SEN2SR Tools

This repository provides a list of comprehensive tools and utilities for the usage of the SEN2SR neural network super-resolution model, developed by the ESAOpenSR team. This model specializes in upscaling Sentinel-2's 10m/px images up to a x4 times improvement of 2.5m/px.

This package implements a feature to crop a polygon directly from the SR image, useful for close-up satellite observation of agricultural states.

Installation

1. Clone the repository

   git clone https://github.com/KhaosResearch/sen2sr-tools.git
   cd sen2sr-tools

2. Create a virtual environment and install all dependencies

   python -m venv venv
   source venv/bin/activate   # On Windows use `venv\Scripts\activate`
   pip install -e .

Quickstart example

The overall workflow can be run directly, since it implements a simple example in it:

cd sen2sr-tools
python -m sen2sr_tools.get_sr_image

The example can also be run using commands:

python -m sen2sr_tools get-sr-image --latitude 42.465226 --longitude -2.292699 --start-date 2025-11-01 --end-date 2025-11-15 --geojson-path ./example.geojson

Output will be at sen2sr_out, where the SR and original images can be seen.

Usage in Python

All relevant methods are stored in the get_sr_image module.

Get a SR image

Using get_sr_image runs the full workflow, from date and location data to super resolved image. Default size is the minimal requested 128px. Default bands are Near-Infrared, Red, Blue, Green and the Scene Classification Layer, that contains cloud density information among others. Cropping a polygon from the image is optional, provided the GeoJSON filepath. The final image outputs are in TIF and PNG formats, for usage and visualization respectively.

from sen2sr_tools.get_sr_image import get_sr_image

lat = 37.265840
lng = -4.593406 
start_date = "2025-11-01"
end_date = "2025-11-15"
# bands = ["B08", "B02", "B03", "B04", "SCL"]       # Default bands
# size = 128                                        # Default size
# geometry = None                                   # Default value

sr_filepath = get_sr_image(lat, lng, bands, start_date, end_date)

print(f"SR image successfully downloaded and saved at: {sr_filepath}")

Download CUBO

Running download_sentinel_cubo serves the requested bands' data directly. It uses the date range to get cloudless (<=1%) images. The crs can be provided (i.e, "EPSG:32630"), but it can also be automatically calculated from the lat and lon arguments. If no requested data is found within the date range, it retries with by expanding it backwards in time. The CUBO data is provided as a Dask.array.

from sen2sr_tools.get_sr_image import download_sentinel_cubo

lat = 37.265840
lng = -4.593406 
start_date = "2025-11-01"
end_date = "2025-11-15"
# crs = None                                   # Default CRS
# bands = ["B08", "B02", "B03", "B04", "SCL"]  # Default bands
# size = 128                                   # Default size
# cloud_threshold= 0.01                        # Default threshold density
# max_retries = 3                              # Default retries
# retry_days_shift = 15                        # Default shift

cloudless_cubo_data_array, sample_date = download_sentinel_cubo(lat, lon, start_date, end_date)

print(f"CUBO data summary:\n\n{cloudless_cubo_data_array.coords}")

# If you need the array as NumPy
# cubo_np_array = (cloudless_cubo_data_array.astype("float32") / 10_000).compute()

Crop parcel from image

The crop_parcel_from_tif method crops the polygon from the given TIF file and returns the cropped image output and metadata.

import json
import geopandas as gpd
from sen2sr_tools.get_sr_image import crop_parcel_from_tif

gdf = gpd.read_file(geometry)

raster_path = "path/to/file.tif"
geometry = {
    "type": "Polygon",
    "coordinates": [[
        [-2.29544172799985, 42.46242962302221],[-2.292531924994703, 42.46538599420777],
        [-2.292180580738519, 42.463584097818824], [-2.286733836169323, 42.464620125420105],
        [-2.29544172799985, 42.46242962302221]
    ]], 
    "CRS": "epsg:4258"}

out_image, out_meta = crop_parcel_from_tif(raster_path, geojson_path)

print(f"Polygon successfully cropped!")

Attribution and License

We appreciate the work done by the ESAOpenSR team in upscaling and super-resolving satellite imagery and making SR models accessible through Open Source code. This module is built upon the core model and concepts of the sen2sr Python package.

• Original Repository: https://github.com/ESAOpenSR/SEN2SR.git

The original work is licensed under the CC0 1.0 Universal License.

Trained model:

The model file (model/model.safetensors) is available for download from HugginFace.

• Official SEN2SR HuggingFace model card: https://huggingface.co/tacofoundation/sen2sr.