LGND Embeddings

Name: LGND Embeddings
Creator: Clay

This repository contains Clay embedding datasets created and maintained by lgnd.ai. Licensed under CC BY 4.0.

Product Details

Visibility: Public
Owner: Clay
Created: 1 Jun 2026
Last Updated: 2 Jun 2026

Product Contents

root

monthly-aggregated

model_version=v1.5

README.md

2.58 KB

README

Monthly-Aggregated Embeddings

A deduplicated time series of Clay v1.5 embeddings, selecting one representative embedding per Major TOM grid cell per month. This removes redundancy from overlapping scene coverage at MGRS tile borders, normalizes observation frequency across regions with varying revisit rates, and simplifies time series analysis by providing consistent monthly snapshots.

Partitioning

The data is hive-partitioned by model version, collection, chip size, embedding dimensions, geohash, year, and month:

1monthly-aggregated/
2  model_version=v1.5/
3    collection=sentinel-2-l2a/
4      chip_size=1280m/
5        dims=256/
6          geohash=ab/year=2024/month=06/{hash}.parquet
1monthly-aggregated/
2  model_version=v1.5/
3    collection=sentinel-2-l2a/
4      chip_size=1280m/
5        dims=256/
6          geohash=ab/year=2024/month=06/{hash}.parquet

1aws s3 ls --no-sign-request s3://us-west-2.opendata.source.coop/clay/lgnd-embeddings/monthly-aggregated/model_version=v1.5/collection=sentinel-2-l2a/chip_size=1280m/dims=256/geohash=9y/year=2025/month=01/ --summarize --human-readable
1aws s3 ls --no-sign-request s3://us-west-2.opendata.source.coop/clay/lgnd-embeddings/monthly-aggregated/model_version=v1.5/collection=sentinel-2-l2a/chip_size=1280m/dims=256/geohash=9y/year=2025/month=01/ --summarize --human-readable

Schema

Column	Type	Description
`chips_id`	string	Spatio-temporally unique identifier for each embedding
`cell_id`	string	Spatially unique identifier for the Major TOM cell
`rasters_id`	string	Unique, internal LGND identifier for the source raster
`stac_item_id`	string	STAC item ID of the source Sentinel-2 scene, from Earth Search. Join key to the rasters dataset.
`collection`	string	STAC collection name (e.g. `sentinel-2-l2a`)
`datetime`	timestamp	Acquisition timestamp of the source imagery
`embedding`	list<float32>	256-dimensional Clay v1.5 embedding vector
`bbox`	struct	Bounding box (`xmin`, `ymin`, `xmax`, `ymax`) for geoparquet spatial filtering
`geometry`	binary (WKB)	Polygon geometry of the chip footprint in EPSG:4326

Data Format

GeoParquet 1.1 with geo and covering metadata
Hilbert curve spatial ordering within each partition
256 MB row groups

Usage

1import duckdb
2
3duckdb.sql("""
4    SELECT *
5    FROM read_parquet(
6        's3://us-west-2.opendata.source.coop/clay/lgnd-embeddings/monthly-aggregated/**/*.parquet',
7        hive_partitioning=true
8    )
9    WHERE model_version = 'v1.5'
10      AND collection = 'sentinel-2-l2a'
11      AND year = 2024
12      AND month = 6
13      AND bbox.xmin > -105 AND bbox.xmax < -104
14      AND bbox.ymin > 39 AND bbox.ymax < 41
15    LIMIT 10
16""")
1import duckdb
2
3duckdb.sql("""
4    SELECT *
5    FROM read_parquet(
6        's3://us-west-2.opendata.source.coop/clay/lgnd-embeddings/monthly-aggregated/**/*.parquet',
7        hive_partitioning=true
8    )
9    WHERE model_version = 'v1.5'
10      AND collection = 'sentinel-2-l2a'
11      AND year = 2024
12      AND month = 6
13      AND bbox.xmin > -105 AND bbox.xmax < -104
14      AND bbox.ymin > 39 AND bbox.ymax < 41
15    LIMIT 10
16""")

License

Licensed under CC BY 4.0.