earthcatalog 0.4.1

earthcatalog is a scalable STAC ingestion library for partitioned GeoParquet catalogs

EarthCatalog

A library for processing STAC items into spatially partitioned GeoParquet catalogs.

Why EarthCatalog?

The Problem: Working with massive collections of geospatial data (satellite imagery, drone surveys, IoT sensors) is challenging because:

• Traditional databases struggle with spatial queries at scale
• Files become too large to process efficiently
• Spatial overlap makes data organization complex
• Updates may require full rebuilds

EarthCatalog transforms STAC items into fast, spatially-partitioned GeoParquet catalogs that:

• Eliminate full table scans - Query only relevant spatial partitions using spatial hive-partition pruning first.
• Scale to terabytes - Each partition is independently manageable
• Support incremental updates - Add new data without rebuilding the whole catalog
• Handle complex geometries - Smart global partitioning for multi-region items

What about Apache Sedona, GeoMesa, Iceberg or PostGIS?

We intentionally avoided introducing a heavier data management layer such as Apache Sedona, Apache Iceberg, or PostGIS. Our use case does not require a catalog or metadata service beyond the files themselves, and keeping the system file-centric significantly reduces complexity. Direct reads from Parquet provide faster access by eliminating additional metadata lookups, while also allowing the system to remain truly serverless, with no long-running services to deploy or maintain.

The dataset already benefits from spatial partitioning, embedded statistics, and automatic schema generation, which cover the primary performance and discovery needs. STAC items are designed for immutable, versioned assets and do not require capabilities like time travel or complex schema evolution that motivate more sophisticated table formats. In this context, a simpler Parquet-based approach is both sufficient and operationally preferable.

Key Features

• Smart Spatial Partitioning: Multiple grid systems (H3, S2, UTM, MGRS, LatLon, Custom geojson)
• Global Partition Schema: Auto-routes large/complex geometries to global partitions
• Temporal Binning: Year, month, or day-based time partitioning
• Distributed Processing: Local multi-threading or Dask distributed
• Incremental Updates: Merge new data with existing partitions

Quick Start

Installation

pip install earthcatalog

# With distributed processing support
pip install "earthcatalog[dask]"

Basic Usage

# Process STAC URLs into a spatial catalog
stac-ingest \
  --input stac_urls.parquet \
  --output ./catalog \
  --scratch ./scratch \
  --workers 4

# Generate schema metadata for efficient querying (enabled by default)
stac-ingest \
  --input stac_urls.parquet \
  --output ./catalog \
  --scratch ./scratch \
  --workers 4

Example: Create Input Data

import pandas as pd

# Sample STAC item URLs
urls = [
    "https://earth-search.aws.element84.com/v1/collections/sentinel-2-l2a/items/S2A_20240101_123456",
    "https://earth-search.aws.element84.com/v1/collections/landsat-8-c2-l2/items/LC08_20240103_345678",
]

df = pd.DataFrame({"url": urls})
df.to_parquet("stac_urls.parquet", index=False)

Configuration Examples

# Use S2 grid with daily partitioning
stac-ingest --input urls.parquet --output ./catalog --scratch ./scratch \
  --grid s2 --grid-resolution 13 --temporal-bin day

# Enable global partitioning with custom thresholds
stac-ingest --input urls.parquet --output ./catalog --scratch ./scratch \
  --global-thresholds-file custom-thresholds.json

# Distributed processing with Dask
stac-ingest --input s3://bucket/urls.parquet --output s3://bucket/catalog \
  --scratch s3://bucket/scratch --processor dask --workers 16

Example: Efficient Spatial Queries

# Traditional approach (slow - scans entire catalog)
import geopandas as gpd
from shapely.geometry import box

roi = box(-122.5, 37.7, -122.0, 38.0)  # San Francisco area
df = gpd.read_parquet("catalog/**/*.parquet")  # Reads EVERYTHING
results = df[df.intersects(roi)]
print(f"Found {len(results)} items (but scanned entire catalog)")

# EarthCatalog approach (fast - scans only relevant partitions)
from earthcatalog.spatial_resolver import spatial_resolver
import duckdb

resolver = spatial_resolver("catalog/catalog_schema.json")
partitions = resolver.resolve_partitions(roi)
paths = resolver.generate_query_paths(partitions)

result = duckdb.sql(f"SELECT * FROM read_parquet({paths})").df()
print(f"Found {len(result)} items (scanned only {len(partitions)} partitions)")

# Remote schema files (S3, GCS, Azure, HTTP) - requires fsspec
resolver = spatial_resolver("s3://my-bucket/catalog_schema.json", "s3://my-bucket/catalog/")
resolver = spatial_resolver("https://example.com/schema.json", "./local-catalog/")

Output Structure

Uses Hive-style temporal partitioning for optimal query pruning in DuckDB, Athena, and Spark:

catalog/
├── {mission}/
│   └── partition=h3/
│       └── level=2/
│           ├── 8928308280fffff/
│           │   └── year=2024/
│           │       ├── month=01/
│           │       │   └── items.parquet  # January 2024 items
│           │       └── month=02/
│           │           └── items.parquet
│           └── global/
│               └── year=2024/
│                   └── month=01/
│                       └── items.parquet  # Large geometries spanning multiple cells
└── catalog_schema.json  # Generated metadata for efficient querying (enabled by default)

Schema Metadata and Efficient Querying

EarthCatalog generates comprehensive metadata about your catalog's partitioning scheme by default:

# Schema is generated by default
stac-ingest --input urls.parquet --output ./catalog --scratch ./scratch

# Use custom schema filename
stac-ingest --input urls.parquet --output ./catalog --scratch ./scratch \
  --schema-filename my_catalog_schema.json

# Disable schema generation
stac-ingest --input urls.parquet --output ./catalog --scratch ./scratch \
  --no-generate-schema

The generated schema includes:

• Grid system details: Type, resolution, cell sizes, coordinate system
• Partition structure: All spatial and temporal partitions created
• Usage examples: DuckDB queries for efficient partition pruning
• Statistics: Item counts, partition counts, processing info

Automatic Global Partition Detection

The resolver intelligently includes the global partition when needed:

# Threshold-based inclusion (queries spanning many cells include global)
large_area = box(-130, 30, -110, 50)  # Multi-state region
partitions = resolver.resolve_partitions(large_area)
# Includes 'global' because query spans > threshold cells

# Geography-based inclusion (continental-scale areas include global)
continental = box(-180, -60, 180, 80)  # Nearly global extent
partitions = resolver.resolve_partitions(continental)
# Includes 'global' because geometry area > large geometry threshold

# Manual control when needed
partitions_no_global = resolver.resolve_partitions(large_area, include_global=False)
partitions_force_global = resolver.resolve_partitions(small_area, include_global=True)

Remote Schema Files

The spatial_resolver() function supports schema files stored in cloud storage or remote locations:

from earthcatalog.spatial_resolver import spatial_resolver

# S3 (requires fsspec[s3])
resolver = spatial_resolver("s3://my-bucket/catalog_schema.json", "s3://my-bucket/catalog/")

# Google Cloud Storage (requires fsspec[gcs])
resolver = spatial_resolver("gs://my-bucket/catalog_schema.json", "gs://my-bucket/catalog/")

# Azure Blob Storage (requires fsspec[azure])
resolver = spatial_resolver("abfs://container/catalog_schema.json", "abfs://container/catalog/")

# HTTP/HTTPS
resolver = spatial_resolver("https://example.com/catalog_schema.json", "./local-catalog/")

# Mixed: Remote schema with local catalog
resolver = spatial_resolver("s3://bucket/schema.json", "/local/catalog/")

Requirements:

• Install fsspec with appropriate extras: pip install fsspec[s3], fsspec[gcs], fsspec[azure]
• The catalog_path parameter is required for remote schema files
• Authentication follows fsspec conventions (AWS credentials, service accounts, etc.)

Grid-Specific Resolution

Key Benefits:

• Automatic Resolution: No need to manually calculate grid intersections
• All Grid Systems: Works with H3, S2, MGRS, UTM, LatLon, and custom GeoJSON
• Configurable Overlap: Control boundary handling and buffer zones
• Performance: Query only relevant partitions instead of full catalog scan
• DuckDB Integration: Generates ready-to-use file path patterns

⚡ Performance Benchmarks

Query Performance Comparison (San Francisco Bay Area query on global dataset):

Metric	Without Pruning	With Spatial Resolution	Improvement
Data Scanned	50GB+	6GB	88.5% reduction
Query Time	45 seconds	5.2 seconds	8.7x faster
Memory Usage	12GB	2.1GB	82% reduction
Files Read	15,000+	1,200	92% fewer files

Grid System Performance (typical regional query):

• H3 Resolution 6: 8-12 cells → ~85-90% data reduction
• MGRS 100km: 1-4 zones → ~95-98% data reduction
• Custom GeoJSON: Variable based on tile design

Documentation

• 📖 Full Documentation - Complete guides and API reference
• 🏁 Quick Start Guide - Get up and running in minutes
• ⚙️ Configuration Guide - All configuration options
• 🌍 Global Partitioning - Handle large/complex geometries
• 🔧 API Reference - Python and CLI documentation

Contributing

# Development setup
git clone https://github.com/betolink/earthcatalog.git
cd earthcatalog
pip install -e ".[dev]"

# Run tests
python -m pytest

# Format and lint
black earthcatalog/ && ruff check earthcatalog/

License

MIT License - see LICENSE file for details.

Acknowledgements

This project was inspired by the need for efficient geospatial data management and builds upon the work of the open-source geospatial community. Special thanks to the developers of STAC, GeoParquet, H3, S2, and other foundational libraries that made this project possible.

Thanks to NSIDC DAAC, NASA ITS_LIVE and NASA Openscapes for supporting open data initiatives that drive innovation in geospatial data processing.