What is Geospatial Data? Types, Formats, Examples

Geospatial data is information tied to a location on the Earth's surface. A store, a road, a storm forecast, and a satellite photo of a city are all geospatial data. Each records where something is and what it is. The US Census Bureau describes geospatial data as data that is "clearly geographic in nature," such as maps and data for geographic information system (GIS) software. This article claims that most questions about a place need both main types in one query. The two types are vector objects, such as buildings, and raster grids, such as a land cover map. A walkthrough over Lake Havasu City, Arizona, tests the claim on five layers.

Key takeaways

• Geospatial data combines a location, attributes, and often a timestamp.
• The two main types are vector data (points, lines, polygons) and raster data (grids of pixels). Point clouds and trajectories are common specialised forms.
• A coordinate reference system ties the numbers in a dataset to real places on Earth.
• Common formats include GeoJSON, Shapefile, GeoTIFF, and GeoParquet. Apache Iceberg tables now store geometry natively.
• Sources range from satellites and sensors to government agencies and open data projects such as Overture Maps.

What is geospatial data?

"Geospatial" combines "geo", the Earth, with "spatial", relating to space and position. Every geospatial record has three parts.

1. Location. A geometry, such as a point at a longitude and latitude, a line, or a polygon, or a cell in a grid. The location only has meaning with a coordinate reference system (CRS) that says which model of the Earth the numbers refer to.
2. Attributes. The facts about the location: a building's height, a road's speed limit, a pixel's reflectance, a parcel's owner.
3. Time. When the observation was made or when it is valid. A satellite scene, a GPS ping, and a weather forecast all carry a timestamp.

Geospatial data vs spatial data vs GIS data

Spatial data is any data with a position in some space, including a warehouse floor plan or a medical scan. Geospatial data is spatial data referenced to the Earth. GIS data is geospatial data packaged as map layers for a geographic information system.

How geospatial data works

Coordinates need a reference system. The pair (-114.35, 34.47) is a spot in Lake Havasu City, Arizona. It points there only because a CRS says the numbers are longitude and latitude on the ellipsoid of WGS 84, the World Geodetic System 1984. The International Association of Oil & Gas Producers (IOGP) maintains a dataset called EPSG Dataset. It assigns codes such as EPSG:4326 to these systems.

Objects or fields. Helen Couclelis framed the two ways people model geography in People manipulate objects (but cultivate fields) (1992). In the object view, the world is made of things with boundaries, such as a building: vector data. In the field view, a value exists everywhere, such as elevation: raster data.

Geometry has a standard shape. The Open Geospatial Consortium (OGC) Simple Feature Access standard defines the core geometry types: Point, LineString, Polygon, and their multi and collection forms. The International Organization for Standardization (ISO) also publishes it as ISO 19125. It ties each geometry to a spatial reference system. PostGIS and Apache Sedona follow it, so ST_Intersects means the same thing in both.

Relations are computed from geometry. Max Egenhofer and Robert Franzosa's point-set topological spatial relations (1991) compare the interiors and boundaries of two shapes to define relations such as contains, touches, and overlaps. Spatial SQL uses those relations as true-or-false tests. A spatial join applies them across whole tables.

Types of geospatial data

Vector data

Vector data represents discrete features with geometry and a row of attributes.

• Points: a single location, such as a store, a sensor, or a point of interest.
• Lines: a sequence of points, such as a road, a river, or a ship's track.
• Polygons: a closed area, such as a building footprint, a property parcel, or a census tract.

Raster data

Raster data divides an area into a regular grid of cells. Each cell holds one value per band, a band being one measured layer. A satellite image stores reflectance per pixel. A digital elevation model stores height, and a land cover map stores a class code. A weather model such as the High-Resolution Rapid Refresh (HRRR) stores temperature or wind speed. Cell size sets the resolution: a 10-meter pixel in Sentinel-2 imagery covers 100 square meters on the ground.

Point clouds

A point cloud is a set of 3D points, each with x, y, and z coordinates and attributes such as return intensity and a class code (ground, building, vegetation). Most come from lidar (light detection and ranging) scanners on aircraft, drones, and vehicles. The LAS format, the standard exchange file for point clouds, comes from the American Society for Photogrammetry and Remote Sensing (ASPRS). The U.S. Geological Survey (USGS) 3D Elevation Program references LAS 1.4 revision 15.

Trajectories

A trajectory is an ordered sequence of positions and timestamps for one moving object. Examples are a delivery van, a phone, a vessel broadcasting its position over the Automatic Identification System (AIS), or an aircraft. Yu Zheng's survey Trajectory Data Mining: An Overview (ACM Transactions on Intelligent Systems and Technology, 2015) covers cleaning, indexing, and mining them.

Other forms

Tables of addresses become geospatial once geocoded. Discrete global grids such as H3 assign every place on Earth to a cell ID, so tables join on the cell. Kevin Sahr, Denis White, and A. Jon Kimerling describe these grids in Geodesic Discrete Global Grid Systems (2003).

Vector vs raster data

	Vector data	Raster data
Structure	Points, lines, and polygons with attributes	Grid of cells, each with one or more values
Model of the world	Objects with boundaries	Fields defined everywhere
Best for	Discrete features	Continuous surfaces and imagery
Examples	Buildings, roads, parcels, places	Satellite imagery, elevation, land cover
Resolution	Set by coordinate precision	Set by cell size
Common formats	GeoJSON, Shapefile, GeoPackage, GeoParquet	GeoTIFF, Cloud Optimized GeoTIFF, NetCDF, Zarr

For a full comparison, see raster vs vector data.

Geospatial data formats and standards

• GeoJSON is a JSON format for vector features. Its specification is Request for Comments (RFC) 7946 (2016), which fixes its coordinates to WGS 84 longitude and latitude in decimal degrees.
• Shapefile is an older Esri vector format made of several files. It limits column names to 10 characters.
• GeoPackage is an OGC standard that stores vector features and raster tiles in one SQLite file.
• GeoTIFF adds georeferencing tags to TIFF (Tagged Image File Format) images. The OGC GeoTIFF standard formalized it in version 1.1. A Cloud Optimized GeoTIFF (COG) orders its internal tiles so a client can fetch only the ones it needs over HTTP.
• GeoParquet adds geometry types to Apache Parquet, a columnar format. Version 1.1 encodes geometry as well-known binary (WKB) or GeoArrow and stores the CRS as PROJJSON, a JSON encoding of coordinate reference systems. WherobotsDB reads and writes it. Its catalog tables are Iceberg tables of Parquet files.
• Apache Iceberg tables now support geometry and geography types, so vector data can live in a data lakehouse with time travel, which reads earlier table versions. See Iceberg v3 gets native geo types.
• NetCDF and Zarr store multidimensional arrays such as climate model output.

Where geospatial data comes from

• Satellites and aircraft. Optical, radar, thermal, and lidar sensors image the Earth on regular schedules. Landsat 1 launched on 23 July 1972 and started the longest-running Earth observation record.
• Sensors and devices. Phones, vehicles, weather stations, and Internet of Things (IoT) devices report positions and readings.
• Government agencies. The Census Bureau's TIGER/Line Shapefiles publish legal and statistical boundaries every year.
• Open data projects. Overture Maps Foundation releases six themes (addresses, base, buildings, divisions, places, and transportation) with stable Global Entity Reference System (GERS) IDs for joining. The European Space Agency (ESA) publishes WorldCover, which maps global land cover at 10 m in 11 classes. Volunteers build OpenStreetMap, as Muki Haklay and Patrick Weber describe (2008). Michael Goodchild called this pattern volunteered geographic information (2007).
• Commercial providers and models. Parcels, high-resolution imagery, mobility data, forecasts, and AI outputs such as footprints extracted from imagery.

Geospatial data examples

• Satellite and aerial imagery. Earth observation data from missions such as Landsat and Sentinel-2, synthetic aperture radar that images through clouds, and hyperspectral imagery with hundreds of narrow bands. Satellite imagery analysis turns the pixels into maps.
• Derived layers. Normalized Difference Vegetation Index (NDVI) for vegetation health, land cover classification maps such as WorldCover, and change detection results that flag where the ground changed.
• Elevation. The Copernicus digital elevation model (DEM), a 30 m global surface model published as Cloud Optimized GeoTIFFs, and lidar-derived terrain models.
• Buildings, places, and roads. Footprints, points of interest, and road networks from open datasets such as Overture Maps.
• Boundaries. Countries, counties, census tracts, ZIP codes, and parcels.
• Hazard and risk layers. Flood risk maps, wildfire perimeters, and the event sets used in catastrophe modeling.
• Movement. Vehicle telematics, vessel positions, and aircraft tracks.

Insurers join buildings and parcels to hazard layers for probable maximum loss estimates and stochastic models. Energy companies screen land for solar farms and data center sites. Geospatial AI models and agents read the same layers to answer questions about real places, the basis of physical AI. Geospatial analysis is the set of methods that turn these layers into answers.

Geospatial data in Wherobots

Wherobots is a cloud platform for geospatial data, built by the creators of Apache Sedona. In WherobotsDB, vector and raster data share one SQL and Python engine, with tables in GeoParquet and Apache Iceberg. The Havasu catalog lists the open datasets ready to query, including Overture places, Sentinel-2 imagery, and National Agriculture Imagery Program (NAIP) aerial imagery. The Wherobots Model Context Protocol (MCP) server lets AI coding tools explore the catalog and run spatial SQL.

The walkthrough uses one box around Lake Havasu City, Arizona, and five layers. Overture buildings, roads, and places are the vector data. ESA WorldCover land cover and the Copernicus DEM, a digital elevation model, are the raster data.

Source	Data model and format	Size	What the queries read
Overture buildings (buildings_building)	Vector polygons, Iceberg table of Parquet files	2,533,842,612 rows worldwide	4,844 footprints in the box: count, area, centroid
Overture transportation segments (transportation_segment)	Vector lines, Iceberg table of Parquet files	352,054,710 rows worldwide, all subtypes	1,931 road segments in the box: count, length
Overture places (places_place)	Vector points, Iceberg table of Parquet files	81,455,423 rows worldwide	2,078 places in the box: count, category, point
ESA WorldCover 2021 v200	Raster, Cloud Optimized GeoTIFF (COG), 10 m pixels	One 3 x 3 degree file	277,200 pixels in the box: class code; class under each feature
Copernicus DEM GLO-30	Raster, COG, 30 m pixels	One 1 x 1 degree file	Surface elevation under 6,922 features

Row counts are COUNT(*) results on the Wherobots open data tables in October 2026. WherobotsDB ran all three queries in one SQL engine, on vector tables and raster files together. Most stacks pick one data model.

Vector features in the Lake Havasu City box

ST_PolygonFromEnvelope builds a rectangle from two corners. ST_Intersects keeps the features that touch it. This query counts the three vector layers and measures them on the WGS 84 ellipsoid:

WITH aoi AS (SELECT ST_PolygonFromEnvelope(-114.375, 34.455, -114.32, 34.49) AS g)
SELECT 'buildings' AS layer, COUNT(*) AS features,
       ROUND(SUM(ST_AreaSpheroid(ST_Intersection(b.geometry, aoi.g))) / 1e6, 3) AS area_km2, CAST(NULL AS DOUBLE) AS length_km
FROM wherobots_open_data.overture_maps_foundation.buildings_building b, aoi
WHERE ST_Intersects(b.geometry, aoi.g)
UNION ALL
SELECT 'road segments', COUNT(*), NULL, ROUND(SUM(ST_LengthSpheroid(ST_Intersection(s.geometry, aoi.g))) / 1000, 1)
FROM wherobots_open_data.overture_maps_foundation.transportation_segment s, aoi
WHERE s.subtype = 'road' AND ST_Intersects(s.geometry, aoi.g)
UNION ALL
SELECT 'places', COUNT(*), NULL, NULL
FROM wherobots_open_data.overture_maps_foundation.places_place p, aoi
WHERE ST_Intersects(p.geometry, aoi.g)

layer	features	area_km2	length_km
buildings	4,844	1.635
road segments	1,931		243.0
places	2,078

The query ran in 33.8 seconds. Grouping places by category profiles the town:

SELECT basic_category, COUNT(*) AS places
FROM wherobots_open_data.overture_maps_foundation.places_place
WHERE ST_Intersects(geometry, ST_PolygonFromEnvelope(-114.375, 34.455, -114.32, 34.49))
  AND basic_category IS NOT NULL
GROUP BY basic_category
ORDER BY places DESC
LIMIT 5

basic_category	places
real_estate_service	264
personal_or_beauty_service	156
restaurant	126
financial_service	119
home_service	76

Land cover pixels in the same box

ESA WorldCover stores one class code per 10 m pixel in 3 by 3 degree COGs. RS_TileExplode splits the file into 512-pixel tiles, and the engine reads a tile only when a query uses it. RS_Clip cuts the tiles that touch the box. RS_BandAsArray turns the pixels into rows to count:

WITH aoi AS (
  SELECT ST_PolygonFromEnvelope(-114.375, 34.455, -114.32, 34.49) AS g
),
wc AS (   -- 512-pixel tiles of the 3 x 3 degree file, read on demand
  SELECT RS_TileExplode(RS_FromPath('s3://esa-worldcover/v200/2021/map/ESA_WorldCover_10m_2021_v200_N33W117_Map.tif'), 512, 512) AS (x, y, rast)
),
c AS (    -- RS_Clip(raster, band, geometry, allTouched, noDataValue, crop)
  SELECT /*+ BROADCAST(aoi) */ RS_Clip(wc.rast, 1, aoi.g, false, 0, true) AS c
  FROM wc CROSS JOIN aoi
  WHERE RS_Intersects(wc.rast, aoi.g)
)
SELECT CAST(v AS INT) AS worldcover_class, COUNT(*) AS pixels,
       ROUND(100 * COUNT(*) / SUM(COUNT(*)) OVER (), 1) AS pct
FROM (SELECT explode(RS_BandAsArray(c, 1)) AS v FROM c)
WHERE v > 0
GROUP BY CAST(v AS INT)
ORDER BY pixels DESC

The box holds 277,200 pixels, counted in 5.0 seconds. Built-up (class 50) covers 40.2%, permanent water (80) 30.2%, bare or sparse vegetation (60) 23.8%, tree cover (10) 2.0%, shrubland (20) 1.9%, grassland (30) 1.9%, and cropland (40) 0.1%.

Vector features joined to both rasters

In the last query, each building becomes its centroid, the point at the center of its footprint. Each place stays a point. RS_Value returns the WorldCover class and the DEM elevation under every feature. BROADCAST copies the few raster tiles to every worker. The query plan has two BroadcastIndexJoin steps, each with an R-tree index on the tiles, and no Cartesian product, the step that pairs every row with every other row. Each feature keeps its Overture id through the grouping:

WITH aoi AS (
  SELECT ST_PolygonFromEnvelope(-114.375, 34.455, -114.32, 34.49) AS g
),
wc AS (   -- ESA WorldCover 2021, 10 m land cover: 512-pixel tiles of one 3 x 3 degree COG
  SELECT RS_TileExplode(RS_FromPath('s3://esa-worldcover/v200/2021/map/ESA_WorldCover_10m_2021_v200_N33W117_Map.tif'), 512, 512) AS (x, y, rast)
),
dem AS (  -- Copernicus DEM GLO-30, 30 m surface elevation: 512-pixel tiles of one 1 x 1 degree COG
  SELECT RS_TileExplode(RS_FromPath('s3://copernicus-dem-30m/Copernicus_DSM_COG_10_N34_00_W115_00_DEM/Copernicus_DSM_COG_10_N34_00_W115_00_DEM.tif'), 512, 512) AS (x, y, rast)
),
wc_aoi  AS (SELECT /*+ BROADCAST(aoi) */ wc.rast  FROM wc  CROSS JOIN aoi WHERE RS_Intersects(wc.rast, aoi.g)),
dem_aoi AS (SELECT /*+ BROADCAST(aoi) */ dem.rast FROM dem CROSS JOIN aoi WHERE RS_Intersects(dem.rast, aoi.g)),
f AS (    -- vector features: building centroids and place points
  SELECT /*+ BROADCAST(aoi) */ 'building' AS kind, b.id, ST_Centroid(b.geometry) AS pt
  FROM wherobots_open_data.overture_maps_foundation.buildings_building b CROSS JOIN aoi
  WHERE ST_Intersects(b.geometry, aoi.g)
  UNION ALL
  SELECT /*+ BROADCAST(aoi) */ 'place', p.id, p.geometry
  FROM wherobots_open_data.overture_maps_foundation.places_place p CROSS JOIN aoi
  WHERE ST_Intersects(p.geometry, aoi.g)
),
v AS (    -- one land cover class and one elevation per feature, keyed by its Overture id
  SELECT /*+ BROADCAST(w), BROADCAST(d) */ f.kind, f.id,
         CAST(RS_Value(w.rast, f.pt) AS INT) AS worldcover_class,
         RS_Value(d.rast, f.pt)              AS elev_m
  FROM f
  JOIN wc_aoi  w ON RS_Intersects(w.rast, f.pt)
  JOIN dem_aoi d ON RS_Intersects(d.rast, f.pt)
)
SELECT kind, worldcover_class,
       COUNT(DISTINCT id)                        AS features,
       ROUND(percentile_approx(elev_m, 0.5), 1)  AS median_elev_m,
       ROUND(MIN(elev_m), 1)                     AS min_elev_m,
       ROUND(MAX(elev_m), 1)                     AS max_elev_m
FROM v
GROUP BY kind, worldcover_class
ORDER BY kind, features DESC

kind	worldcover_class	features	median_elev_m	min_elev_m	max_elev_m
building	50 built-up	4,477	176.0	136.0	244.7
building	60 bare / sparse	330	171.1	136.0	242.9
building	10 tree cover	24	142.9	138.3	172.5
building	80 water	7	136.8	136.0	138.3
building	20, 40, 30	3, 2, 1
place	50 built-up	1,851	176.0	136.0	235.9
place	60 bare / sparse	195	156.2	138.2	226.2
place	80 water	14	136.0	136.0	176.0
place	30, 10, 40	9, 8, 1

The Copernicus DEM is a surface model, so its heights include roofs and trees. Every feature matched a pixel in both rasters. The building rows sum to 4,844 and the place rows to 2,078, the totals from the first query. Built-up land covers 40.2% of the box and holds 4,477 of the 4,844 buildings, 92.4%. The 7 buildings on water pixels sit at 136.0 to 138.3 m, the lowest values in the table. They are on the shoreline, where a 10 m pixel of lake can cover a dock or boathouse. Buildings reach 244.7 m in the hills to the northeast, about 109 m above the shoreline.

The 330 buildings on bare or sparse pixels are places where the two sources disagree. Possible causes are construction after 2021, or structures smaller than a 10 m pixel. The same pattern scales to a county or country by swapping the box for tract or parcel polygons from the census tracts table, or by summarising features into H3 cells.

Geospatial data research and what changes at scale

The walkthrough's box filter, spatial index, and raster-vector join come from six decades of research.

From a land inventory to GIS

Digital geospatial data starts with the Canada Land Inventory, set up in 1962 to map what rural land could support. Roger Tomlinson led the system that automated its map analysis, the Canada Geographic Information System, the first operational GIS. His 1968 paper is the first known published use of the term. The Royal Geographical Society's obituary of Roger Tomlinson (2014) calls him the father of GIS.

Landsat 1 added a gridded stream, the raster archive behind Earth observation data. Michael Goodchild argued in Geographical information science (1992) that representation, accuracy, and spatial analysis formed a research field of their own.

Spatial databases and indexes

Rows have no natural order in two dimensions, so finding the features in a box needs a spatial index. Raphael Finkel and Jon Bentley's quad trees (Acta Informatica, 1974) split space into four quadrants recursively. Antonin Guttman's R-trees: a dynamic index structure for spatial searching (SIGMOD 1984) groups nearby objects under bounding rectangles in a balanced tree. A search visits only the rectangles that overlap its box. Norbert Beckmann, Hans-Peter Kriegel, Ralf Schneider, and Bernhard Seeger's R*-tree (SIGMOD 1990) improved how the tree splits its nodes. R-tree variants remain a common spatial index.

Thomas Brinkhoff, Kriegel, and Seeger processed spatial joins with R-trees (SIGMOD 1993) by walking two trees together. Jignesh Patel and David DeWitt's partition based spatial-merge join (SIGMOD 1996) split both inputs into spatial partitions. It then joined each pair of partitions. Distributed engines still follow this pattern. Ralf Hartmut Güting's survey An introduction to spatial database systems (1994) listed what a spatial database must offer: spatial types, spatial indexing, and spatial joins. The Open Geospatial Consortium, founded in 1994, turned those types into the Simple Features standard. PostGIS, Apache Sedona, and cloud warehouses implement it.

Distributed engines GeoSpark and Apache Sedona

In the 2010s, global datasets grew too large for one database server. Ahmed Eldawy and Mohamed Mokbel built SpatialHadoop (ICDE 2015) on MapReduce. Jia Yu, Jinxuan Wu, and Mohamed Sarwat introduced GeoSpark (ACM SIGSPATIAL 2015) on Apache Spark. Yu, Zongsi Zhang, and Sarwat described its design in Spatial data management in Apache Spark: the GeoSpark perspective and beyond (2019). GeoSpark splits data into spatial partitions, indexes each partition, and runs range, nearest-neighbour, and join queries across a cluster. GeoSpark entered the Apache incubator as Sedona in 2020 and became a top-level Apache project in February 2023. Yu and Sarwat went on to found Wherobots.

What changes at scale

• Volume and layout. At billions of rows, file layout sets speed. Columnar files and table formats keep per-file statistics, so a spatial filter skips most files.
• Two data models in one query. The walkthrough samples two rasters at 6,922 vector features. With tools built for one model, teams export, convert, and re-import data between them. One SQL engine that reads COG tiles on demand removes that step, as the out-of-database raster architecture post explains.
• Skew. Cities hold most buildings and places, so equal-area partitions overload a few workers. GeoSpark's R-tree, quadtree, and KDB-tree (k-dimensional B-tree) partitioners split dense areas into smaller pieces.
• CRS errors grow with coverage. A local projection that works for one city distorts a continent, so area and distance at scale need geodesic functions.

Open problems

• Identity across datasets. The same building appears in Overture, OpenStreetMap, a parcel file, and a lidar survey with different IDs and shapes. Stable identifiers such as Overture's GERS reduce this matching work, called conflation in the ontology article.
• Quality of volunteered and AI-derived data. Goodchild's 2007 paper raised accuracy as the open question for volunteered data. It now applies to model-made footprints and land cover maps too. WorldCover 2021 reports an overall accuracy of 76.7%, so a class read at one pixel is an estimate.
• Time. The walkthrough joins a 2021 land cover map to current buildings. Trajectory and change data need indexes over time as well as space.
• Rasters and vectors at different resolutions. A 30 m elevation cell covers several small buildings, and a 10 m land cover pixel can split a house. Sampling at a centroid, averaging over a footprint, and weighting by area give different answers.

Conclusions

The Lake Havasu City walkthrough shows that geospatial data comes as vector features and raster grids on shared coordinates, and that one question about a place can draw on both.

1. Geospatial data comes in two models, and a question about a place often needs both. One query read a land cover class and an elevation for all 6,922 buildings and places. If a question mixes objects and a surface, run it in an engine that holds vector and raster data together.
2. Geospatial layers match up only when they share a coordinate reference system. At scale, measure area and length on the curved Earth, as ST_AreaSpheroid and ST_LengthSpheroid do in the inventory query.
3. A join between geospatial layers can drop or duplicate features, so check totals. The distinct IDs summed to 4,844 buildings and 2,078 places, so the join lost no feature. Distinct counts cannot show a feature joined to two tiles, which would count its elevation twice. Compare joined rows with distinct IDs, as the tract check in geospatial analysis does.
4. When two geospatial layers disagree about one place, treat the gap as a fact about the data. 330 buildings sit on bare pixels and 7 on water in a 2021 map with 76.7% overall accuracy. Check dates and resolution first. Treat a class read at one pixel as an estimate.
5. Geospatial data at global scale needs formats that let a query skip files. The inventory query filtered three global Overture layers to one town in 33.8 seconds. Columnar files carry per-file statistics, so the engine skipped files outside the box. The rasters were read as COG tiles on demand. For data read by location, prefer GeoParquet, Iceberg tables, and COG.
6. A value read from geospatial data depends on how it was sampled, so state the method. The walkthrough reads each raster at a building's centroid. Averaging over the footprint gives different answers for small buildings.

Related terms

• Geospatial database: geometry types, a spatial index and spatial SQL in one system
• Geospatial analysis: the methods that turn geospatial data into answers
• Raster vs vector data: the two main data models compared
• Coordinate reference system: how coordinates map to places on Earth
• GeoTIFF: the georeferenced raster format behind COGs
• H3: a hexagonal grid for indexing and joining geospatial data
• Spatial join: combining datasets by location
• Earth observation data: data collected about the Earth from space and the air
• Digital elevation model: the height raster in the walkthrough
• Land cover classification: how maps such as WorldCover are made
• Ontology: the classes and relations that give geospatial data shared meaning

Read more from Wherobots

• Spatial data, GeoParquet, and Apache Sedona
• Making Overture Maps data more efficient with GeoParquet and Apache Sedona
• El Niño 2026, atmospheric rivers, and California's burn scars: one SQL engine, two data models
• The medallion architecture for geospatial data
• Raster processing at scale: the out-of-database architecture behind WherobotsDB

Query vector and raster geospatial data together with a Wherobots free trial at cloud.wherobots.com.