Open · Reproducible · Explainable Remote Sensing
Wildfire Danger & Burn Severity Mapping with Google Earth Engine
An open-access workflow for Mediterranean ecosystems that combines pre-fire danger mapping, post-fire USGS-style burn severity assessment, and interpretable statistical modeling. The approach uses freely available satellite, meteorological, and ancillary data to produce transparent hazard products that can be audited, explained, and transferred across regions.
Why this workflow?
Mediterranean wildfires are shaped by climate stress, vegetation condition, wind, terrain, fuel dryness, and human exposure. Operational products must be accurate, but also transparent enough to justify decisions before and after fire events.
Mediterranean fire pressure
Increasing fire frequency, size, and severity require fast, repeatable workflows for monitoring danger before ignition and assessing damage after fire.
Explainability first
Instead of relying only on black-box prediction, each pixel-level danger class is traceable to physical conditions such as fuel, dryness, heat, humidity, wind, and antecedent moisture deficit.
Open data stack
The workflow uses freely available Sentinel-2, Landsat 8, ERA5-Land, and OpenStreetMap layers for reproducible hazard mapping.
Pre-fire danger mapping.
Danger is calculated as a physically interpretable rule-based score from six binary indicators. Each indicator captures one necessary condition for elevated wildfire danger.
1. Fuel availability
Vegetation presence is estimated from optical satellite indices such as NDVI. Pixels without sufficient vegetation cannot sustain severe vegetation fire spread.
2. Fuel dryness
SWIR-sensitive moisture proxies capture dry vegetation and low fuel moisture, which increase ignition probability and spread potential.
3. Heat
Elevated 2 m air temperature from ERA5-Land or land surface temperature from optical thermal products indicates heat stress and stronger fire-conducive conditions.
4. Atmospheric dryness
Relative humidity is used to represent atmospheric moisture stress. Low humidity dries fuels and increases fire spread potential.
5. Wind speed
Wind accelerates fire spread, increases ember transport, and can transform local ignitions into fast-moving events.
6. Antecedent moisture deficit
Recent precipitation and/or soil moisture deficits capture preconditioning: dry weeks before ignition can make landscapes highly flammable.
Danger score
Danger Score = Fuel + Dryness + Heat + Atmospheric Dryness + Wind + Moisture Deficit
The score ranges from 0 to 6. A pixel with score 6 satisfies all six danger conditions. This makes the map immediately explainable: the class is not only a number, but a count of triggered physical risk factors.
0
Minimal
Minimal
1
Very Low
Very Low
2
Low
Low
3
Moderate
Moderate
4
High
High
5
Very High
Very High
6
Extreme
Extreme
Post-fire burn severity.
Burn severity is standardized using NBR and dNBR. This enables consistent comparison between fires and regions using established USGS-style thresholds.
NBR and dNBR
NBR = (NIR - SWIR2) / (NIR + SWIR2)
dNBR = pre-fire NBR - post-fire NBR
dNBR_scaled = dNBR × 1000
dNBR = pre-fire NBR - post-fire NBR
dNBR_scaled = dNBR × 1000
Healthy vegetation generally has high NIR reflectance and lower SWIR response. Burned areas typically show reduced NIR and increased SWIR reflectance, causing NBR to decrease after fire. Larger positive dNBR values usually indicate higher burn severity.
Why dNBR?
- It compares pre-fire and post-fire vegetation condition.
- It is widely used for burn severity mapping.
- It can be implemented efficiently in Google Earth Engine.
- It provides a transparent, auditable post-fire impact metric.
Sentinel-2: B8 and B12
Landsat 8: B5 and B7
Output: continuous dNBR
Output: severity class
USGS-style dNBR severity classes
| dNBR × 1000 range | Class | Interpretation |
|---|---|---|
| ≤ -251 | Enhanced Regrowth, High | Strong post-fire greening or vegetation increase. |
| -250 to -101 | Enhanced Regrowth, Low | Moderate post-fire greening or low vegetation increase. |
| -100 to 99 | Unburned | No strong burn signal. |
| 100 to 269 | Low Severity | Light burn effect or partial vegetation change. |
| 270 to 439 | Moderate-Low Severity | Clear burn signal with moderate vegetation impact. |
| 440 to 659 | Moderate-High Severity | Strong burn impact and substantial vegetation loss. |
| ≥ 660 | High Severity | Severe vegetation and surface change. |
Enhanced
Regrowth High
Regrowth High
Enhanced
Regrowth Low
Regrowth Low
Unburned
Low
Severity
Severity
Moderate-Low
Severity
Severity
Moderate-High
Severity
Severity
High
Severity
Severity
Google Earth Engine implementation.
The workflow creates pre-fire and post-fire composites, masks clouds, computes NBR/dNBR, classifies burn severity, and exports maps and pixel-level samples for open-source modeling.
Remote-sensing workflow
1. Define fire event and bounding box
Each fire is represented by event dates and a spatial analysis region.
2. Build pre-fire composite
A multi-week period before ignition is used to represent baseline vegetation condition.
3. Build post-fire composite
A post-fire window after smoke and active fire effects is used to estimate damage.
4. Compute NBR and dNBR
NBR is calculated for both periods, then differenced to produce dNBR.
5. Classify severity
dNBR × 1000 is converted into USGS-style severity classes.
6. Export maps and samples
Raster outputs and pixel samples are exported for visualization, validation, GAMs, and hierarchical ordinal logistic regression.
Core dNBR logic
var preNBR = preComposite
.normalizedDifference(['B8', 'B12'])
.rename('pre_NBR');
var postNBR = postComposite
.normalizedDifference(['B8', 'B12'])
.rename('post_NBR');
var dNBR = preNBR
.subtract(postNBR)
.rename('dNBR');
var dNBRScaled = dNBR
.multiply(1000)
.rename('dNBR_USGS_scaled');
var severity = ee.Image(0)
.where(dNBRScaled.lte(-251), 1)
.where(dNBRScaled.gt(-251).and(dNBRScaled.lte(-101)), 2)
.where(dNBRScaled.gt(-101).and(dNBRScaled.lte(99)), 3)
.where(dNBRScaled.gt(99).and(dNBRScaled.lte(269)), 4)
.where(dNBRScaled.gt(269).and(dNBRScaled.lte(439)), 5)
.where(dNBRScaled.gt(439).and(dNBRScaled.lte(659)), 6)
.where(dNBRScaled.gt(659), 7)
.rename('USGS_Burn_Severity');Greek wildfire contexts.
The workflow is designed for comparative analysis across Mediterranean fire regimes, including Attica, Euboea, Rhodes, Evros, and additional Greek fire events.
Example fire-event database
| Fire event | Dates | Region | Workflow purpose |
|---|---|---|---|
| Varympompi | Aug 4–26, 2021 | Attica | High human exposure. |
| Northern Evia | Aug 3–11, 2021 | Euboea / Evia | Large forest-fire footprint and extended burn progression. |
| Rhodes Megafire | July 18–28, 2023 | Rhodes | Island-scale fire spread under summer dryness. |
| Evros / Dadia | Aug 19–Sept 4, 2023 | Thrace / Evros | Large protected-area fire context with major ecological relevance. |
The same analysis structure can be applied to Ancient Olympia, Schinos, Agia Paraskevi, Varnavas, Patra/Achaia, and additional fire events by changing event dates and bounding boxes in the GEE event list.
Explainable modeling layer.
The rule-based score is complemented by interpretable statistical models that quantify driver effects, nonlinear thresholds, regional differences, and cross-region transferability.
Generalized Additive Models
GAMs use continuous predictors to recover nonlinear response curves. They show how danger or burn severity changes along gradients of dryness, heat, wind, vegetation, and antecedent moisture.
Ordinal logistic regression
Hierarchical ordinal logistic regression models ordered severity classes while allowing baseline levels and selected driver effects to vary by region.
Leave-one-region-out validation
Transferability is evaluated by training on all but one region and testing on the withheld region. This reveals where the workflow generalizes and where local recalibration is required.
Interpretable outputs
- Pixel-level danger trigger maps.
- Continuous dNBR burn-severity maps.
- USGS-style classified burn-severity maps.
- GAM response curves for continuous driver effects.
- Ordinal logistic odds ratios with uncertainty intervals.
- Cross-region validation diagnostics.
Physical drivers linked to decisions
Because the workflow separates fuel, dryness, heat, humidity, wind, and antecedent moisture conditions, the final danger product can be explained in plain language: a high-risk pixel is high-risk because specific environmental conditions co-occurred there before ignition.
Transparent
Auditable
Reproducible
Operational
Open-source ready
Suggested citation text
Copy this text into conference materials, posters, slides, or project pages.
Notas, A., Frousiou, M.-S., and Papadomarkakis, D.: An Open and Explainable Google Earth Engine Workflow for Wildfire Danger and Burn Severity Mapping in Mediterranean Ecosystems, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-6499, https://doi.org/10.5194/egusphere-egu26-6499, 2026.