Skip to main content

Cloud-Native Geospatial Data

“Cloud-native” geospatial formats arrange data and metadata so clients can discover and read relevant portions efficiently. The property is a contract among file layout, indexes or partitions, metadata, object server, and reader—not a guarantee created by uploading an ordinary file.

Prerequisites​

Complete windowed raster processing, geospatial file formats, and MAUP and the ecological fallacy.

Learning goals​

  • Explain how COG tiles, overviews, GeoTIFF keys, and HTTP byte ranges work together.
  • Explain GeoParquet geometry metadata, column projection, row groups, and optional bounding-box covering.
  • Inspect metadata and read a raster window before loading whole data.
  • Distinguish a local window demonstration from remote performance evidence.

Raster and vector contracts​

OGC COG 1.0 builds on TIFF/BigTIFF and GeoTIFF. A conforming COG uses small internal tiles, reduced-resolution subfiles where required, and georeference keys; a serving system also supports HTTP byte ranges. Overviews avoid reading full-resolution pixels for a zoomed-out view. Range support alone does not turn any GeoTIFF into a COG, and compression, tile size, overview levels, access pattern, latency, cache, and request cost still determine performance.

GeoParquet stores geometry with declared encoding and metadata for the primary geometry column, geometry types, CRS, edge interpretation, and optional bounding box. Parquet column projection can avoid unrelated attributes. Row-group statistics and a conforming bounding-box covering can help readers prune data, but row-group size and spatial ordering matter. Metadata must be validated after round trips; an absent CRS field has a specification-defined meaning and must not be treated casually as “unknown.”

The local fixture below is intentionally an ordinary, untiled GeoTIFF with no overviews. It demonstrates metadata-first inspection, a two-by-two logical window, and a local byte-slice analogy. It is not labeled a COG and makes no network or performance claim.

Worked example​

from pathlib import Path

import rasterio
from rasterio.windows import Window

path = Path("tests/fixtures/geospatial-data-science/tiny-multiband.tif")
byte_range = path.read_bytes()[:64]
window = Window(col_off=0, row_off=0, width=2, height=2)
with rasterio.open(path) as dataset:
values = dataset.read(1, window=window, masked=True)
bounds = tuple(round(value) for value in dataset.window_bounds(window))
tiled = bool(dataset.profile.get("tiled", False))
print(f"driver={dataset.driver} crs={dataset.crs} tiled={tiled} overviews={dataset.overviews(1)}")
print(f"window=2x2 bounds={bounds} values={values.filled(dataset.nodata).tolist()}")
print(f"local_range=bytes=0-63 returned={len(byte_range)} file_bytes={path.stat().st_size} network_reads=0")
print("cog_conformance_claim=False performance_claim=False")

Expected output:

driver=GTiff crs=EPSG:32633 tiled=False overviews=[]
window=2x2 bounds=(391000, 5819000, 392000, 5820000) values=[[180, 185], [175, 180]]
local_range=bytes=0-63 returned=64 file_bytes=1384 network_reads=0
cog_conformance_claim=False performance_claim=False

This read proves only the local fixture contract. A remote benchmark would also record server headers, request count, bytes transferred, cold/warm caches, latency, reader version, and the exact query window.

Metadata-first workflow​

Discover an asset's media type, size, checksum, spatial and temporal extent, CRS, resolution, bands or columns, NoData semantics, license, and provenance. Select the smallest relevant partition, columns, overview, and window. Validate decoded CRS, transform, geometry types, bounds, schema, and missing values. Cache with identity and version metadata, then record the exact asset and query in evidence.

Limitations and common mistakes​

  • File extension and object storage location do not prove COG or GeoParquet conformance.
  • A bounding box is a candidate filter; exact geometry predicates still require refinement.
  • Antimeridian-crossing geometry needs special handling and is not supported by every bounding-box pruning scheme.
  • Tiny local reads are not benchmarks for remote services or production-scale files.
  • Public object access still requires licensing, privacy, integrity, availability, and cost controls.

Practice​

Design a remote-window benchmark without running it. Specify asset checksum, server range headers, cache state, window, overview, bytes, requests, latency repetitions, and the conclusion the benchmark cannot support.

Next steps​

Next, combine validation, network distance, uncertainty, and claim boundaries in the spatial accessibility project.

Primary sources​

Checked 2026-10-18: OGC, Cloud Optimized GeoTIFF 1.0; GeoParquet Standards Working Group, GeoParquet format specification; Apache Parquet, file format and metadata; Rasterio, windowed reading and writing.