Wildfire Probability Analysis

01 — Project Overview

Mapping Wildfire Risk
in Marin County

Marin County sits at the intersection of mountainous terrain, dense native forests, shrubland, and one of California's most populated coastal corridors — making it uniquely vulnerable to the compounding effects of climate change and historic fire suppression. This MA capstone research investigated which environmental factors most strongly correlate with wildfire occurrence and assessed how accurately FEMA's National Risk Index (NRI) predicts historical fire patterns.

Using ArcGIS Pro as the primary platform, I constructed a multi-layer raster analysis pipeline that integrated seven environmental stress variables — precipitation, temperature, soil moisture, drought index, elevation, wind, and land cover — into a Boolean wildfire probability raster for Marin County. This was refined using Wildland-Urban Interface (WUI) data to prioritize high-risk areas near populated zones, and validated against a historical fire boundaries dataset spanning from the 1920s to 2020.

A linear regression analysis, conducted in GeoDa, then correlated 222 MODIS thermal anomaly points against all seven environmental variables to statistically identify which factors had the strongest relationship to actual fire occurrence.

Figure 1 — Marin County National Risk Index (NRI) · FEMA Census Tract Analysis · Displaying Expected Annual Loss, Social Vulnerability, Community Resilience, and Overall Risk scores across the region.

03 — Methodology

The Analytical
Pipeline

Step 01 — Data Acquisition

Multi-Source Environmental Dataset Assembly

Assembled seven environmental datasets across raster and vector formats from national and local sources: precipitation and temperature rasters at 4km resolution from PRISM; a 1m DEM and 100m wind raster from USGS and Global Wind Atlas; soil moisture vectors from USDA NRCS; land cover from the Golden Gate National Parks Conservancy; and drought index vectors from FEMA. MODIS thermal anomaly points (2018–2020) served as the fire occurrence dataset, filtered to 222 confirmed points within Marin County.

Step 02 — Data Preparation

Clip, Select by Attribute & Vector-to-Raster Conversion

Each dataset was clipped to the Marin County boundary using the ArcGIS Pro Data Management Clip tool — with separate vector and raster versions applied to each format. Environmental stress thresholds were then isolated using Select by Attribute: soil moisture ≤20%, temperature ≥19°C, precipitation ≤10%, wind ≥7 (>50th percentile), and elevation ≤295m. Fire-prone land cover types (native/non-native forests, shrubland, cropland) were selected and exported as new layers. All vector layers were then converted to raster format at 50m cell size using Polygon to Raster for compatibility with the Raster Calculator.

Step 03 — Raster Math

Boolean Wildfire Probability Surface

Each environmental raster was processed through two rounds of Raster Calculator operations. Round 1 extracted Boolean layers (1 = stress condition present, 0 = absent) for each variable. Round 2 multiplied all layers together, producing a composite probability surface highlighting areas where all environmental stress factors converge. The DEM layer was retained at continuous values to enforce the elevation threshold. A second raster calculation incorporated the WUI dataset to narrow results to high-risk areas near populated zones.

Step 04 — Linear Regression

Spatial Regression in GeoDa

Environmental attribute values were spatially joined to the 222 MODIS thermal anomaly points: soil moisture, drought index, and land cover were joined via Spatial Join; temperature, precipitation, and wind values were manually entered per point due to raster format incompatibilities. Land cover types were translated to numerical factorial values to enable regression. The three annual TA point datasets (2018, 2019, 2020) were merged into a single shapefile and exported to GeoDa for ordinary least squares linear regression.

Wildfire Risk Analysis Map — Figure 2 — Wildfire Risk Analysis · All environmental stress factors applied · High-risk zones in southern interior and northwestern Marin County.

Wildfire Risk with WUI — Figure 3 — Wildfire Risk with Wildland-Urban Interface · Risk refined to populated corridors · Bay coast, southern ridgeline, and northern lagoon regions.

04 — Data Sources

Environmental
Datasets Used

Dataset	Source	Format	Resolution	Role in Analysis
Precipitation	PRISM Climate Group	Raster	4 km	Annual precip ≤10% stress threshold
Temperature	PRISM Climate Group	Raster	4 km	High temp ≥19°C stress threshold
DEM / Elevation	USGS National Map	Raster	1 m	Elevation threshold ≤295m (amendable zones)
Wind	Global Wind Atlas	Raster	100 m	Wind ≥7 (>50th percentile) stress threshold
Soil Moisture	ArcGIS / USDA NRCS	Vector	Local	Water retention ≤20% (0–100cm depth)
Land Cover	Golden Gate NPS Conservancy	Vector	Local	Fire-prone vegetation classification
Drought Index	FEMA	Vector	County-level	Severe–Exceptional drought zones
MODIS Thermal Anomalies	NASA MODIS	Vector	Point	Fire occurrence validation (2018–2020)
Historical Fire Boundaries	Cal Fire / FRAP	Vector	Local	Cross-reference validation (1920s–2020)
WUI Zones	Cal Fire / FRAP	Vector	Local	Prioritize high-risk populated areas

05 — Key Findings

Results &
Conclusions

Finding 01 — Environmental Analysis

High-Risk Zones Confirmed by Historical Record

The composite raster analysis identified high-risk wildfire zones concentrated in Marin County's southern interior mountain regions, with additional risk along the northwest and scattered southeastern sections. When cross-referenced against FRAP historical fire boundaries spanning the 1920s through 2020, calculated high-risk areas matched recorded fire locations from multiple decades — semi-confirming that multi-variable environmental analysis can spatially predict wildfire zones.

Finding 02 — WUI Refinement

WUI Integration Refocused Risk to Populated Corridors

Adding the Wildland-Urban Interface layer to the raster calculation shifted the highest-priority risk areas from broad southern zones to the northwest interior mountains, inland bay coastline, and southern ridgeline forest areas — aligning risk output with the areas most critical for emergency management response.

Finding 03 — Linear Regression

Soil Moisture & Land Cover Most Strongly Correlated

Of the seven variables tested, soil moisture and vegetation land cover showed positive statistical relationships with thermal anomaly occurrence. Temperature, precipitation, and drought produced negative coefficients — attributed largely to data resolution limitations at 4km rather than a true inverse relationship with fire.

Finding 04 — NRI Comparison

NRI Accurately Reflects Recent Fires, Not Historical Patterns

FEMA's National Risk Index flagged the county's north-northwest as highest risk — which aligned with 2020 fire locations but diverged significantly from the century-long historical record, where the majority of fires clustered in the southern regions. The NRI's economic-demographic model produces a fundamentally different spatial pattern than ecologically-grounded environmental analysis.

06 — Statistical Results

Linear Regression
Summary

Variable	Coefficient	Std. Error	t-Statistic	Probability
Constant	365.74	28.05	13.04	0.000 ✦
Temperature	−4.63	1.86	−2.49	0.013 ✦
Precipitation	0.014	0.016	0.91	0.365
Soil Moisture ✦	0.253	0.389	0.65	0.516
Land Cover ✦	19.50	6.33	3.08	0.002 ✦
Drought Index	0.522	3.96	0.13	0.895
Wind	−4.21	2.22	−1.90	0.059

13.6%

R² — Variance Explained

Interpreting the Results

The R² value of ~13.6% reflects the model's explanatory power given the constraints of available data. Coarser-resolution datasets for temperature and precipitation (4km) and the county-level drought index likely suppressed statistical relationships that finer data would reveal.

Land cover was the strongest statistically significant predictor (p=0.002), while soil moisture showed a positive relationship consistent with fire ecology theory. The researcher noted that higher-resolution drought and climate datasets, combined with machine learning approaches, could substantially improve model performance.

✦ Positive relationship with fire occurrence · ✦ p < 0.05 significant

07 — Technical Toolkit

Software &
Methods Used

ArcGIS Pro

Primary GIS platform — raster analysis, clipping, attribute selection, raster calculator, polygon-to-raster conversion, spatial joins, and thematic map production

GeoDa

Spatial statistics platform used for ordinary least squares linear regression of thermal anomaly points against seven environmental variables

Raster Calculator

Boolean multiplication of seven environmental stress rasters to produce composite wildfire probability surface across Marin County

MODIS Thermal Anomalies

NASA MODIS satellite-derived fire detection points filtered to 222 confirmed thermal anomalies within Marin County for 2018–2020

WUI Analysis

Cal Fire / FRAP Wildland-Urban Interface data integrated to refine risk outputs toward populated areas requiring priority emergency response

FEMA NRI Validation

National Risk Index compared against environmental analysis outputs and historical fire boundaries to assess model accuracy and limitations

Interested in
this research?

GIS Engineer & Geospatial Scientist · Portland, Oregon · Open to new roles

View LinkedIn Get in Touch

Mapping Wildfire Riskin Marin County

Three CoreQuestions

The AnalyticalPipeline

EnvironmentalDatasets Used

Results &Conclusions

Linear RegressionSummary