EUMETSAT Meteosat Third Generation (MTG) Preprocessed Dataset

Dataset Summary

This dataset contains preprocessed satellite imagery data from the EUMETSAT Meteosat Third Generation (MTG) satellites. The data has been downloaded, reprojected, and formatted into patches of 128x128 pixels, making it suitable for machine learning applications.

The dataset is generated by a pipeline that performs the following steps:

Data Download: Downloads Level 1C Full Disk High Spectral Imagery (FDHSI) data from the EUMETSAT Data Store.
Reprojection: Reprojects the data to the target EPSG:27704 coordinate system with a resolution of 2000 meters per pixel.
Patch Generation: Extracts 128x128 pixel patches from the reprojected imagery.
Formatting: Saves the patches in Parquet format, with metadata including coordinates and shape.

Data Fields

Each row in the dataset corresponds to a single patch and contains the following fields:

data: A byte string representing the patch's pixel data. This can be loaded into a NumPy array using the shape and dtype fields.
shape: The shape of the patch data, typically (6, C, 128, 128), where C is the number of channels.
dtype: The data type of the pixel values (e.g., float16).
epsg: The EPSG code of the coordinate system (e.g., 27704).
x_coord: The x-coordinate of the top-left corner of the patch in the specified EPSG.
y_coord: The y-coordinate of the top-left corner of the patch in the specified EPSG.
patch_x_idx: The index of the patch along the x-axis.
patch_y_idx: The index of the patch along the y-axis.

Data Splits

The dataset is not split into predefined training, validation, or test sets. You can create your own splits based on your specific needs. The data is organized by the date of the satellite imagery, which can be used to create time-based splits.

Usage

To use this dataset, you can load the Parquet files using a library like pyarrow or pandas. Here's an example of how to load a single Parquet file and access the patch data:

import pandas as pd
import numpy as np

# Load a Parquet file
df = pd.read_parquet("path/to/your/file.parquet")

# Get the first patch
patch_info = df.iloc[0]

# Reshape the data into a NumPy array
patch_data = np.frombuffer(patch_info["data"], dtype=patch_info["dtype"]).reshape(patch_info["shape"])

print(f"Patch shape: {patch_data.shape}")
print(f"Patch coordinates: ({patch_info['x_coord']}, {patch_info['y_coord']})")

Source Data

The source data is the MTG FCI Level 1c Full Disk High Spectral Imagery (FDHSI) product from EUMETSAT.

Preprocessing

The preprocessing steps include:

Selecting the closest available satellite image to a target timestamp.
Extracting specified channels.
Reprojecting the data to EPSG:27704.
Combining visible/near-infrared (vis_nir) and infrared/water vapor (ir_wv) channels.
Creating 128x128 patches from the reprojected data.

Limitations

The dataset is limited to the geographical area covered by the reprojection, which is focused on Europe.
The temporal resolution of the data is determined by the availability of the source imagery from EUMETSAT.

License

The code used to generate this dataset is licensed under the MIT License. The source data from EUMETSAT has its own terms and conditions, which you should consult before using the data.