EarthForge

Working with cloud-native geospatial data means juggling gdalinfo for COGs, stac-client for discovery, geopandas for GeoParquet, xarray for Zarr, and a collection of one-off scripts to glue them together. Each tool has its own CLI conventions, its own output format, and its own assumptions about how you authenticate to cloud storage.

EarthForge is a single composable toolkit that unifies these workflows. One CLI. One config system. One output contract. Every command works locally, against S3, GCS, or Azure — and every command produces both human-readable tables and machine-parseable JSON.

# Inspect any cloud-native geospatial file — format auto-detected
earthforge info s3://bucket/image.tif
earthforge info buildings.parquet
earthforge info climate.zarr

# Search STAC catalogs
earthforge stac search sentinel-2-l2a --bbox -85,37,-84,38 --datetime 2025-06/2025-09

# Generate a quicklook preview from a remote COG without downloading it
earthforge raster preview s3://bucket/scene.tif -o preview.png

# Convert legacy formats to cloud-native
earthforge vector convert buildings.shp --to geoparquet
earthforge raster convert image.tif --to cog

# Query GeoParquet with spatial predicate pushdown
earthforge vector query buildings.parquet --bbox -85,37,-84,38

# Inspect and slice Zarr datacubes
earthforge cube info s3://era5-pds/zarr/2025/01/data/air_temperature_at_2_metres.zarr
earthforge cube slice s3://era5-pds/zarr/ --var t2m --bbox -85,37,-84,38 --time 2025-06/2025-06 -o ky_june.zarr

# Pipe structured JSON into other tools
earthforge stac search sentinel-2-l2a -o json | jq '.items[].assets.B04.href'

What EarthForge Is

EarthForge is a library-first, CLI-first developer toolkit. Install it as a Python library and call functions directly, or use the CLI from shell scripts and pipelines. Every CLI command is a thin wrapper around a library function, so anything you can do from the terminal you can also do from Python, a Jupyter notebook, or a pipeline runner.

from earthforge.raster.info import inspect_raster
from earthforge.stac.search import search_catalog

# Library usage — same logic as the CLI, no subprocess needed
items = await search_catalog("sentinel-2-l2a", bbox=(-85, 37, -84, 38))
metadata = await inspect_raster("s3://bucket/scene.tif")

Real-World Output

The samples below are actual outputs from EarthForge commands run against public geospatial data. Sample files live in data/samples/.

KyFromAbove 3-inch Orthoimagery — fetched thumbnail

earthforge stac fetch \
  https://spved5ihrl.execute-api.us-west-2.amazonaws.com/collections/orthos-phase3/items/N097E305_2024_Season1_3IN_cog \
  --assets thumbnail --output-dir data/kyfromabove_fetch
# → 78,026 bytes in 2.34s

3-inch orthoimagery, KyFromAbove Phase 3 (2024). Public domain. Full COG available at kyfromabove.s3.us-west-2.amazonaws.com.

---

Sentinel-2 STAC Search — --output json

earthforge stac search sentinel-2-l2a \
  --bbox -85,37,-84,38 --datetime 2025-06/2025-09 --max-items 5 \
  --output json

{
  "collection": "sentinel-2-l2a",
  "matched": 47,
  "returned": 5,
  "elapsed_seconds": 1.243,
  "items": [
    {
      "id": "S2A_18SYJ_20250914_0_L2A",
      "datetime": "2025-09-14T16:28:43Z",
      "properties": { "eo:cloud_cover": 4.2, "platform": "sentinel-2a" }
    }
  ]
}

Full sample: data/samples/stac_search.json

---

COG Metadata — earthforge raster info

earthforge raster info \
  https://sentinel-cogs.s3.us-west-2.amazonaws.com/.../B04.tif \
  --output json

{
  "format": "COG",
  "width": 10980, "height": 10980,
  "crs": "EPSG:32618",
  "is_tiled": true, "tile_width": 512, "tile_height": 512,
  "overview_count": 6,
  "compression": "deflate"
}

Full sample: data/samples/raster_info.json

---

GeoParquet Metadata — earthforge vector info

earthforge vector info ky_wildlife_management_areas.parquet --output json

{
  "format": "geoparquet",
  "row_count": 83,
  "geometry_types": ["MultiPolygon"],
  "crs": "EPSG:4326",
  "bbox": [-89.57, 36.49, -81.96, 39.15],
  "compression": "SNAPPY",
  "file_size_bytes": 142863
}

Full sample: data/samples/vector_info.json

---

Output Gallery

All images below were generated from real-world data using EarthForge example scripts. No synthetic or simulated data. Each output includes a .txt sidecar with alt text, data provenance, and generation metadata. See examples/outputs/ for full details.

Grand Canyon — DEM with Hillshade + Cross-Section

SRTM 30m elevation data via OpenTopography API. Shows 1,844m of relief from river to rim with elevation cross-section profile.

Swiss Alps — Matterhorn/Zermatt Elevation Analysis

Copernicus DEM 30m via OpenTopography. Elevations from 1,868m (valley) to 4,330m (peaks) with statistics sidebar.

Colorado Front Range — Sentinel-2 NDVI

Vegetation gradient from plains to alpine tundra, showing elevation-driven ecology. BrBG colorblind-safe diverging palette.

Netherlands — Urban/Water/Vegetation NDVI

Sentinel-2 scene over Rotterdam/Delft showing water (NDVI < 0), urban (low NDVI), and agricultural areas (high NDVI).

Amazon Rainforest — Tropical NDVI

Sentinel-2 scene near Manaus, Brazil showing dense tropical forest canopy with uniformly high NDVI.

Copernicus DEM — Elevation Statistics + Histogram

Raster statistics computed from a Copernicus DEM 30m tile with elevation distribution histogram. Viridis colorblind-safe palette.

Yellowstone — Landsat STAC Search Footprints

Landsat Collection 2 Level-2 scene footprints from Earth Search, color-coded by cloud cover percentage.

Yosemite — Multi-Collection STAC Query

Two-panel figure querying both Sentinel-2 scenes and Copernicus DEM tiles from a single STAC API.

STAC-to-NDVI Pipeline

Complete pipeline workflow: STAC search, range-read Sentinel-2 bands, NDVI computation via safe expression evaluator, rendered output with pipeline summary.

Format Detection Matrix

EarthForge's three-stage format detection chain (magic bytes, extension, content inspection) tested across 12 geospatial file formats.

---

What EarthForge Is Not

EarthForge is not a platform. It does not include a web server, a tile cache, a database, an ML pipeline, or a Kubernetes deployment. It is not a replacement for QGIS, ArcGIS, or Google Earth Engine. It does not try to be everything — it is a focused set of tools that integrate with existing workflows via structured output, stdin/stdout piping, and Python imports.

If you need a tile server, use TiTiler. If you need a STAC API, use stac-fastapi. If you need a geospatial database, use PostGIS. EarthForge is the CLI toolkit you reach for alongside those tools, not instead of them.

Install

# Full toolkit
pip install earthforge[all]

# Just what you need
pip install earthforge[stac]        # STAC discovery only
pip install earthforge[raster]      # COG operations only
pip install earthforge[vector]      # GeoParquet operations only
pip install earthforge[cube]        # Zarr datacube operations only
pip install earthforge[cli]         # CLI framework only

Cloud Storage

EarthForge uses named profiles for cloud storage authentication, similar to AWS CLI profiles:

# Initialize config
earthforge config init

# Search with a specific profile
earthforge stac search sentinel-2-l2a --profile planetary

Profiles are defined in ~/.earthforge/config.toml:

[profiles.default]
stac_api = "https://earth-search.aws.element84.com/v1"
storage = "s3"

[profiles.planetary]
stac_api = "https://planetarycomputer.microsoft.com/api/stac/v1"
storage = "azure"

Architecture

EarthForge is built as a monorepo with independently installable workspace packages. The architecture is documented in detail — not as an afterthought, but as the foundation the implementation is built on.

• ARCHITECTURE.md — System design, dependency graph, module interfaces
• ai-dev/decisions/ — Architectural decision records with alternatives considered and tradeoffs acknowledged
• ai-dev/spec.md — Requirements and acceptance criteria per milestone

Key architectural decisions:

Decision	Record	Summary
Monorepo structure	DL-001	Single repo with Hatch workspace packages, not 15 separate repos
Async-first I/O	DL-002	All network I/O is async via httpx; sync wrappers for convenience
obstore for storage	DL-003	Rust-backed S3/GCS/Azure abstraction over fsspec
Rust extension boundary	DL-005	Rust for format detection and range reads; Python for everything else
Engineering credibility	DL-006	Nothing ships empty; decisions before code; scope boundaries enforced
promptfoo evaluation	DL-007	Agent prompts and guardrails regression-tested in CI via promptfoo

Formats

Format	Support	Operations
COG (Cloud Optimized GeoTIFF)	Full	info, validate, convert, preview, band math, tile
GeoParquet	Full	info, validate, convert, query, clip, tile
Zarr	Full	info, validate, convert, slice, stats
FlatGeobuf	Read/Write	info, validate, convert
COPC (Cloud Optimized Point Cloud)	Info	info
STAC (SpatioTemporal Asset Catalog)	Full	search, info, validate, fetch, publish

Contributing

See CONTRIBUTING.md. EarthForge has specific engineering standards — please read the contribution guide before opening a PR.

Code of Conduct

SeeCODE_OF_CONDUCT

Security

See SECURITY

License

GNU General Public License v3.0. See LICENSE.