Cloud-Optimized GeoTIFF (COG) and STAC Catalog Architecture
Eliminate multi-gigabyte scene downloads by pairing internally tiled Cloud-Optimized GeoTIFF rasters with STAC metadata APIs.
How Cloud-Optimized GeoTIFF Enables HTTP Range Streaming
Traditional GeoTIFFs store scanlines sequentially, forcing users to download a 2 GB scene to view one 5 km site. A Cloud-Optimized GeoTIFF (COG) remains fully compatible with desktop GIS while reorganizing internal bytes for partial HTTP streaming.
Valid COGs place Image File Directory (IFD) headers and tile offsets at offset 0, followed by overview pyramids and 256x256 or 512x512 internal tiles. Clients issue HTTP Range requests to fetch only tiles intersecting an Area of Interest.
- Header IFD bytes sit at offset 0 so one small read retrieves the tile index.
- 256x256 or 512x512 internal tiles avoid scanning full-width image rows.
- Embedded overviews (2x, 4x, 8x) accelerate zoomed-out map rendering.
- DEFLATE or ZSTD lossless compression lowers cloud egress without data loss.
STAC Specification: Catalogs, Collections, and Items
The SpatioTemporal Asset Catalog (STAC) specification standardizes Earth observation scene JSON metadata across providers. Each STAC Item stores a GeoJSON footprint, timestamp, sensor properties, and signed URLs linking directly to COG band assets.
STAC Collections group related Items with license and band schemas, while STAC APIs support CQL2 filtering so applications query cloud cover, off-nadir angle, and EPSG projections in milliseconds.
- Extensions (`eo`, `sar`, `view`, `proj`) normalize cloud, incidence, and EPSG fields.
- CQL2 JSON filters push predicate evaluation to the server.
- Earth SAT Data exposes unified STAC-aligned metadata across 700+ sensors.
Windowed Cloud Processing with GDAL and Rasterio
Using GDAL `/vsicurl/` or Python `rasterio` windowed reads, pipelines stream clipped raster chips straight into machine learning or ArcGIS Pro without local staging. Separate single-band COGs cut transfer bytes when computing narrow indices like NDVI, whereas pixel-interleaved COGs speed up RGB rendering.
Validating COG Compliance Before Production Ingestion
Files ending in `.tif` without header-first IFDs or internal overviews trigger silent full-file downloads. Running `rio cogeo validate` at ingestion enforces predictable latency and cloud egress costs.
Comparison of Raster Delivery Formats for Cloud Geospatial Workflows
| Format | Partial AOI Read | Desktop GIS Support | Best Cloud Engineering Use Case |
|---|---|---|---|
| Cloud-Optimized GeoTIFF (COG) | Yes (HTTP Range) | Native (QGIS, ArcGIS, GDAL) | Dynamic tiling, AOI clipping, ML chip extraction |
| Legacy Striped GeoTIFF | Poor (full row scans) | Native | Local archival only; avoid for cloud streaming |
| JPEG2000 (JP2) | Limited / CPU heavy | Requires OpenJPEG | Compact archive distribution; slower decode |
| Zarr / NetCDF-4 | Chunked N-D arrays | Plugin / Xarray | Multi-year datacubes and climate time series |
| WMTS / XYZ Raster Tiles | Pre-rendered only | Basemap layers | Fast visual basemaps; not raw 16-bit analytics |
COG combines 16-bit quantitative pixel fidelity with HTTP Range partial reads and universal GIS compatibility.
Worked Illustration: Estimating COG HTTP Range Tile Reads and Egress
A 4-band uint16 COG measures 16,384 x 16,384 pixels with 256 x 256 internal tiles (~2:1 DEFLATE compression). A client reads an aligned 512 x 512 AOI window.
- Full scene uncompressed size: 16,384 x 16,384 x 4 x 2 bytes = 2.0 GiB (~1.0 GiB compressed).
- Single 256 x 256 tile size: 256 x 256 x 4 x 2 = 512 KiB (~256 KiB compressed).
- Tiles intersecting 512 x 512 AOI: (512 / 256) x (512 / 256) = 4 tiles.
- Total transferred: ~32 KiB header + (4 x 256 KiB) = ~1.03 MiB instead of 1,024 MiB.
Calculated Result & Takeaway: HTTP Range streaming transfers ~1.03 MiB (~0.1% of the compressed file), cutting egress and latency by ~99.9%.
COG & STAC Production Deployment Checklist
- Validate rasters with `rio cogeo validate` for header-first IFDs.
- Use 256x256 or 512x512 internal tiles and power-of-two overviews.
- Publish valid GeoJSON footprints and `proj:epsg` in every STAC Item.
- Enable `Accept-Ranges: bytes` and CORS headers on object storage.
- Set `CPL_VSIL_CURL_ALLOWED_EXTENSIONS=.tif` to block unneeded probes.
- Include `eo:cloud_cover` and `view:off_nadir` for fast CQL2 filtering.
Frequently Asked Questions
Can I open a COG in QGIS or ArcGIS Pro without downloading it?
Yes. Both stream COG URLs via GDAL /vsicurl/ using HTTP Range requests, fetching only visible viewport tiles.
Should I choose 256x256 or 512x512 internal tiles?
256x256 tiles reduce over-fetching for small web tiles and ML chips; 512x512 tiles cut HTTP request counts for regional jobs.
Does STAC store the satellite pixel data itself?
No. STAC is a lightweight JSON metadata index linking to external COG assets via URLs.
Why separate band COGs for spectral index pipelines?
Pipelines computing NDVI only need Red and NIR bands, avoiding network transfer of unused Blue or SWIR bytes.
Primary References & Standards
- OGC Cloud Optimized GeoTIFF (COG) Standard 1.0 (Open Geospatial Consortium).
- SpatioTemporal Asset Catalog (STAC) Specification v1.0.0 & STAC API.
- GDAL COG Driver & /vsicurl/ Virtual File System Documentation.