Spatial data models and spatial databases

In short

A spatial data model is the formal way the continuous real world is simplified into discrete, computer-storable objects — points, lines, polygons, cells or surfaces — together with their attributes and their spatial relationships. A spatial database is a DBMS (usually relational) extended with geometry data types, spatial indexes and spatial operators so that location can be queried in SQL alongside ordinary attributes. Getting the model right decides whether a GIS can answer questions about adjacency, connectivity and containment at all, and whether a national database of millions of parcels performs. This topic is the conceptual backbone of GIS papers in B.Tech Geoinformatics, carries marks in GATE Geomatics (GE) under "GIS: data models, data structures, topology, database queries", and is tested in technical rounds at Survey of India, NRSC/ISRO, NIC, NSDI/state GIS cells, RMSI, Esri India, Genesys, CyberTech, MapmyIndia (Mappls), Tata Consultancy Services geospatial units and smart-city/AMRUT consultancies.

Explained

Levels of abstraction: from reality to bits

• Reality → conceptual model → logical model → physical structure → file/DBMS. The conceptual stage asks *what* entities exist; the logical stage fixes vector/raster/object; the physical stage fixes the file format and index.
• Two world views:
  • Discrete object (entity) view — the world is empty space occupied by countable things with sharp boundaries (wells, roads, plots). Maps naturally to vector.
  • Continuous field view — every location has a value of some attribute (elevation, rainfall, reflectance, LST). Maps naturally to raster or to a field model such as TIN.
• Spatial data has three components: *geometry* (where), *attributes* (what), and *time* (when) — plus *topology* (how objects relate). Loss of any one makes analysis impossible.
• Nature of measurement scale for attributes: nominal (land-use class), ordinal (road hierarchy), interval (temperature °C), ratio (population, area) — determines which analyses and which raster resampling (nearest neighbour for categorical, bilinear/cubic for continuous) are legitimate.

Vector data model

• Geometry is built from coordinate pairs (x, y) or triples (x, y, z) referenced to a datum and projection (in India, WGS-84/ITRF via NNRMS standards, or the older Everest spheroid with polyconic Survey of India topo sheets, Lambert conformal conic at 1:1 M).
• Primitives: 0-D point/node; 1-D line/arc/polyline/chain (an ordered list of vertices between a *from-node* and a *to-node*); 2-D polygon/area (a closed ring; interior rings = holes); 2.5-D surface; multi-geometries and geometry collections.
• Spaghetti (non-topological) model: every feature stores its own complete coordinate list. Simple and fast to draw, but shared boundaries are digitised twice → double digitising, sliver polygons, mismatch, and no adjacency information. The ESRI shapefile is essentially a spaghetti model.
• Topological (arc-node / georelational) model: coordinates stored once per arc; relationships held in tables:
  • *Arc-coordinate list* — vertices of each arc.
  • *Node-arc list* — which arcs meet at a node → connectivity (feature attributes themselves live in the AAT/PAT attribute tables).
  • *Left-right polygon table* — polygon on each side of an arc → adjacency/contiguity.
  • *Polygon-arc list* — arcs that bound each polygon → area definition/containment, with a negative arc id for islands and the universe polygon (id 1) for the outside world.
  • Classic implementations: ARC/INFO coverage, DIME and TIGER files (US Census), POLYVRT.
• Topological consistency rules enforced at build/clean: no dangling nodes (undershoot), no overshoots, polygons closed, label point in every polygon, no duplicate arcs. Tolerances used: *fuzzy/cluster tolerance*, *dangle length*, *snapping distance*.
• Advantages of vector: compact, accurate boundaries, true topology, excellent for networks, discrete entities, cadastre and cartography, scale-independent plotting.
• Limitations: complex data structure, overlay is computationally heavy, continuous surfaces awkward, simulation and image algebra difficult.

Raster data model

• Space is tessellated into a regular grid of cells (pixels); each cell holds one value; position is implicit from row, column, cell size and the origin/geo-transform (upper-left x, y, cell size x, cell size y, rotation).
• Tessellation types: square (universal — simple neighbourhood, but 4 vs 8 connectivity ambiguity), triangular, hexagonal (equidistant neighbours, used in some ecological and H3-style grids).
• Cell value semantics: *centre-point* (value at cell centre), *dominant/majority* (most-area class), *presence-absence*, *percentage occurrence*.
• Resolution vs file size: halving the cell size multiplies cell count — and storage — by four. Mixed pixels at class boundaries are the fundamental raster error.
• Compression / storage structures:
  • Run-length encoding (RLE) — store (value, run length) per row; very effective for class maps.
  • Chain codes — trace a boundary as directions; good for single regions.
  • Block codes — 2-D extension of RLE using square blocks (medial-axis transform).
  • Quadtree — recursively divide the grid into four quadrants until homogeneous; variable resolution, also a spatial index.
• Advantages: simple structure, natural for remote-sensing imagery and DEMs, fast map algebra and overlay, easy modelling/simulation, direct input from satellites (Resourcesat LISS-III/IV, Cartosat, Sentinel-2).
• Limitations: large volumes, boundaries blocky, topology must be inferred, coordinate transformation requires resampling (which alters values), poor for networks.
• Formats: GeoTIFF/BigTIFF, Cloud Optimized GeoTIFF (COG), Erdas IMG, NetCDF/HDF (multidimensional), ASCII grid; GDAL is the de-facto translator library.

Surface, network, object and time models

• DEM/DTM/DSM: DEM = bare-earth elevation grid; DSM includes buildings and canopy; DTM adds breaklines/mass points. Indian sources: CartoDEM (Cartosat-1, ~30 m) from NRSC/Bhuvan, SRTM, ASTER GDEM, ALOS PALSAR, and high-resolution LiDAR/UAV DSMs.
• TIN (Triangulated Irregular Network): irregular mass points joined into non-overlapping triangles, usually by Delaunay triangulation (maximises the minimum angle; each triangle's circumcircle contains no other point — the dual of the Voronoi/Thiessen diagram). Stores slope, aspect and breaklines efficiently; variable density follows terrain complexity; harder to overlay than a grid.
• Grid vs TIN vs contour vs point cloud: four interchangeable surface representations; LAS/LAZ point clouds are now first-class (LiDAR, photogrammetric dense matching).
• Network model: a topological graph of links (edges) and junctions (nodes) with impedance, turn tables, one-way flags, connectivity rules; used for routing, utilities (water, sewer, power) and transport.
• Object-oriented model: features as objects with geometry, attributes, behaviour (validation rules, relationships, subtypes, domains) — basis of the Esri geodatabase (personal/file/enterprise) and of OGC's feature model.
• Spatio-temporal models: *snapshot* (a layer per date), *space-time composite*, *event/amendment-based* (store only changes — used for cadastral mutations), *object-with-lifespan*. Bitemporal databases keep both *valid time* and *transaction time*.
• 3-D models: 2.5-D surface vs true 3-D (voxels, B-rep solids), CityGML LoD 0-4, BIM/GIS integration for smart cities.

Spatial databases: architecture, types, indexing

• Architecture evolution:
  1. Hybrid/georelational — geometry in proprietary files, attributes in an RDBMS, joined by a feature ID (ARC/INFO coverages, early shapefile + dBASE).
  2. Integrated/extended relational (object-relational) — geometry is a *column type* inside the DBMS: PostgreSQL + PostGIS, Oracle Spatial (SDO_GEOMETRY), SQL Server geometry/geography, SpatiaLite/GeoPackage, MySQL spatial.
  3. NoSQL / cloud — MongoDB 2dsphere, Elasticsearch geo types, BigQuery GIS; useful for very large, loosely structured location streams.
• Why a DBMS at all: multi-user concurrency, ACID transactions, versioning, roles and permissions, referential integrity, backup/recovery, SQL and set-based querying, data independence — all absent in flat files.
• Relational essentials examined: entity-relationship modelling, primary and foreign keys, normalisation to 3NF (remove repeating groups, partial and transitive dependencies), cardinality (1:1, 1:M, M:N via a junction table), joins, views, and the spatial join (join on a spatial predicate rather than a key).
• Why ordinary indexes fail: a B-tree orders one dimension; geometry is 2-D and extents overlap. Spatial access methods used instead:
  • R-tree / R*-tree — hierarchy of minimum bounding rectangles (MBRs); PostGIS GiST and Oracle default.
  • Quadtree / grid (fixed or multi-level grid) — Oracle's older option, SQL Server's grid index.
  • Space-filling curves — Morton/Z-order, Hilbert curve, geohash; map 2-D to 1-D so a B-tree works.
  • k-d tree, BSP tree, cell/grid files — research and in-memory use.
• Two-step query processing: (1) filter step — index returns candidates whose MBRs satisfy the predicate (cheap, approximate); (2) refinement step — exact geometry computation on candidates. This is why && (bbox overlap) is fast and ST_Intersects is slower.
• Topological predicates — the 9-intersection and DE-9IM model: relations between interior, boundary and exterior of two geometries give a 3×3 matrix; named predicates are *equals, disjoint, touches, contains, within, covers, overlaps, crosses, intersects*.
• Standards: OGC/ISO 19125 Simple Features Access (SFA, Part 1 common architecture, Part 2 SQL) defines the geometry hierarchy and functions; ISO 19107 spatial schema; ISO 19115/19115-1 metadata; ISO 19136 = GML; serialisations WKT/WKB, EWKB, GeoJSON (RFC 7946, always in WGS-84 lon-lat), GeoPackage (OGC, SQLite-based), TopoJSON; EPSG/SRID registry for CRS (EPSG:4326 geographic WGS-84, EPSG:3857 Web Mercator, EPSG:7755 etc. for Indian zones; older Indian grids also carry EPSG codes under the Everest spheroid).
• Raster in the database: PostGIS raster type, Oracle GeoRaster, or out-of-DB tiles with a catalogue table; tiling + overviews (pyramids) are essential for performance.

Data quality, metadata and the Indian policy frame

• Components of spatial data quality: positional accuracy (RMSE), attribute accuracy, logical consistency (topology obeyed), completeness, lineage, temporal accuracy, semantic accuracy, resolution/scale.
• Error sources: source map scale and generalisation, digitising error, datum/projection mismatch, edge-matching across sheets, classification error, rasterisation, interpolation; errors propagate and compound through overlays.
• Metadata answers who, what, where, when, how accurate, what CRS, what use constraints — mandated in India through NSDI metadata standards aligned to ISO 19115, and published via NSDI/Bhuvan/India-WRIS portals and the NDSAP (National Data Sharing and Accessibility Policy, 2012) with data.gov.in.
• Indian standards and practice:
  • NNRMS standards (ISRO/DOS) for thematic mapping, layer naming, accuracy and scale-vs-resolution mapping.
  • National Map Policy, 2005 — Open Series Maps (OSM) on WGS-84/UTM for public use, Defence Series Maps restricted; the Guidelines for acquiring and producing geospatial data, 15 February 2021 deregulated Indian-entity collection and processing, removed prior approval for most mapping, and set threshold values for finer-than-permitted accuracy regulation for foreign entities. The National Geospatial Policy, 2022 targets a national integrated data and knowledge framework, high-resolution topographic and bathymetric data and a Geodetic CORS network.
  • Digital India Land Records Modernization Programme (DILRMP) — cadastral maps linked to Records of Rights, ULPIN/Bhu-Aadhaar 14-digit parcel ID, SVAMITVA drone survey of abadi land; state portals (Bhulekh, Maha Bhu-Naksha, Dharani, e-Dhara).
  • AMRUT GIS master plans, PM Gati Shakti National Master Plan (multi-layer NMP on Bhuvan), Survey of India CORS network.
  • DPDP Act, 2023 governs personal data; its Rules of 2025 place the Data Fiduciary duties (notice, consent, breach intimation, erasure) in force only from 13 May 2027 — until then IT Act s.43A and the SPDI Rules apply, which matters when a GIS holds address or household-level personal data.
  • Copyright and licensing: government geospatial data increasingly under open licences (GODL-India); OpenStreetMap under ODbL — licence terms must be recorded in metadata.

Key points

• Spatial data model - the abstraction that turns continuous reality into storable geometry plus attributes plus topology.
• Entity vs field view - discrete countable objects → vector; value-everywhere phenomena → raster/TIN.
• Vector primitives - point (0-D), line/arc (1-D), polygon (2-D), surface (2.5-D); vertices, nodes, rings.
• Spaghetti model - independent coordinate strings per feature; no topology; causes slivers and double digitising (the shapefile).
• Arc-node topology - connectivity (node-arc), adjacency (left-right polygon), containment (polygon-arc); coordinates stored once.
• Raster geo-transform - origin x, y plus cell size defines position; row/column order is implicit, so no coordinates are stored per cell.
• Quarter-rule - halving cell size quadruples the number of cells and the storage.
• Raster compression - RLE, chain code, block code, quadtree; quadtree is both a compression and an index.
• TIN - Delaunay triangles over irregular mass points; dual of the Thiessen/Voronoi diagram; efficient on variable terrain.
• DEM vs DSM vs DTM - bare earth, top-of-feature, and bare earth with breaklines; CartoDEM ~30 m is the Indian national grid.
• Database architectures - hybrid georelational, integrated object-relational (PostGIS, Oracle Spatial), NoSQL/cloud.
• Spatial index - R-tree (MBRs), quadtree/grid, Z-order/Hilbert/geohash; a plain B-tree cannot order two dimensions.
• Two-step query - MBR filter from the index, then exact geometry refinement.
• DE-9IM - 3×3 interior/boundary/exterior intersection matrix behind touches, contains, overlaps, crosses.
• OGC Simple Features (ISO 19125) - standard geometry types and SQL functions; WKT/WKB, GML, GeoJSON, GeoPackage are the serialisations.
• SRID/EPSG - every geometry column needs a declared CRS; 4326 is lon-lat WGS-84, 3857 is Web Mercator.
• Quality components - positional, attribute, logical consistency, completeness, lineage, temporal; metadata per ISO 19115/NSDI.
• Indian frame - National Map Policy 2005, the 2021 geospatial guidelines, National Geospatial Policy 2022, NNRMS standards, DILRMP/ULPIN, NDSAP.

Formulas and examples

Example 1 — raster volume and the resolution penalty. A 10 km × 10 km LISS-IV tile is to be stored at 5 m resolution, 4 spectral bands, 16-bit (2 bytes) per cell.

• Rows = 10,000/5 = 2,000; columns = 2,000 → n = 4,000,000 cells.
• S = 4,000,000 × 2 × 4 = 32,000,000 bytes ≈ 32 MB (≈30.5 MiB).
• Resample to 2.5 m: cells = 16,000,000 → S = 16,000,000 × 2 × 4 = 128 MB, i.e. 4× for halving the cell size.
• The same area as a vector land-use layer with 1,200 polygons and ~40 vertices each ≈ 1,200 × 40 × 16 bytes ≈ 0.77 MB — about 40 times smaller, which is the vector-raster storage trade-off for class maps.

Example 2 — run-length encoding gain. A single-band 8-bit land-use raster, 1,000 × 1,000 cells, averages 25 runs per row.

• Uncompressed = 1,000 × 1,000 × 1 = 1,000,000 bytes.
• RLE: 25 runs × 1,000 rows = 25,000 runs; each run = 1 byte value + 2 bytes run length = 3 bytes → 75,000 bytes.
• CR = 1,000,000 / 75,000 ≈ 13.3 : 1. If the map were salt-and-pepper noise with 500 runs per row, RLE = 500 × 1,000 × 3 = 1,500,000 bytes — larger than the original, which is why RLE suits blocky class maps only.

Example 3 — topology check with Euler's formula. A cleaned polygon coverage reports 148 nodes, 236 arcs and 90 polygons (including the universe polygon).

• N − A + P = 148 − 236 + 90 = 2 → the planar graph is consistent.
• If the build reported 89 polygons instead, 148 − 236 + 89 = 1 → one polygon failed to close (an unsnapped arc or a missing label point); the dangling-node report must be re-run with a larger snap tolerance.

Example 4 — TIN size and a facet slope. A UAV survey gives 1,500 mass points, of which 60 lie on the convex hull.

• T = 2(1,500) − 60 − 2 = 2,938 triangles; E = 3(1,500) − 60 − 3 = 4,437 edges.
• One facet has A(0, 0, 100.0), B(30, 0, 101.8), C(0, 40, 98.4) m.
• ∂z/∂x = (101.8 − 100.0)/30 = 0.0600; ∂z/∂y = (98.4 − 100.0)/40 = −0.0400.
• Gradient magnitude = √(0.0600² + 0.0400²) = √(0.0036 + 0.0016) = √0.0052 = 0.07211.
• Slope = arctan(0.07211) = 4.125°, i.e. 7.21%; aspect = direction of steepest descent, here towards +y and −x (north-west quadrant in a standard grid, exact bearing from arctan of the components).

Example 5 — spatial index and query cost. A PostGIS table holds 2,000,000 building footprints for a city; a query asks for buildings within 200 m of a metro alignment, and the search window retrieves 3,500 candidate MBRs.

• Sequential scan: 2,000,000 exact geometry tests.
• GiST R-tree: ~log-depth descent + 3,500 MBR hits → 3,500 exact ST_DWithin tests, i.e. about 0.175% of the work — roughly a 570× reduction in exact geometry computations.

CREATE TABLE buildings (
  gid      serial PRIMARY KEY,
  ulpin    varchar(14),
  use_code varchar(10),
  geom     geometry(MultiPolygon, 32644)   -- UTM 44N, WGS-84
);
CREATE INDEX buildings_geom_gix ON buildings USING GIST (geom);
ANALYZE buildings;

-- spatial join with a buffer, areas in m² because the CRS is projected
SELECT b.ulpin, ROUND(ST_Area(b.geom)::numeric, 2) AS area_m2
FROM   buildings b
JOIN   metro_line m ON ST_DWithin(b.geom, m.geom, 200)
WHERE  b.use_code = 'RES';

-- topology/validity audit before loading into production
SELECT gid, ST_IsValidReason(geom) FROM buildings WHERE NOT ST_IsValid(geom);

Where it is used

• Land records and cadastre: DILRMP/SVAMITVA parcel databases with ULPIN, state Bhu-Naksha portals; topologically clean polygon layers are mandatory because a sliver is a disputed boundary. Roles: GIS analyst, cadastral data QC, database administrator in NIC and state revenue departments.
• Utilities and smart cities: water, sewer, gas and electricity network geodatabases under AMRUT and Smart Cities Mission; network models drive valve-isolation and outage tracing for discoms and municipal corporations.
• Remote sensing and resource mapping: NRSC/ISRO raster archives, Bhuvan, Bharatmaps, India-WRIS, FSI forest cover, CartoDEM; raster models and tiled pyramids underpin Bhuvan's services.
• Navigation and location products: MapmyIndia/Mappls, Ola/Uber, Zomato/Swiggy and Delhivery route databases — network topology, geohash/H3 bucketing and PostGIS or NoSQL geo-indexes at very high query rates.
• Infrastructure and consultancy: PM Gati Shakti layers, NHAI alignment and ROW databases, railway and pipeline asset databases, RMSI/Esri India/Genesys/CyberTech production work; roles include spatial database developer, data modeller, geospatial back-end engineer.
• Disaster management and defence: NDMA/state SDMA hazard and exposure databases, flood inundation rasters, Survey of India topographic and OSM/DSM series — spatio-temporal models track change between events.

Interview questions

• Q: What is the difference between a spatial data model and a spatial data structure? A: The model is the conceptual abstraction (vector, raster, TIN, network) of how reality is represented; the structure is how that model is physically organised in storage — arc-node tables, run-length encoding, an R-tree.
• Q: A shapefile and a coverage both hold polygons — what is the real difference? A: A shapefile is spaghetti: each polygon carries its own full boundary, so shared edges are duplicated; a coverage stores each arc once with node and left-right polygon tables, so adjacency and connectivity are explicit.
• Q: Why can't a normal B-tree index serve a geometry column? A: A B-tree sorts values along one dimension, but geometries occupy two or three dimensions and their extents overlap, so there is no single ordering that keeps spatially near objects near in the index; R-trees or space-filling curves are used instead.
• Q: How does an R-tree actually speed up ST_Intersects? A: It indexes minimum bounding rectangles hierarchically, so the filter step returns only candidate geometries whose MBRs overlap the query window, and the costly exact geometry test runs on that short list.
• Q: What is DE-9IM and where would you quote it? A: It is the dimensionally extended 9-intersection model — a 3×3 matrix of intersections between the interior, boundary and exterior of two geometries — and it formally defines predicates like touches, contains, overlaps and crosses in OGC Simple Features.
• Q: When would you prefer a TIN over a raster DEM? A: When terrain complexity varies and breaklines matter — ridges, embankments, road edges — because a TIN puts points only where needed and honours linear features, whereas a grid wastes cells on flat ground and smooths breaks.
• Q: What goes wrong if the SRID is unknown or mismatched? A: Spatial predicates fail or return nonsense, distances and areas come out in degrees instead of metres, and overlays silently misregister; always set the SRID and transform both layers to a common projected CRS before measuring.
• Q: Explain why halving raster cell size is expensive. A: The cell count grows as the inverse square of cell size, so halving it quadruples cells, storage, I/O and the cost of every map-algebra pass.
• Q: A polygon build reports a non-zero Euler check and 12 dangling nodes — what do you do? A: Treat it as a topology failure: inspect undershoots and overshoots, raise the snap/fuzzy tolerance carefully, snap or extend the offending arcs, re-add missing label points, then rebuild and re-validate rather than editing geometry blindly.
• Q: Is GeoJSON suitable for a production cadastral database? A: No — it is a text interchange format, fixed to WGS-84 lon-lat by RFC 7946, with no indexing, typing, concurrency or transactions; use PostGIS or a GeoPackage and export GeoJSON only for web delivery.
• Q: How would you model a parcel that is subdivided over time? A: Use an event or amendment-based spatio-temporal model: keep the parent parcel with a valid-time end date, create child parcels with start dates and a lineage/parent-id link, so the mutation history is auditable — exactly what land-record systems with ULPIN need.
• Q: Which Indian policy governs how finely a private company may collect geospatial data? A: The Department of Science & Technology Guidelines of 15 February 2021 liberalised collection, storage and processing for Indian entities and removed prior approvals for most mapping, with threshold accuracies regulated mainly for foreign entities; the National Geospatial Policy, 2022 sets the longer-term national framework.

Common mistakes

• Calling the shapefile "topological" — it stores no adjacency or connectivity, which is why slivers and double-digitised boundaries appear after joining sheets.
• Computing area or distance on geometries in EPSG:4326 and reporting the answer in metres; degrees must first be projected (UTM 42N-46N or an Indian equal-area/conformal CRS).
• Confusing DEM, DTM and DSM, and then subtracting the wrong pair when extracting building or canopy heights.
• Choosing bilinear or cubic resampling for a categorical layer (land use, soil class) — it invents class values; nearest neighbour is the only valid choice.
• Forgetting that spatial indexes need to be created *and* the statistics refreshed (ANALYZE/DBMS_STATS); a built index the planner does not know about is unused.
• Assuming every raster compression saves space — RLE and quadtrees expand fragmented, noisy rasters.
• Mixing up the Delaunay triangulation with its dual Thiessen/Voronoi polygons, and quoting the wrong triangle count or neighbourhood definition.
• Loading attributes without normalising — repeating the same owner or class name in thousands of rows instead of a lookup table with a foreign key, then losing integrity on update.

Related topics

Spatial analysis and geostatistics · GIS software and web GIS: QGIS, ArcGIS and GeoServer · Map projections, datums and coordinate systems · Remote sensing and digital image processing · GNSS and modern surveying · Photogrammetry, LiDAR and DEM generation