earthfetch 0.6.1

Download USGS 3DEP DEMs and Sentinel-2 L2A imagery with zero API keys

earthfetch

Analysis-ready Earth data in one line. Zero API keys. Zero accounts.

📖 Docs: https://ethan-m2024.github.io/earthfetch/

import earthfetch as ef

# Cloud-free composite of any place on Earth — geocoded, cloud-masked,
# mosaicked across scene boundaries, reflectance-scaled, in your UTM zone:
rgb = ef.composite("Moab, Utah", bands=["B04", "B03", "B02"],
                   start="2026-05-01", end="2026-06-15")
ef.preview(rgb, "moab.png")

# NDVI for a farm polygon, clipped to its boundary:
ds = ef.composite("field.geojson", bands=["B08", "B04"],
                  start="2026-05-01", end="2026-06-01")
ndvi = ef.ndvi(ds)

# Terrain anywhere (Alps -> Copernicus DEM, auto-UTM):
terr = ef.terrain((6.85, 45.82, 6.90, 45.87))   # dem, slope, aspect, hillshade
ef.to_cog(terr.hillshade, "hillshade.tif")

One scene replaces an afternoon: no EarthExplorer queues, no Copernicus tokens, no manual SCL masking, no tile mosaicking, no gdalwarp. Only the AOI window travels over the network — everything is read from Cloud-Optimized GeoTIFFs with HTTP range requests.

AOI: pass anything

Every function takes bboxes, GeoJSON dicts, .geojson files, shapely geometries, or place names:

ef.composite((-111.9, 40.7, -111.8, 40.8), ...)   # bbox tuple
ef.composite("Yosemite National Park", ...)        # geocoded (Nominatim)
ef.composite("watershed.geojson", ...)             # file; clips to polygon
ef.composite(shapely_polygon, ...)                 # __geo_interface__

crs="utm" (default in composite/terrain/load_naip) picks the right UTM zone. Explicit polygons clip results to their boundary; geocoded place names return the full rectangle (pass clip=True to cut to the boundary).

Data sources (all free, no auth)

Source	What	Coverage	Resolutions
USGS 3DEP (The National Map)	DEM	United States	1 m, 10 m, 30 m, 5 m (AK)
Copernicus GLO-30 (AWS)	DEM	Global	30 m
Sentinel-2 L2A (Earth Search / AWS)	Multispectral imagery	Global	10 / 20 / 60 m
NAIP (Planetary Computer)	Aerial photography (RGBN)	United States	0.6-1 m

Looking for "Google Earth"-quality imagery? That's NAIP: actual aerial photos where you can see individual cars and trees.

img = ef.load_naip("Moab, Utah", res=1)          # 1 m RGB mosaic
nir = ef.load_naip(aoi, bands=["N","R","G"])     # false-color infrared

Install

pip install earthfetch              # search + download only (requests)
pip install "earthfetch[xarray]"    # + load_dem / load_sentinel2 / stack

Composites, indices, terrain

# method: "median" (robust), "mean", "first" (fastest)
da = ef.composite(aoi, bands=["B04","B03","B02"], start=..., end=...,
                  method="median", mask_clouds=True, max_scenes=8)

ef.ndvi(ds); ef.ndwi(ds); ef.nbr(ds); ef.evi(ds); ef.savi(ds)

terr = ef.terrain(aoi, products=["dem","slope","aspect","hillshade"],
                  resolution="10m")   # USGS in US, Copernicus elsewhere

ef.to_geotiff(obj, "out.tif"); ef.to_cog(obj, "out.tif")
ef.preview(obj, "look.png")           # percentile-stretched quicklook

Array API (pipelines)

import earthfetch as ef

bbox = (-111.90, 40.70, -111.85, 40.75)  # (min_lon, min_lat, max_lon, max_lat)

# DEM as xarray.DataArray — any CRS, any pixel size
dem = ef.load_dem(bbox, resolution="10m", crs="EPSG:32612", res=10)

# Outside the US? Copernicus GLO-30 is used automatically (source="auto"),
# or ask for it explicitly:
alps = ef.load_dem((6.85, 45.82, 6.88, 45.85), crs="EPSG:32632", source="copernicus")

# Sentinel-2 bands, clearest scene in a date range
s2 = ef.load_sentinel2(bbox, bands=["B04", "B08"], crs="EPSG:32612",
                       start="2026-05-01", end="2026-06-01", max_cloud=20)

# Everything aligned on one grid (ML-ready xarray.Dataset)
ds = ef.stack(bbox, crs="EPSG:32612", res=30, bands=["B04", "B08"],
              start="2026-05-01", end="2026-06-01")

All arrays are float32 with NaN nodata and carry crs, transform, and source URLs in attrs. Bands accept ESA ids (B02–B12, B8A, SCL) or Earth Search asset keys (red, nir, ...).

Search + download API (files)

tiles = ef.search_dem(bbox, resolution="10m")           # metadata only
paths = ef.download_dem(bbox, resolution="10m", out_dir="dem")

scenes = ef.search_sentinel2(bbox, "2026-05-01", "2026-06-01", max_cloud=15)
files = ef.download_sentinel2(scenes[0], bands=["B04", "B08"], out_dir="s2")

from earthfetch import clip_reproject                    # needs [raster]
clip_reproject(paths, bbox, "EPSG:32612", "dem_utm.tif")

Downloads default to a per-user cache ($EARTHFETCH_CACHE to override), stream to .part files, verify byte counts, skip files already on disk, and run in parallel. The library never prints — it logs to the earthfetch logger; pass progress=earthfetch.utils.print_progress for a progress bar in scripts.

CLI

earthfetch dem --bbox -111.9 40.7 -111.8 40.8 --search-only --json
earthfetch dem --bbox -111.9 40.7 -111.8 40.8 --resolution 10m --out dem/
earthfetch s2  --bbox -111.9 40.7 -111.8 40.8 \
    --start 2026-05-01 --end 2026-06-01 --bands B04,B08 --out s2/

--json emits machine-readable search results; -v logs library activity.

Errors

Everything raises a subclass of earthfetch.EarthfetchError: TileNotFoundError (bbox outside coverage), NoScenesError (no clear scenes — widen dates or raise max_cloud), DownloadError, BandNotFoundError, MissingDependencyError.

License

MIT