sen2sr 0.3.0

A Python package to super-resolve Sentinel-2 satellite imagery up to 2.5 meters.

SEN2SR

A Python package for enhancing the spatial resolution of Sentinel-2 satellite images up to 2.5 meters 🚀

---

GitHub: https://github.com/ESAOpenSR/sen2sr 🌐

PyPI: https://pypi.org/project/sen2sr/ 🛠️

---

Table of Contents

• Overview
• Installation ⚙️
• Predict 10m and 20m bands
• Predict only RGBNIR bands
• Predict on large images
• Estimate the Local Attention Map of the model

Overview

sen2sr is a Python package designed to enhance the spatial resolution of Sentinel-2 satellite images to 2.5 meters using a set of neural network models.

Installation

Install the SEN2SRLite version using pip:

pip install mlstac sen2sr

For the full version, which use Mamba arquitecture, install as follows:

pip install mlstac sen2sr[full]

From 10m and 20m S2 bands to 2.5m

This example demonstrates the use of the SEN2SRLite model to enhance the spatial resolution of Sentinel-2 imagery. A Sentinel-2 L2A data cube is created over a specified region and time range using the cubo library, including both 10 m and 20 m bands. The pretrained model, downloaded via mlstac, takes a single normalized sample as input and predicts a HR output. The visualization compares the original RGB composite to the super-resolved result.

import matplotlib.pyplot as plt
import numpy as np
import torch
import cubo

import sen2sr
import mlstac

# Download the model
mlstac.download(
  file="https://huggingface.co/tacofoundation/sen2sr/resolve/main/SEN2SRLite/main/mlm.json",
  output_dir="model/SEN2SRLite",
)

# Load the model
model = mlstac.load("model/SEN2SRLite").compiled_model()
model = model.to(device)


# Create a Sentinel-2 L2A data cube for a specific location and date range
da = cubo.create(
    lat=39.49152740347753,
    lon=-0.4308725142800361,
    collection="sentinel-2-l2a",
    bands=["B02", "B03", "B04", "B05", "B06", "B07", "B08", "B8A", "B11", "B12"],
    start_date="2023-01-01",
    end_date="2023-12-31",
    edge_size=64,
    resolution=10
)


# Prepare the data to be used in the model, select just one sample 
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
original_s2_numpy = (da[11].compute().to_numpy() / 10_000).astype("float32")
X = torch.from_numpy(original_s2_numpy).float().to(device)

# Apply model
superX = model(X[None]).squeeze(0)

# Visualize the results
fig, ax = plt.subplots(1, 2, figsize=(10, 5))
ax[0].imshow(X[[2, 1, 0]].permute(1, 2, 0).cpu().numpy()*4)
ax[0].set_title("Original S2")
ax[1].imshow(superX[[2, 1, 0]].permute(1, 2, 0).cpu().numpy()*4)
ax[1].set_title("Enhanced Resolution S2")
plt.show()

From 10m S2 bands to 2.5m

This example demonstrates the use of the SEN2SRLite NonReference_RGBN_x4 model variant to enhance the spatial resolution of only the 10 m Sentinel-2 bands: red (B04), green (B03), blue (B02), and near-infrared (B08). A Sentinel-2 L2A data cube is created using the cubo library for a specific location and date range. The input is normalized and passed to a pretrained non-reference model optimized for RGB+NIR inputs.

import matplotlib.pyplot as plt
import numpy as np
import mlstac
import torch
import cubo

#mlstac.download(
#  file="https://huggingface.co/tacofoundation/sen2sr/resolve/main/SEN2SRLite/NonReference_RGBN_x4/mlm.json",
#  output_dir="model/SEN2SRLite_RGBN",
#)

# Create a Sentinel-2 L2A data cube for a specific location and date range
da = cubo.create(
    lat=39.49152740347753,
    lon=-0.4308725142800361,
    collection="sentinel-2-l2a",
    bands=["B04", "B03", "B02", "B08"],
    start_date="2023-01-01",
    end_date="2023-12-31",
    edge_size=128,
    resolution=10
)

# Prepare the data to be used in the model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
original_s2_numpy = (da[11].compute().to_numpy() / 10_000).astype("float32")
X = torch.from_numpy(original_s2_numpy).float().to(device)

# Load the model
model = mlstac.load("model/SEN2SRLite_RGBN").compiled_model()
model = model.to(device)

# Apply model
superX = model(X[None]).squeeze()

From 20m S2 bands to 10m

This example demonstrates the use of the SEN2SRLite Reference_RSWIR_x2 model variant to enhance the spatial resolution of the 20 m Sentinel-2 bands: red-edge (B05, B06, B07), shortwave infrared (B11, B12), and near-infrared (B8A) to 10 m.

import matplotlib.pyplot as plt
import numpy as np
import torch
import cubo

import sen2sr
import mlstac

# Create a Sentinel-2 L2A data cube for a specific location and date range
da = cubo.create(
    lat=39.49152740347753,
    lon=-0.4308725142800361,
    collection="sentinel-2-l2a",
    bands=["B04", "B03", "B02", "B08"],
    start_date="2023-01-01",
    end_date="2023-12-31",
    edge_size=64,
    resolution=10
)


# Prepare the data to be used in the model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
original_s2_numpy = (da[11].compute().to_numpy() / 10_000).astype("float32")
X = torch.from_numpy(original_s2_numpy).float().to(device)


# Download the model
#mlstac.download(
#  file="https://huggingface.co/tacofoundation/sen2sr/resolve/main/SEN2SRLite/NonReference_RGBN_x4/mlm.json",
#  output_dir="model/SEN2SRLite_RGBN",
#)

# Load the model
model = mlstac.load("model/SEN2SRLite_RGBN").compiled_model()
model = model.to(device)

# Apply model
superX = model(X[None]).squeeze(0)

Predict on large images

This example demonstrates the use of SEN2SRLite NonReference_RGBN_x4 for super-resolving large Sentinel-2 RGB+NIR images by chunking the input into smaller overlapping tiles. Although the model is trained to operate on fixed-size 128×128 patches, the sen2sr.predict_large utility automatically segments larger inputs into these tiles, applies the model to each tile independently, and then reconstructs the full image. An overlap margin (e.g., 32 pixels) is introduced between tiles to minimize edge artifacts and ensure continuity across tile boundaries.

import matplotlib.pyplot as plt
import numpy as np
import torch
import cubo

import sen2sr
import mlstac

# Create a Sentinel-2 L2A data cube for a specific location and date range
da = cubo.create(
    lat=39.49152740347753,
    lon=-0.4308725142800361,
    collection="sentinel-2-l2a",
    bands=["B04", "B03", "B02", "B08"],
    start_date="2023-01-01",
    end_date="2023-12-31",
    edge_size=1024,
    resolution=10
)


# Prepare the data to be used in the model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
original_s2_numpy = (da[11].compute().to_numpy() / 10_000).astype("float32")
X = torch.from_numpy(original_s2_numpy).float().to(device)

# Load the model
#mlstac.download(
#  file="https://huggingface.co/tacofoundation/sen2sr/resolve/main/SEN2SRLite/NonReference_RGBN_x4/mlm.json",
#  output_dir="model/SEN2SRLite_RGBN",
#)
model = mlstac.load("model/SEN2SRLite_RGBN").compiled_model()

# Apply model
superX = sen2sr.predict_large(
    model=model,
    X=X, # The input tensor
    overlap=32, # The overlap between the patches
)

Estimate the Local Attention Map of the model 📊

This example computes the Local Attention Map (LAM) to analyze the model's spatial sensitivity and robustness. The input image is scanned with a sliding window, and the model's attention is estimated across multiple upscaling factors. The resulting KDE map highlights regions where the model focuses more strongly, while the robustness vector quantifies the model's stability to spatial perturbations.

import matplotlib.pyplot as plt
import numpy as np
import torch
import cubo

import sen2sr
import mlstac

# Create a Sentinel-2 L2A data cube for a specific location and date range
da = cubo.create(
    lat=39.49152740347753,
    lon=-0.4308725142800361,
    collection="sentinel-2-l2a",
    bands=["B04", "B03", "B02", "B08"],
    start_date="2023-01-01",
    end_date="2023-12-31",
    edge_size=128,
    resolution=10
)


# Prepare the data to be used in the model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
original_s2_numpy = (da[11].compute().to_numpy() / 10_000).astype("float32")
X = torch.from_numpy(original_s2_numpy).float().to(device)


kde_map, complexity_metric, robustness_metric, robustness_vector = sen2sr.lam(
    X=X.cpu(), # The input tensor
    model=model.srx4, # The SR model
    h=240, # The height of the window
    w=240, # The width of the window
    window=128, # The window size
    scales = ["2x", "3x", "4x", "5x", "6x"]
)

# Visualize the results
plt.imshow(kde_map)
plt.title("Kernel Density Estimation")
plt.show()

plt.plot(robustness_vector)
plt.title("Robustness Vector")
plt.show()