--- title: "One Geospatial AI Window: Trace GeoTIFF Resolution and Storage Tiles" description: "Leviathan Systems — Technical content Published 2026-10-20; updated 2026-10-20." canonical: "https://www.leviathansystems.co/articles/geospatial-ai-cog-overview-resolution-window-tiles" last-updated: "2026-10-20" --- # One Geospatial AI Window: Trace GeoTIFF Resolution and Storage Tiles Leviathan Systems — Technical content Published 2026-10-20; updated 2026-10-20. Choose the sample grid required by the AI task before comparing how a Cloud Optimized GeoTIFF reader accesses it. The same geographic area can contain 262,144 base-image positions or 16,384 overview positions in the worked example. Storage tiles help locate those windows in a file; their count does not establish retained detail, transferred bytes or model performance. An overview is appropriate when the agreed task permits its coarser samples. ## CO-GRID: one COG area at two resolutions A provider offering a geospatial AI input can identify the right geographic area while leaving the supplied resolution unresolved. Start with that difference. This explanation uses the OGC Cloud Optimized GeoTIFF Standard, version 1.0, document 21-026, published July 14, 2023. Its section 7.2.1 describes an overview as a reduced-resolution image stored within the same TIFF file. Original fictional case CO-GRID has a single-band base image of 4,096 x 4,096 pixels and a selected overview of 1,024 x 1,024 pixels. Both selected levels have 512 x 512-pixel storage tiles. Their grids share the same geographic extent and top-left origin, with an exact 4:1 dimension ratio in each direction. Other resolution levels may exist; the case identifies only the two selected grids. Use zero-based, half-open pixel windows: [768, 1280) includes positions 768 through 1279. Coordinates below belong to the selected image level, rather than geographic coordinate units or file-byte offsets. No real raster, software execution or measured transfer is supplied. References: [OGC1](https://docs.ogc.org/is/21-026/21-026.html), [COG-LEVEL](https://docs.ogc.org/is/21-026/21-026.html#_requirement_reduced_resolution_subfiles) ## The base window crosses four storage tiles At base resolution, request x = [768, 1280) and y = [768, 1280). Each dimension contains 1280 - 768 = 512 positions, giving 512 x 512 = 262,144 spatial sample positions. Because this case has one band, each position has one band value. The base level's tile grid has 4096 / 512 = 8 columns and eight rows. Tile column 1 covers x = [512, 1024), and column 2 covers [1024, 1536). The requested x interval intersects both. The same calculation selects tile rows 1 and 2. Their combinations are four candidate tiles: (1, 1), (2, 1), (1, 2) and (2, 2), written as (column, row). The window occupies a 256 x 256 region inside each candidate. Four such regions contain 4 x 256 x 256 = 262,144 positions, agreeing with the window calculation. The four complete tiles contain 4 x 512 x 512 = 1,048,576 positions. That larger population describes the intersected storage units, rather than the requested window. A window with the same dimensions as one storage tile can therefore cross four tiles when it straddles a tile boundary in both directions. Section 7.1.2 describes the rectangular tile organization; CO-GRID supplies this particular placement. The four candidates establish spatial intersection without prescribing how a reader fetches or decodes them. References: [COG-TILE](https://docs.ogc.org/is/21-026/21-026.html#_tiles) ## The same area becomes a 128 x 128 overview window In CO-GRID's exact 4:1 mapping, divide the base-window boundaries by four. The same area is x = [192, 320) and y = [192, 320) in the selected overview. Its 128 x 128 window contains 16,384 sample positions. That overview has 1024 / 512 = 2 tile columns and two tile rows. Both window intervals lie inside [0, 512), so only tile (0, 0) intersects. The complete candidate tile contains 512 x 512 = 262,144 positions, while the requested region uses 16,384 of them. Compare like quantities. The requested-position ratio is 262,144 / 16,384 = 16:1: four times as many positions along each of two dimensions gives sixteen times as many positions in the area. The intersecting-tile ratio is 4:1. The complete candidate-tile populations also have a 4:1 ratio, 1,048,576 / 262,144. Neither ratio supplies a byte, elapsed-time or GPU-work multiplier. The overview's unchanged 512 x 512 storage-tile dimensions do not make its samples as fine as the base samples. In this case, a tile spans four times the geographic distance in each direction at the overview level. Sections 7.3.1-7.3.2 explain the shared georeference and dimension-based pixel scales. Exact division here follows the stipulated grids; a real image's dimensions and georeference must establish its own mapping. References: [COG-GRID](https://docs.ogc.org/is/21-026/21-026.html#_requirement_georeference_keys) ## A matching extent leaves the original values unresolved The overview covers the same place with fewer samples. A simple invented averaging example shows why the original values cannot generally be inferred from the coarse result. For this illustration only, define each overview value as the arithmetic mean of one aligned 4 x 4 base block. Block A contains sixteen values of 1. Its sum is 16 and its mean is 1. Block B contains fifteen zeros and one value of 16. Its sum is also 16 and its mean is also 1, wherever the 16 sits within the block. The coarse value 1 cannot distinguish a uniform block from that concentrated value, or locate the concentration. Even an exactly known mean leaves several original arrangements possible. This is a many-to-one transformation under the assumed averaging rule. It is not a requirement that COG use this rule, and it predicts no AI accuracy result. Expanding the 128 x 128 overview window to a 512 x 512 array can restore the requested array dimensions. The expanded dimensions alone supply none of the missing original distinctions demonstrated above. Such an array and the base window can have equal shapes while leaving their sample provenance materially different. A task allowed to use the coarse population may be well served by the overview. A task requiring the base sample population needs that requirement established separately. ## TIFF metadata connects tile locations to file bytes After selecting the image level and window, a capable reader can use TIFF metadata to locate the relevant tile data in the file. OGC 1.0 section 8 explains the mapping from tile indices to byte ranges using TIFF headers and image file directories. This connects a spatial request to encoded file locations. A pixel coordinate alone is insufficient to calculate an encoded byte offset. CO-GRID leaves the encoded layout, compression, metadata access, cache state and reader/server behavior unspecified. Its four candidate tiles do not establish four HTTP requests, complete fetching or decoding of every candidate tile, or a transferred-byte total. A real reader may reuse cached data or organize its accesses differently. Byte and time claims need evidence for the actual file and implementation at a stated measurement boundary. Sections 7.1.3-7.1.4 discuss compression and component layout. CO-GRID deliberately uses one band; a multiband offer also needs to explain the selected bands and their storage organization. A count of spatial tiles alone does not settle that additional access question. The related HTTP representation guide addresses continuity when combining partial deliveries. That remains a separate concern from choosing a raster level and finding the window's intersecting tiles. References: [COG-ACCESS](https://docs.ogc.org/is/21-026/21-026.html#HTTPRangeSupportRequirements), [COG-ENCODING](https://docs.ogc.org/is/21-026/21-026.html#_compression) ## Read the conformance class behind the format claim The standard separates requirements from recommendations. Section 2 defines distinct conformance classes. Section 7.1.2 requires tiled organization for the GeoTIFF Tiles class. Section 7.2.1 discusses optional overview presence across the standard and the obligations of the GeoTIFF Overviews class; section 7.4 adds requirements for the Optimized GeoTIFF class and its dependencies. In particular, section 7.4.2 requires a progression of reduced-resolution subfiles and a final level with one tile across or down. Its Recommendation 4 addresses preferred tile sizes and resolution progression. Compression in section 7.1.3 is also a recommendation. These statements do not make 512-pixel tiles or CO-GRID's selected 4:1 mapping universal settings. CO-GRID does not enumerate a complete overview pyramid or evaluate a file against these classes. Ask which class supports a supplier's format claim. The presence of an overview, or a successful window read, leaves that broader claim to its own evidence. References: [COG-CONFORMANCE](https://docs.ogc.org/is/21-026/21-026.html#optimized_geotiff-requirements-class), [COG-TILE](https://docs.ogc.org/is/21-026/21-026.html#_tiles), [COG-LEVEL](https://docs.ogc.org/is/21-026/21-026.html#_requirement_reduced_resolution_subfiles), [COG-ENCODING](https://docs.ogc.org/is/21-026/21-026.html#_compression) ## Decide which population the provider must deliver Return to the application requirement before comparing access offers. Suppose offer A supplies CO-GRID's 512 x 512 base window. Offer B supplies the same area's 128 x 128 overview window. Their matching geographic extent establishes where the inputs refer to; it leaves the two sample populations different. If B expands its output to 512 x 512, matching array dimensions still do not establish delivery of A's base samples. If the agreed task requires the base-level sample population, an overview window or resized copy does not establish that deliverable. If the task explicitly permits the coarser population and its value-generation method, B can be appropriate. The requirement decides which offers remain eligible; the tile count cannot decide on its behalf. Have the appointed data or software provider identify the source file or release, image level, dimensions, georeference, selected band, data type, value units or scaling, window bounds and any transformation before the application receives the array. For a real task, also resolve how invalid or missing samples are represented so a position count is not silently treated as a count of usable observations. Obtain the resulting array dimensions and sample provenance together. CO-GRID's simple counts do not resolve those dataset-specific properties. Once that input is agreed, compare byte or timing evidence for the same deliverable and stated reader path. The related Parquet guide examines a different format's logical selection and storage layout. The vision patch-grid guide continues downstream from supplied pixels into encoder positions. Neither replaces the resolution choice made here. ## Scope a separate physical connection requirement If the surrounding storage and compute equipment also needs physical connections, Leviathan's Structured Cabling & Fiber Installation service is the relevant route for agreed routing, dressing, endpoint labels and connection records. Use the related project inquiry with a nonconfidential equipment and site outline, approved endpoints, material-supply responsibilities and required physical handoff. Raster preparation, geospatial processing, reader behavior and model evaluation remain with the appointed data and software providers. The physical inquiry concerns the selected equipment's cabling scope, requirements and availability. References: [L1](https://www.leviathansystems.co/services/structured-cabling) ## Primary sources - [OGC1 — OGC Cloud Optimized GeoTIFF Standard 1.0, document 21-026](https://docs.ogc.org/is/21-026/21-026.html). Checked 2026-10-20. Published July 14, 2023. Provides the historical version and terminology used for this explanation; no claim about a later revision is made. - [COG-LEVEL — OGC COG 1.0, section 7.2.1: Reduced-Resolution Subfiles](https://docs.ogc.org/is/21-026/21-026.html#_requirement_reduced_resolution_subfiles). Checked 2026-10-20. Describes reduced-resolution images within TIFF and overview-class obligations. It supplies no mandated averaging method or task-specific AI quality result. - [COG-TILE — OGC COG 1.0, section 7.1.2: Tiles](https://docs.ogc.org/is/21-026/21-026.html#_tiles). Checked 2026-10-20. Defines rectangular tile organization and its requirement within the GeoTIFF Tiles class. CO-GRID's tile dimensions and window placement are fictional assumptions. - [COG-GRID — OGC COG 1.0, sections 7.3.1-7.3.2: Georeference and pixel scale](https://docs.ogc.org/is/21-026/21-026.html#_requirement_georeference_keys). Checked 2026-10-20. Explains shared origin, geographic reference and extent for linked overviews, with pixel scales derived from width and height ratios. No location or geodetic accuracy is assessed. - [COG-ACCESS — OGC COG 1.0, sections 6.1 and 8: Spatial selection and HTTP ranges](https://docs.ogc.org/is/21-026/21-026.html#HTTPRangeSupportRequirements). Checked 2026-10-20. Connects TIFF metadata and tile indices to file-byte ranges. It provides no measured request count, byte total or latency for the fictional case. - [COG-ENCODING — OGC COG 1.0, sections 7.1.3-7.1.4: Compression and planar configuration](https://docs.ogc.org/is/21-026/21-026.html#_compression). Checked 2026-10-20. Distinguishes the compression recommendation and component-storage arrangements relevant to actual file access. No codec or band layout is selected here. - [COG-CONFORMANCE — OGC COG 1.0, sections 2 and 7.4: Conformance classes and optimized organization](https://docs.ogc.org/is/21-026/21-026.html#optimized_geotiff-requirements-class). Checked 2026-10-20. Separates class-specific requirements from recommendations for tile sizes and overview progression. A two-level arithmetic example cannot establish complete file conformance. - [L1 — Leviathan Systems: Structured Cabling & Fiber Installation](https://www.leviathansystems.co/services/structured-cabling). Checked 2026-10-20. Defines the existing physical cabling service, agreed material responsibilities, endpoint labels and connection records for a separate equipment project. ## Related Leviathan resources - [HTTP byte ranges and continuity of the same representation](/articles/http-byte-range-resume-representation-validator) - [Parquet layout, selected populations and remaining scan work](/articles/ai-parquet-row-groups-projection-and-filter-scan) - [From supplied image pixels to vision encoder positions](/articles/vision-transformer-patch-grid-input-sequence) - [Structured Cabling & Fiber Installation](/services/structured-cabling) - [Discuss a separate physical cabling requirement](/contact?service=structured-cabling&from=%2Farticles%2Fgeospatial-ai-cog-overview-resolution-window-tiles) ## Does the surrounding equipment need physical cabling? Share a nonconfidential equipment, site, endpoint and physical handoff outline to discuss cabling scope, requirements and availability. [Discuss the scope](/contact?service=structured-cabling&from=/articles/geospatial-ai-cog-overview-resolution-window-tiles)