satellite tile vault / v0.9.0 / MIT

GEOVAULT

Polygon in, covering tiles out. 26 free public datasets, fetched once, stored lossless in a handful of Parquet files, read back as numpy.

pip install git+https://github.com/mlhypc/GEOVAULT

01 / usage

Ingest once, read from anywhere.

Give it a polygon and a date range. It finds the scenes, fetches exactly the tiles that cover the polygon, skips everything already stored, and lets you read the result back clipped to the polygon.

shellingest
# polygon in, covering tiles out (true intersection, not bbox)
geovault ingest --dataset s2 --geojson parcel.geojson \
    --start 2024-09-01 --end 2026-09-01
geovault ingest --dataset s1 --geojson parcel.geojson \
    --start 2024-09-01 --end 2026-09-01

# same arguments: report what is still missing, fetch nothing
geovault ingest ... --dry-run
geovault coverage s2 --date 2024-06
pythonread
from geovault.reader import clip, series, acquisitions, mosaic

arr, transform, crs = clip(
    "s2", "parcel.geojson", "B4", "2025-07-14"
)

# pixel time series: one record per overpass
for rec in series("s2", lon, lat, "B4"):
    print(rec["date"], rec["acq"], rec["value"])

# two overpasses on one date: name the one to read
for acq in acquisitions("s1", bbox, "VV", "2026-03-01"):
    arr, transform, crs = mosaic("s1", bbox, "VV",
                                 "2026-03-01", acq=acq)
  1. 01polygon + datesGeoJSON or bbox
  2. 02STAC searchscenes that touch it
  3. 03tile plancovering tiles on a fixed lattice
  4. 04catalog checkfull tiles skipped, partial ones completed
  5. 05range readsonly those bytes of the COG
  6. 06Parquet storelossless GeoTIFF blob per row

Re-running an ingest is safe and cheap: the same command serves the first backfill and every later top-up. An interrupted run keeps what it fetched.

02 / datasets

26 sources, best free one per dataset.

22 need no account at all. The other 4 need a free one (Copernicus CDS or CDSE). Each swatch draws the same patch of ground at that dataset's pixel size (log scale); each strip is one cell per year since 1940.

dataset product pixel 1940today source
Optical 2
s2 Sentinel-2 L2A13 bands, surface reflectance 10 / 20 / 60 m dated archive AWS earth-search (Element 84)no account
landsat Landsat Collection 2 Level-2TM, ETM+, OLI/TIRS: surface reflectance and surface temperature 30 m 1982 to today Microsoft Planetary Computerno account
Radar 1
s1 Sentinel-1 RTCterrain-corrected gamma0, VV / VH 10 m dated archive Microsoft Planetary Computerno account
Vegetation indices 4
modis_vi MODIS MOD13Q1 + MYD13Q116-day NDVI / EVI and reliability 250 m 2000 to today Microsoft Planetary Computerno account
modis_vi_1km MODIS vegetation indices, 1 kmsame composites from the 4x overview, about 9x less download 926 m 2000 to today Microsoft Planetary Computerno account
s3_syn_v10 Sentinel-3 SYNERGY V1010-day NDVI (OLCI + SLSTR), S3A and S3B as separate bands 1/112° (~1 km) 2019 to today Microsoft Planetary Computerno account
cgls_ndvi300 Copernicus Global Land NDVI 300 V310-daily NDVI, uncertainty, observation count, quality flags 1/336° (~300 m) 2019 to today Copernicus Data Spacefree CDSE S3 keys
Elevation 7
glo30 Copernicus GLO-30 DEMsurface model 30 m static layer AWS open datano account
glo90 Copernicus GLO-90 DEMbasemap-grade elevation, about 1/8 the transfer of GLO-30 90 m static layer AWS open datano account
nasadem NASADEMSRTM 2000 reprocessed, surface model, public domain 30 m static layer Microsoft Planetary Computerno account
aw3d30 ALOS World 3D 30 v3.2PRISM optical stereo 2006 to 2011, surface model 30 m static layer Microsoft Planetary Computerno account
gedtm30 GEDTM30 v1.2bare-earth terrain from an ML ensemble: DTM, STD, MASK 30 m static layer OpenGeoHubno account
fabdem FABDEM V1-2 non-commercial onlyGLO-30 with forests and buildings removed 30 m static layer University of Bristolno account
3dep_1m USGS 3DEP 1 mlidar bare earth, United States only 1 m static layer USGS The National Mapno account
Land cover and soil 2
worldcover ESA WorldCover 2021land cover classes 10 m static layer AWS open datano account
soilgrids SoilGrids11 soil properties x 6 depths 250 m static layer ISRICno account
Fire 2
modis_fire MODIS MOD14A1 + MYD14A1daily fire mask, maximum fire radiative power, QA 1 km 2000 to today Microsoft Planetary Computerno account
modis_burned MODIS MCD64A1monthly burned area: day of burn, uncertainty, QA 500 m 2000 to today Microsoft Planetary Computerno account
Weather and climate 8
modis_lst MODIS MOD11A1 + MYD11A1daily land surface temperature, day and night 1 km 2000 to today Microsoft Planetary Computerno account
chirps CHIRPS 2.0daily precipitation, mm/day 0.05° (~5 km) 1981 to today UCSB Climate Hazards Centerno account
era5 ERA5 single levelshourly reanalysis: temperature, dewpoint, precipitation, wind, radiation 0.25° (~28 km) 1940 to today Google ARCO-ERA5no account
era5land ERA5-Landhourly reanalysis, land only 0.1° (~11 km) 1950 to today Copernicus Climate Data Storefree CDS account
era5land_daily ERA5-Land dailydaily mean, min and max in local time 0.1° (~11 km) dated archive Copernicus Climate Data Storefree CDS account
agera5 AgERA5 v2daily agrometeorological indicators 0.1° (~11 km) 1979 to today Copernicus Climate Data Storefree CDS account
pardig_nrt PARDIG-NRT experimental15-minute precipitation rate (Google WeatherNext) 0.1° (~11 km) 2016 to today Google Cloud public bucketno account
pardig_lr PARDIG-LR experimental15-minute precipitation rate, long record 0.1° (~11 km) 1999 to 2023 Google Cloud public bucketno account

03 / design

Few big files, a fixed grid, no remote compute.

layoutdata/
data/
  catalog/<dataset>.parquet                      one row per stored tile
  store/dataset=<d>/<YYYY>/<MM>/r<res>.parquet   sealed month
  store/dataset=<d>/static/r<res>.parquet        static layers

row: band, date, x, y, crs, res, tile_px, west/south/east/north,
     nodata, cloud_pct, scene_id, data (GeoTIFF blob), size_bytes, fetched_at,
     acq, valid_px
key: (band, date, acq, crs, x, y)   acq: the overpass, two a day can share a date

04 / responsible use

It hosts no data.

GEOVAULT is a personal research and education tool. It automates downloads anyone can make from the public sources above and grants no rights to that data. You are the one accessing those services: read each provider's terms before fetching or publishing, attribute the provider, and keep concurrency modest. Some sources are non-commercial only.

Licences and attribution, per dataset →