Skip to content

Latest commit

 

History

History

earth_index

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
21MNT.parquet
21MNT.parquet
 
 
readme.md
readme.md
 
 
requirements.txt
requirements.txt
 
 
sample_metadata.json
sample_metadata.json
 
 
schema.json
schema.json
 
 
validate.py
validate.py
 
 

readme.md

Earth Index Embeddings

Submission Details

Overview

Earth Index is a platform for tile-level geospatial search and classification of satellite imagery. Earth Index works by “pre-indexing” the planet for search using earth observation foundation models.

Use Cases

We've been using embeddings for human-in-the-loop tile-level search (approximate nearest neighbors) and classification (linear or other lightweight models). Some downstream applications include monitoring targets like:

  • Poultry CAFOs
  • Artisinal gold mining
  • Infrastucture development (roads, dams, etc.)
  • Landfills and waste sites

Data

  • URL: [tbd, on source.coop]
  • Documentation: [tbd]
  • Projection: EPSG:4326
  • License: CC-BY

Samples

Validation

  • GeoParquet validation: gpq validate embeddings.parquet
  • Emb sample metadata and parquet metadata validation: python3 validate.py

Columns

Field Name Type Description
geometry geometry Geometry of tile used to generate embeddings
id string Unique identifier for tile
embedding array The vector embedding

File structure

The files originate from MGRS 100km x 100km imagery tile, e.g. 21MNT, which the parquet files inherit. That being said, using a filename as metadata is an easy way to lose context so we make sure that all relevant metadata from the filename also gets embedding in the file itself.

Metadata

The embedding.parquet file is also a valid Geoparquet file, with the expected metadata stored under the geo key. Additionally we have added an emb key for the embeddings metadata - also in JSON format.

The metadata adheres to the json schema provided. For interoperability with STAC this schema references some STAC schema elements, specifically provider, datetime and licensing

EMB Metadata

Field Name Type Description
version string fixed at 0.0.1 (required)
model object<string, object> see Model metadata
providers object<string, object> provider metadata -- see provider
licensing object<string, object> licensing metadata -- see licensing
datetime object<string, object> datetime metadata -- see datetime
source_datasets object<string, object> datasets used to generate embeddings - see Dataset metadata
embedding object<string, object> Embeddings metadata

Model metadata

Field Name Type Description
id string id of model (required)
source string URI for model (required)
version string version of model
family string model family
name string Name of model
description string human readable description of model
config string Configuration of model; includes information needed to generate embeddings (e.g. what layers were extracted)

Dataset metadata

Field Name Type Description
id string Dataset id (required)
name string Dataset used to generate the embeddings
description string Dataset used to generate the embeddings
source string URI for dataset that was used by the embeddings model (required)

Embeddings metadata

Field Name Type Description
dim int Embeddings size
quantization string Description of quantization scheme, if any

Design thoughts / open questions

  • We also duplicate most of this metadata in STAC and have reused STAC metadata definitions to ease interoperability between parquet and STAC .
  • We've opted to put the metadata within the emb key, in the same style as geoparquet. But we're not committed to this and are interested in other opinions.