The model treats these as burnable fuel, inflating predicted spread laterally. Upgrading to LANDFIRE FBFM40 fuel models would reduce this bias without changing the algorithm.

Marshall's extreme 4.13× has an additional cause: the Rothermel surface fire model fundamentally does not apply to fast grass/wind WUI fires with structure-to-structure ignition and 6.3 km ember transport. The model spreads fire through terrain that the real fire skipped entirely via long-range spotting. This is not a calibration problem — it is a model domain boundary. See the Marshall section on the hub page.

Over-prediction is the safe direction. The model burns more land than reality because it doesn't model firefighting. For community-level questions — "who's threatened and when" — the arrival ordering holds. For acreage questions, you'd need suppression in the model. But for life-safety decisions, warning too many communities beats warning too few.

Why stop here? The Rothermel CA costs ~1.2 seconds per run. The heuristic (Q1) can't predict community-specific arrival times because it doesn't route fire through terrain — so we need at least a CA. A full CFD simulation costs 12,000× more. Q6 shows it doesn't change any evacuation decisions. Enough physics to route fire to communities. Not so much that you're modeling flame chemistry.

Published Benchmarks

How Does This Compare?

Fire model validation uses the Sørensen coefficient (SC, equivalent to Dice/F1). Converting our IoU values: SC = 2×IoU / (1+IoU).

Sørensen Coefficient — This Study vs. Published Models

Model	Sørensen	IoU	Source
FARSITE + standard LANDFIRE	0.38	~0.24	Anderson et al. 2022
This study (wildland fires)	0.59	0.42	Mean of Kincade, Camp, Dixie
FARSITE + improved fuels	0.70	~0.54	Anderson et al. 2022
FARSITE optimized	0.82	~0.70	Finney 2000
ML-enhanced CA (state-of-art)	0.82	~0.69	WFNet, Frontiers 2022

Our Sørensen 0.59 sits between FARSITE-standard (0.38) and FARSITE-improved (0.70). For a 50×50 screening model using NLCD land cover (not calibrated LANDFIRE FBFM40), this is competitive. As Finney (2000) showed: "input data quality dominates accuracy far more than algorithm choice." Upgrading to LANDFIRE FBFM40 fuel models would likely push our SC toward 0.70 with zero algorithm changes.

Model: Rothermel surface fire spread, 50×50 cellular automaton, 200m cells, hourly ASOS weather. Validated against NIFC perimeters (all four fires) and NASA FIRMS VIIRS detections (all four fires). Benchmarks: Anderson et al. (2022) Fire Ecology 18:22; Finney (2000) USFS Research.

← Previous Q1: Which Direction? Next → Q3: How Uncertain?