Field Lab 06 · Change Detection · Northern Evia, Greece
Investigate how satellite-observed vegetation condition changed after the August 2021 northern Evia wildfire—and whether the subsequent spectral recovery was spatially uniform.
A return of NDVI or NBR toward the pre-fire spectral baseline can indicate renewed greenness or changes in canopy and moisture structure. Sentinel-2 alone cannot prove recovery of species composition, biomass, forest structure, habitat quality or ecosystem function.
Investigation path
Define the event, evidence and non-claims before processing imagery.
Anticipate seasonal confounding, cloud artefacts and index behaviour.
Connect NIR–SWIR and red–NIR responses to NBR and NDVI.
Create masked, matched-season composites and transparent metrics.
Audit valid observations, ranges, grids, denominators and exports.
Describe spatial and temporal spectral patterns cautiously.
Communicate uncertainty and separate spectral from ecological recovery.
Authoritative spatial evidence
Copernicus Emergency Management Service activated EMSR527 on 4 August 2021 for wildfires in Greece. This project bundles the AOI01 Monit03 observed-event vector delivered on 11 August 2021 and filters the feature whose notation is Burnt area. The source product records that feature as 50,909.887828 hectares.
The outline below is drawn from the bundled authoritative GeoJSON, not from an illustrative hand-drawn polygon. The vector remains an emergency-mapping observation with its own acquisition date and method—not timeless ground truth.
Question before pixels
notation = Burnt areaCOPERNICUS/S2_SR_HARMONIZEDThis analysis does not directly measure species composition, tree survival, above-ground biomass, habitat quality, ecosystem function or causal drivers of recovery.
Matched seasons
Every period uses the same 1 September–15 October window. The 2019–2020 baseline reduces dependence on one year but does not remove interannual variability. The 2021 window begins after the fire and remains seasonally comparable. The project stops at 2025 because the 2026 window is not complete.
Neither example automatically proves a fair fire comparison, but matching the same seasonal window is more defensible. Spring–late-summer differences can exaggerate disturbance through phenology and moisture seasonality.
Surface reflectance
Level-2A provides atmospherically corrected surface reflectance. Earth Engine’s harmonized collection adjusts newer scenes across the Sentinel-2 processing-baseline shift, supporting a more internally consistent 2019–2025 series. Reflectance bands are stored scaled by 10,000.
B4 Red and B8 NIR: native 10 mB11 SWIR1 and B12 SWIR2: native 20 mSCL: Scene Classification Layer used for core maskingQuality assurance
The core code removes SCL 3 (cloud shadow), 8 and 9 (medium/high cloud), 10 (cirrus), 11 (snow/ice), and 6 (water) for this terrestrial analysis. Learners compare masked and unmasked imagery and map valid-observation counts. Masking changes the analysis population.
Each period is a median of valid observations, reducing sensitivity to residual cloud and outliers. It is not a single “2024 image.” Contributing dates and observation counts vary spatially and must be reported.
No. Check cloud shadow, terrain shadow, water, acquisition support, residual haze, registration and the valid-observation layer before interpreting disturbance.
Disturbance and vegetation evidence
(B8 − B12) / (B8 + B12)Tracks the contrast between near-infrared and shortwave-infrared response.
NBRpre − NBRpostRetained as a continuous result. Generic thresholds are not presented as validated Evia burn-severity classes.
(B8 − B4) / (B8 + B4)Supports greenness interpretation but can saturate and does not measure complete ecological recovery.
Green vegetation may have returned while SWIR-sensitive canopy, structure or moisture characteristics remain different. This is a spectral hypothesis, not proof of a mechanism.
Longitudinal metric
The code masks pixels where the absolute NBR denominator is below 0.05. Values near 0 resemble the immediate post-fire reference; values near 1 have returned toward the pre-fire spectral baseline; values above 1 exceed it; and negative values are farther from baseline. This is never reported as an “ecosystem recovery percentage.”
Spatial heterogeneity
The workflow divides continuous dNBR inside the EMS perimeter at its own 33rd and 67th percentiles. These are descriptive lower, medium and higher initial spectral-change strata—not universal severity classes. Annual medians and IQRs are compared without treating neighbouring pixels as independent replicates.
A conservative lag-candidate screen requires higher initial dNBR, 2025 relative recovery below 0.6, and at least two valid observations. Its output is labelled candidate persistent spectral departure, not failed ecological recovery.
Required visual evidence
Orient the event and perimeter provenance.
Inspect comparable true- and false-colour evidence.
Explain the disturbance-sensitive spectral contrast.
Show matched-season spatial recovery patterns.
Compare NDVI, NBR and valid observation support.
Test whether recovery appears spatially uniform.
Bound what the maps and summaries can establish.
Connect methods, evidence and a defensible conclusion.
Satellite composites, dNBR maps and trajectories are deliberately not pre-populated here. Learners generate them by executing the supplied workflow; the Academy does not fabricate result values or imagery.
Scientific defence
Residual cloud, haze and shadow can survive masking.
Matched dates reduce but do not eliminate seasonal or interannual differences.
B8 is 10 m; B12 is 20 m. A common grid requires resampling decisions.
A median mixes dates and spatially varying observation counts.
EMSR527 is a dated rapid-mapping product with a stated extraction method.
Similar index values can arise from different vegetation communities or structures.
Two pre-fire seasons cannot describe the full natural range.
Neighbouring pixels are spatially autocorrelated, not independent replicates.
Build and export
The script is organized as small functions for masking, index creation, seasonal compositing, period summaries, recovery fractions and exports. Upload the bundled GeoJSON to your Earth Engine project, replace the single asset ID, inspect every QA layer, then run the exports.
config.maskS2(image)
addIndices(image)
seasonalComposite(year)
recoveryFraction(current)
summarisePeriod(image)
summariseStratum(image, stratum)
Export.table.toDrive(...)
Export.image.toDrive(...)Mandatory checks
Portfolio artifact
Authoritative references