quasara-io/Quasara-MajorTOM-DE-SigLIP

Dataset Card for Dataset Name

We downloaded satellite images from Major-TOM, provided by the European Space Agency, filtered for Germany, and used our vectorisation engine to extract vector embeddings with one of the latest embedding model.

Datasource Details

	Value
Datasource	Major-TOM/Core-S2L2A
Region	box(5.98865807458, 47.3024876979, 15.0169958839, 54.983104153) (Covers whole of Germany)
Date Range	('2020-01-01', '2025-01-01')
Cloud Cover	(0, 10)
No Data	(0.0, 0.0)

Organisation: https://huggingface.co./Major-TOM

Base Dataset: https://huggingface.co./datasets/Major-TOM/Core-S2L2A

Metadata.paraquet File

This dataset shows the relationship between our embeddings/vectors and Major TOM images for fast linking to other Major TOM datasets.

Embedding.dat

This dataset entails the vector embeddings calculated by Quasara.

What we did on our side is:

a) download the Major-TOM dataset and filter it for images showing Germany;

b) vectorising the entire Major-TOM image data;

c) using the OPENCLIP_SIGLIP_400M and our scalable Vectorisation Engine 'Synapsis' for embedding extraction.

There was no pre-training, labelling or finetuning happening to prepare the vector embeddings of the Major TOM dataset.

Uses

Potential use cases for the dataset we came up with is the data exploration of the data using text prompts, image prompts, unsupervised clustering of images, building a RAG or even building a chat bot on top of the vector embeddings. What can you come up with to build?

MajorTOM-DE Dataset

The MajorTOM-DE dataset also provides embeddings derived from high-resolution satellite images of the Germany region, generated using the OpenCLIP SigLIP model. These embeddings, extracted from images covering a range of geographic coordinates across Germany, provide a powerful tool for various applications.

Dataset Information

• Coordinates Info: The embeddings cover a range of geographic coordinates across the Germany region.
• Related Dataset: The MajorTOM-DE dataset is closely related to the original S2L2A dataset.

Features

The MajorTOM-DE dataset leverages CLIP's ability to relate textual descriptions to visual data, enabling more intuitive searches and analysis. This allows users to search among images using text-based queries effectively.

Applications

The MajorTOM-DE dataset can be utilized for various applications, including:

• Monitoring Changes in Land Use and Land Cover:

  • Track deforestation
  • Observe urban expansion
  • Monitor water body dynamics
  • Finding countless objects from airports, golf courses to wind farms

• Precision Agriculture:

  • Analyze crop health
  • Predict yields
  • Plan harvests

• Climate Research:

  • Study climate patterns
  • Monitor changes and impacts on regional and local levels

Dataset Structure

Metadata.paraquet

Column	Explanation
grid_cell	Coordinates in the Major TOM grid, enabling fast linking to other Major TOM datasets.
grid_row_u	Row identifier in the Major TOM grid for linking purposes.
grid_row_r	Another row identifier in the Major TOM grid for linking purposes.
centre_lat	Latitude of the center of the image portion for which embedding has been computed.
centre_lon	Longitude of the center of the image portion for which embedding has been computed.
timestamp	Date and time of the original product in the %Y%m%dT%H%M%S format.
dat_row	Row number in the .dat file associated with the data entry.
unique_id	Unique identifier combining grid_cell, timestamp, and possibly other parameters (e.g., parquet).

Embedding.dat

Column	Explanation
ID	ID of the image/image part for which the embedding was calculated.
dat_row	Row in the .dat file that can be used to match the embeddings to the MajorTOM datasets via metadata.paraquet dataset.
image_type	Each image is split into 70 segments and vectorized.
coordinates	Coordinates in the image that define the segment that was vectorized. Full images have no coordinates.
split_configs	Quasara auto split configuration
embeddings	Vectors calculated from the image/image segment.