S2DR3 SDK Manual (2025-05-22)

Installation

Requires: Linux, Python 3.12 or later
GPUs tested: NVIDIA T4, P100, V100, A100; 16GB VRAM

1. Install NVIDIA Drivers: https://developer.nvidia.com/cuda-downloads
2. Install GDAL and S2DR3 module wheel as follows:

% apt install -qq pip gdal-bin
% pip -q install https://storage.googleapis.com/0x7ff601307fa5/s2dr3-20250905.1-cp312-cp312-linux_x86_64.whl

3: In infer.py configure default paths for input data (downloaded Sentinel-2 data), output data (super-resolved data), as well as logs to the desired locations on your file system. Note: Sentinel-2 will be automatically downloaded to the <datapath> folder.

parser.add_argument('--datapath', type=str, default='/home/ubuntu/data/datapath', help='Input image or folder')
parser.add_argument('--savepath', type=str, default='/home/ubuntu/data/savepath', help='Output image or folder')
parser.add_argument('--logpath', type=str, default='/home/ubuntu/data/logs', help='Path for logging')

Inference (GPU)

% cd <path-to-the-infer.py>
% python infer.py -f --date <YYYY-MM-DD> --aoi <W S E N>

Examples:

% python infer.py -f --date 2021-08-15 --aoi 30.8520 50.3993 30.8750 50.4131
% python infer.py -f --date 2022-07-18 --aoi -47.8103 -21.2284 -47.7000 -21.1312
% python infer.py -f --date 2022-07-18 --geojson /content/input.geojson
% python infer.py -f --date_range 2021-08-01 2021-09-30 --aoi 30.8520 50.3993 30.8750 50.4131
% python infer.py -f --date_range 2022-10-01 2022-10-30 --mgrs 32TLS
% python infer.py -f --date 2022-10-30 --mgrs 32TLS --NM 3 4 --output_bands B02 B03 B04 B05 B08
% python infer.py -f --date 2022-10-30 --geojson /input-path/aoi.geojson

Main parameters

• --date — Exact date of the Sentinel-2 scene (will fail if the scene does not exist)
• --aoi — Bounding box of the Area-of-interest. Should fit entirely within the Sentinel-2 scene. Should not exceed 10×10 km (100 sq. km) in this dev version.
• --geojson — Input GeoJSON file with collection of AOI polygons
• --mgrs — MGRS tile index
• --iso — 2-digit ISO country code
• --NM — Indices of MGRS subsubtile on the 9×9 grid
• --UV — Indices of MGRS subtile on the 3×3 grid
• --datapath — Path for storing original Sentinel-2 data (default: /content/datapath)
• --savepath — Path to the output super-resolved data (default: /content/savepath)
• --logpath — Log file path (default: /content)
• --bands_out — Spectral bands to generate in the Sentinel-2 format (defaults to BGRN: B02 B03 B04 B08)

Options

• -g, --coreg — Coregistration with ESRI basemap
• -f, --force — Force re-processing of existing products
• -t, --get_tile — Force downloading entire S2 MGRS tile
• -d, --debug — Print debugging information
• -q, --quiet — Run completely quiet
• -s, --simulate — Simulate on a small test tile

AWS Credentials

AWS credentials may be updated by setting the following environmental variables:

$AWS_ACCESS_KEY_ID = *****
$AWS_SECRET_ACCESS_KEY = *************

Inputs

Automatically downloaded Sentinel-2 L2A

Outputs

• 10-channel MS 16-bit uint16 GeoTIFF in original UTM projection
• 4-channel BGRN (B2,B3,B4,B8) 16-bit uint16 GeoTIFF in original UTM projection
• 4-channel RGBA 8-bit GeoTIFF in Web-mercator projection

Output logfile example

{
    "PID": "T33VXL-20220625",
    "mode": "mgrs",
    "date": "20220625",
    "MGRS": "33VXL",
    "ISO": "SE",
    "dimensions": [11430, 11340],
    "bbox": [17.80895790985397, 63.460773781104685, 18.018729265308803, 63.54632126474991],
    "bands": ["B02", "B03", "B04", "B08"],
    "save_path_MS": "/content/savepath/SE/T33VXL/T33VXL-20220625/S2L2Ax10_T33VXL-20220625-00_MS.tif",
    "save_path_TCI": "/content/savepath/SE/T33VXL/T33VXL-20220625/S2L2Ax10_T33VXL-20220625-00_TCI.tif",
    "overlap": 1.0,
    "boa_bias": "false"
}

System requirements

• Ubuntu Linux (tested on Linux-4.19.112+-x86_64-with-Ubuntu-18.04-bionic)
• GPU 16GB (tested on Tesla P100-PCIE-16GB)
• RAM 16GB (at least, preferably 32GB)

Limitations

• Exact date of the Sentinel-2 scene (will fail if the scene does not exist)
• Bounding box of the Area-of-interest should fit entirely within the Sentinel-2 scene.
• AOI should not exceed 10×10 km (1000×1000px, 1MPx, 100 sq. km) in this dev version.
• SentinelHub has a limitation of width/height = 2500px