Analysis of the 2024 Achaia-Ilia wildfire in Greece using Sentinel-2 Level-2A imagery from the Copernicus Data Space Ecosystem. Wildfires are one of the most impactful phenomena on the natural equilibrium of the Earth's surface, resulting in damage to the environment with effects on both the economic and social aspects of a nation. The European Forest Fires Information System (EFFIS) published data highlighting how the average area affected by fires in 2025 in European Union member states was more than double the average across 2006-2024.
This investigation is based on the application of spectral indices to pre- and post-fire scenes to map burn extent and classify severity based on the United States Geological Survey (USGS) ranges. Results were validated against the Copernicus Emergency Management Service (CEMS).
For a detailed description of the methodology, please refer to the associated article: "How to detect and assess wildfires with satellite data"
The analysis has been structured in sequential steps:
- Sensitive elements selection: The first step was the definition of the Area of Interest (AOI) and its sensitive elements. In this phase, multiple geospatial vectors were analyzed to extract the information used to construct a GeoPackage file for the sensitive elements present in the AOI. Roads, land use, and protected areas were extracted from public databases like OpenStreetMap, the Database of Global Administrative Areas, and Geofabrik.
- Raster data analysis: The second phase of the analysis used the STAC catalog exposed by the Copernicus project to download the pre- and post-fire scenes. This step included the generation of a raster mosaic from two different tiles and its pre-processing to reduce the analysis data to only what is required. In particular, raster data is clipped using the bounding box of the vector data, and only bands B8A and B12 are retained.
-
CEMS data retrieval: The CEMS report number 730 was analyzed through its exposed API service to reconstruct the AOI used for the validation.
-
Burned area detection: The burned scar detection is based on the widely used spectral index called Normalized Burn Ratio:
$$ \text{NBR} = \frac{\text{NIR} - \text{SWIR}}{\text{NIR} + \text{SWIR}} $$ Vegetation strongly reflects in the NIR band, so healthy forests and crops appear very bright. Water appears dark, and bare soil sits in the middle. In the SWIR band instead, healthy vegetation does not have high reflectance, while bare soil appears brighter. Water is dark here, as in the NIR.
The reflectance of the two bands used for the pre- and post-fire scenes is shown in the picture below:
By leveraging the different spectral responses of these Earth surfaces, and by computing the difference between the pre- and post-fire values, we are able to segment the burned area and assess the wildfire impact. The difference of the two NBR values is called dNBR, and the burn severity map is based on the classification levels provided by the United States Geological Survey (USGS):
| dNBR range | Severity class |
|---|---|
| [-1, -0.251] | Enhanced regrowth (high) |
| [-0.250, -0.101] | Enhanced regrowth (low) |
| [-0.100, 0.099] | Unburned |
| [0.100, 0.269] | Low severity |
| [0.270, 0.439] | Moderate-low severity |
| [0.440, 0.659] | Moderate-high severity |
| >= 0.660 | High severity |
The generated burn severity map with associated severity level distribution is reported below:
A comparison of the segmented area and the one provided by CEMS is shown below:
The region of the post-fire scene that was covered by clouds during data acquisition is clearly visible as a white hole in the segmentation map, close to the coordinates 556 E and 4199 N. The reason is that clouds have high reflectance in both NIR and SWIR, which causes the dNBR close to 0.
The accuracy of the methodology has been evaluated with 3 common computer vision metrics:
| Precision | Recall | IoU |
|---|---|---|
| 0.87 | 0.86 | 0.76 |
During the analysis, the sensitivity of the dNBR index to specific Earth surfaces and atmospheric elements was analyzed and mitigated with different techniques and the final segmentation has been achieved by selecting the polygon with the biggest interconnected area:
Other indices were computed for the wildfire segmentation, namely the Relativized Differenced Normalized Burn Ratio (RdNBR) and the Relativized Burn Ratio (RBR), but neither of them provided a significant improvement compared to the dNBR for this specific scene.
An interesting result was the effectiveness in the visual evaluation of the scar by using a false-color image composed of the bands (SWIR, NIR, Green):
The high reflectivity of ash in the SWIR, and its low reflectivity in the NIR, allows the scar to stand out in the image.
notebooks: analysis files containing a detailed description of the adopted workflow.scripts: bash files used to automatically download files used in the analysis.data: data downloaded with the scripts or generated during the analysis.src: custom helper functions and types used in the notebook.
- Python 3.13+.
- uv for environment management.
- A Copernicus Data Space Ecosystem (CDSE) account to access remote Sentinel-2 imagery.
Create a virtual environment containing the packages used in the notebooks:
make installThe notebooks that pull data directly from the Copernicus S3 bucket require an access key and secret key from your CDSE account. Copy the example env file and fill in your credentials:
cp .env.example .envThen edit .env:
ACCESS_KEY=your_access_key
SECRET_KEY=your_secret_key
Create an IPython kernel for the project and start the Jupyter-lab server.
make create-kernel
make start-jupyterNotebook names are prepended with a number to highlight the proper order of execution.
| Notebook | Description |
|---|---|
00-geospatial-vectors |
Analyze Greece vectors data to create the GeoPackage files for the area of interest and the critically identified elemennts. |
01-raster-exploration-and-mosaic |
Create raster mosaics for the area of interest by using Sentinel-2 L2A data. |
02-cems-exploration |
Analyze the CEMS report via the exposed API and reconstruct the analysis final product. |
03-wildfire-evaluation |
Uses data generate in previous notebooks to segment the burned area and assess severity classes. |
Run tests and format code:
make format- Perform automatic segmentation by introducing CORINE classes to filter agricultural lands.
- Create specific scripts out of each notebook to parametrically perform the analysis with a generic area of interest.
- Average across multiple scenes to remove noise due to clouds and other phenomena.
- Evaluate how sensitive the segmentation is to the time elapsed between the event and the scene acquisition.