Python pipeline that produces training data from Sentinel-2 L2A satellite imagery for change detection models. It automates the download, processing, and tiling of multi-temporal image pairs from a user-defined area of interest (AOI).
Change detection models learn to identify landscape changes by comparing before and after satellite images. This pipeline:
- Finds the best cloud-free images for two time periods (T1 and T2) using the STAC API
- Downloads and crops only the AOI at native 10m resolution
- Filters clouds and shadows using the Scene Classification Layer (SCL)
- Normalizes spectral bands to a consistent [0, 1] range
- Tiles the output into fixed-size patches suitable for training
- Creates a manifest linking T1/T2 pairs with their metadata for reproducibility
The pipeline is designed to be containerized and orchestrated (e.g., via Airflow), with all parameters configurable via CLI or environment variables.
For more information about the architecture: system_design.md
- Python 3.10+
- GDAL/rasterio system libraries
- Clone the repository
git clone https://github.com/etienne912/sentinel-2-data-processing.git
cd sentinel-2-data-processing- Install dependencies
Using uv (recommended):
uv syncOr with pip:
pip install -e .- Run the pipeline
python -m src.main --helpusage: main.py [-h] --aoi AOI --t1-start YYYY-MM-DD --t1-end YYYY-MM-DD --t2-start YYYY-MM-DD --t2-end YYYY-MM-DD [--tile-size TILE_SIZE]
[--bands BAND [BAND ...]] [--output-dir OUTPUT_DIR]
Change detection pipeline for Sentinel-2 L2A imagery.
options:
-h, --help show this help message and exit
--aoi AOI Path to a GeoJSON file defining the area of interest.
--t1-start YYYY-MM-DD
Start date for the T1 (before) search window.
--t1-end YYYY-MM-DD End date for the T1 (before) search window.
--t2-start YYYY-MM-DD
Start date for the T2 (after) search window.
--t2-end YYYY-MM-DD End date for the T2 (after) search window.
--tile-size TILE_SIZE
Output patch size in pixels. Default: 256.
--bands BAND [BAND ...]
List of Sentinel-2 bands to process (e.g. B02 B03 B04 B08).
--output-dir OUTPUT_DIR
Directory where processed patches and manifest will be saved. Default: ./output.
output/
├── manifest.json # Links T1/T2 tiles + metadata
├── t1/
│ └── S2*_..._{col}_{row}.tif
└── t2/
└── S2*_..._{col}_{row}.tif
Each GeoTIFF contains:
- CRS: Projected coordinate system (e.g., UTM)
- Data type: float32, values in [0, 1], NaN for masked pixels
- Band count: Number of requested bands
- Tags: Source product ID, acquisition date, band names
The manifest provides:
- T1 and T2 source product IDs
- Acquisition dates and cloud cover
- Tile file paths and their geospatial bounds
On 25 March 2026, a new bridge (Anne de Bretagne) was installed in Nantes, France. This is a perfect test case because:
- Small AOI (~1 km²) = single Sentinel-2 tile
- Clear before/after imagery available
- Visible change in both optical and infrared bands
You can find more information about the bridge installation in Ouest-France (French local newspaper).
An example GeoJSON AOI and expected outputs are in docs/exemple/:
nantes_bridge.geojson— A small polygon around the bridgeS2B_30TXT_20260318_0_L2A.png— Before image (18 March 2026)S2A_30TXT_20260407_1_L2A.png— After image (7 April 2026)
Selected AOI: A small polygon around the bridge, and construction site.
This is a plot of the AOI over high-quality satellite imagery (Maxar).
python -m src.main --aoi docs/example/nantes_bridge.geojson --t1-start 2026-03-01 --t1-end 2026-03-24 --t2-start 2026-04-01 --t2-end 2026-04-17 --bands blue green red --tile-size 256Before (18 March 2026) — The construction site is visible, but the bridge deck is not yet in place:
RGB composite image from Sentinel-2.
The image shows:
- Urban area
- River and surrounding vegetation
- Construction equipment and staging area
After (7 April 2026) — The bridge structure is complete and visible:
RGB composite image from Sentinel-2.
The image shows:
- Bridge deck clearly visible in the center (linear structure across the river)
- Same urban and vegetation patterns
- Cloud cover is low in both cases, making them ideal training data
Running the command above produces:
- 128×128 pixel tiles (GeoTIFF format, float32) covering the bridge area
- Two sets: one from T1 (before) and one from T2 (after)
- Three bands: Blue, Green, Red
- manifest.json: Metadata linking each T1/T2 pair with source product IDs and acquisition dates
A change detection model can now learn that:
- Before: River + vegetation + urban area
- After: River + vegetation + urban area + bridge structure
The presence/absence of the bridge is the "change" the model learns to detect.
To integrate into an Airflow DAG, import the pipeline function and pass parameters:
from src.main import run_pipeline
import shapely
from airflow.operators.python import PythonOperator
run_task = PythonOperator(
task_id="process_sentinel2",
python_callable=run_pipeline,
op_kwargs={
"aoi": shapely.from_wkt("POLYGON((...))"),
"t1_date_range": ("2026-03-01", "2026-03-24"),
"t2_date_range": ("2026-04-01", "2026-04-17"),
"bands_keys": ["blue", "green", "red"],
"tile_size_px": 128,
"output_dir": "/mnt/data/tiles_output/",
}
)The function returns a dict with manifest path and tile locations, suitable for downstream XCom.
Build and run in a containerized environment:
docker build -t sentinel2-pipeline .
docker run -v $(pwd)/output:/output sentinel2-pipeline \
--aoi /input/aoi.geojson \
--t1-start 2026-03-01 \
--t1-end 2026-03-24 \
--t2-start 2026-04-01 \
--t2-end 2026-04-17 \
--bands blue green red \
--tile-size 256 \
--output-dir /output- Assumes small AOIs (< 1 km²): Typically falls within a single Sentinel-2 MGRS tile
- No co-registration: Images are used as-is; assumes Sentinel-2 L2A processing is sufficient
- Cloud filtering: Basic threshold on cloud cover percentage; can be extended with per-pixel SCL masking
See docs/system_design.md for architectural trade-offs and future work.
The content of this repository is released under the MIT LICENSE.
Please see the LICENSE file for more information.