6.5 Raster Formats — GeoTIFF and Cloud-Optimised GeoTIFF

Key takeaways

• GeoTIFF is a georeferenced variant of TIFF — the universal raster container.
• Compression, tiling, and overviews are orthogonal choices that massively affect performance.
• A Cloud-Optimised GeoTIFF (COG) rearranges the file so clients can stream only the bytes they need.

Introduction

If GeoJSON is the web's default vector format, GeoTIFF is the universal raster format. The cloud-optimised variant (COG) makes GeoTIFFs stream efficiently from S3-style object storage, powering modern platforms like Google Earth Engine, AWS Open Data, and Microsoft Planetary Computer.

TIFF and GeoTIFF

TIFF (Tagged Image File Format, 1986) is a container for raster data with flexible tags — a dictionary of metadata per image. GeoTIFF (1995) specifies which tags carry georeferencing: the CRS, the affine transform, and optional nodata.

A GeoTIFF is still a valid TIFF; a non-spatial TIFF viewer ignores the geo tags. This backward-compatibility is why GeoTIFF won.

Inspecting a GeoTIFF

gdalinfo -stats scene.tif

Returns:

• Driver (should say "GTiff").
• Size (width × height).
• Band count and data type.
• CRS (in WKT).
• Affine transform / pixel size.
• Nodata value.
• Statistics per band.

Compression

GeoTIFF supports many compression algorithms:

Choose based on bit depth and use case:

• 16-bit scientific data (Sentinel-2, Landsat surface reflectance) → ZSTD or DEFLATE.
• 8-bit natural colour orthophotos → JPEG at 75–90 % quality.
• Classified 8-bit land cover → LZW (keep exact class values).

Tiling vs stripes

Two internal layouts:

• Stripes — rows or groups of rows stored contiguously. Default; inefficient for windowed reads.
• Tiles — 256×256 or 512×512 square blocks stored independently. Enables fast region-of-interest reads.

Tiled is the right default for any analysis. Strided is fine for purely sequential reads (e.g., display).

Overviews (pyramids)

Large rasters benefit from pre-computed downsampled versions — overviews — that let viewers display low-zoom quickly without reading the full resolution. GDAL creates them:

gdaladdo -r average scene.tif 2 4 8 16 32

This creates five overview levels (2×, 4×, 8×, 16×, 32× downsampled). The numbers are multiplicative factors: each overview has half the linear resolution of the previous.

External overviews live in a sidecar .ovr file; internal overviews (recommended for COG) live inside the TIFF.

Cloud-Optimised GeoTIFF (COG)

A COG is a regular GeoTIFF with a specific internal layout:

1. Header first — CRS, transform, IFDs at the start of the file.
2. Internal tiles — not stripes.
3. Internal overviews — not sidecar .ovr.
4. Tiles + overviews ordered bottom-up in the file so a client can read low-resolution first, then zoom in.

Crucially, a COG is still a valid GeoTIFF — every existing tool reads it without modification. But a cloud-aware client can fetch only the byte ranges it needs over HTTP range requests, skipping unneeded tiles entirely.

Create a COG:

1gdal_translate scene.tif scene_cog.tif \
2  -of COG \
3  -co COMPRESS=ZSTD \
4  -co PREDICTOR=2 \
5  -co BIGTIFF=IF_SAFER

Validate:

python -m validate_cloud_optimized_geotiff scene_cog.tif

Reading COGs from the cloud

1import rasterio
2from rasterio.windows import Window

Rasterio and GDAL transparently use HTTP range requests on remote URLs.

Other raster formats you'll meet

• JP2 / JPEG 2000 — wavelet compression, used by historical Sentinel-2 distribution. Patent status now free.
• NetCDF / HDF5 — multi-dimensional scientific formats (time + variables); common in climate / oceans.
• Zarr — cloud-native chunked N-dim arrays; the Python / Xarray ecosystem favourite.
• ERDAS IMG (.img) — ESRI-friendly raster; mostly legacy.
• BIL / BIP / BSQ — band-interleaved formats for multispectral; legacy.

A practical workflow

1. Receive data in the format provider gives you (often uncompressed TIFF or JP2).
2. Inspect with gdalinfo — record CRS, resolution, data type, nodata.
3. Reproject if needed with gdalwarp.
4. Convert to COG for cloud serving or analysis pipelines.
5. Document in a sidecar metadata file.

Self-check exercises

Summary

• GeoTIFF is the universal raster format; COG is its cloud-native layout.
• Compression, tiling, and overviews are orthogonal choices with big performance implications.
• COGs enable bytes-on-demand streaming — the foundation of modern raster platforms.

Further reading

• OGC — Cloud-Optimized GeoTIFF standard.
• cogeo.org — curated COG resources.
• gdal_translate -of COG documentation.
• Planet and AWS open data documentation on COG publishing.