VECTIQ
← GeoAI & Remote Sensing

Vietnam Digital Map & Geospatial Data

Satellite imagery shows where a building stands; it does not say what the street is called, which way it runs or what the speed limit is. We build the platform that closes that gap: imagery catalogued by period and served as basemap tiles, a map database where every road, building and POI keeps its id and its history, an editor built on Meta's open-source Rapid where AI suggestions wait for a person to confirm them, and a task system that splits review work between staff, paid contractors and volunteers. The whole stack can be deployed in-country, and the basemap shows Hoàng Sa and Trường Sa as Vietnamese territory.

Smart City & TransportEnvironment & ResourcesEnterprise
Vietnam Digital Map & Geospatial Data

Technical capabilities

01

Imagery by Source and Period

Sentinel-2 scenes, cloud-free composites by year, quarter or month, and purchased high-resolution imagery sit in one STAC catalogue. Users pick a source, then a period, the way they pick a background in an OSM editor.

02

Basemap Tiles at Every Zoom

Low zooms are pre-rendered to PMTiles, high zooms are rendered on demand by TiTiler, and Redis caches both. The vector basemap is served by Martin from our own PostGIS, with self-hosted fonts.

03

A Map Database With History

Roads, buildings, water and POIs are versioned objects with stable ids. Edits arrive as changesets with optimistic locking, releases are snapshots, and the map can be queried as it was on any date.

04

AI-assisted Editor

The Geo Editor is built on Rapid, Meta's open-source OSM editor, and talks to our map database through an OSM API facade. Model outputs show up as orange dashed candidates until a person accepts or deletes them.

05

Tasks With Consensus

Work is cut into small tasks, locked while someone is on them, seeded with gold questions and decided by weighted majority or Dawid–Skene. Decided answers go back into the map and out as training labels.

06

Three Contributor Tiers

Volunteers answer micro-tasks. Paid contractors edit the map, but only inside the H3 cells MapOps assigned them. MapOps run projects, payouts and releases. Rewards, payouts and a leaderboard are part of the system.

07

Map Quality per H3 Cell

Completeness of roads, buildings and POIs is scored cell by cell against Overture and other sources, so new tasks go where the map is weakest rather than where someone happens to look.

08

Signs From Street-level Imagery

In development: traffic signs detected in street-level imagery, placed on the road despite phone GPS noise, checked by a reviewer, then resolved into speed-limit zones along each road and direction.

Processing pipeline

01

Sources

Sentinel-2 / OSM / Overture

02

Catalogue

pgSTAC / COG

03

Tiles

TiTiler / Martin

04

AI Candidates

U-Net / YOLO

05

Tasks & Editor

Dawid–Skene / Rapid

06

Map Release

Map DB / .osm.pbf

Deployment scenarios

Transport & Navigation

A Road Network You Can Route On

Routing needs roads that connect, carry names and keep their ids between releases. The platform builds road topology from the map, finds disconnected islands, proposes names for unnamed lanes, and exports every release in a format a routing engine loads directly.

  • Road topology with shared junctions and island detection
  • Name candidates for unnamed ngõ, kiệt and hẻm, confirmed by people
  • Split and merge that keep object ids, with lineage
  • OSM sync that never overwrites human corrections
  • .osm.pbf export attached to every release
TopologyOSM PBFRoad NamesReleases
Environment & Land

Watch an Area Through Time

Draw or upload an area and every Sentinel-2 scene over it becomes a timeline: true colour or vegetation and water indices, two dates side by side or under a swipe, and an index series over the area or at a single point. Results leave as data, not screenshots.

  • Areas drawn on the map or uploaded as GeoJSON or zipped Shapefile
  • True colour, NDVI, NDWI and EVI with notes on how to read them
  • Swipe and side-by-side comparison with the change measured
  • Index series over the whole area or at one point
  • Area downloads as GeoTIFF in WGS 84, VN-2000 or UTM
Sentinel-2NDVITime SeriesVN-2000
Enterprises & Data Owners

Purchased Imagery, Handled Properly

Commercial imagery is expensive and comes with a licence. Each delivery is archived untouched with checksums before anything else happens, then orthorectified, pansharpened and converted to COGs, tagged with its licence and order number, and shown only to the tenants allowed to see it.

  • Untouched archive of every delivery with a sha256 manifest
  • RPC orthorectification, pansharpening and COG output
  • Licence and order number recorded on every image
  • Private layers per tenant, read with an API key
  • View-only layers; original files are never offered for download
COGSTACLicensingMulti-tenant
Traffic Authorities

Speed Limits on the Map, Not Just Signs

A driver does not ask where the sign is; they ask what the limit is on this stretch. Our street-level pipeline, still in development, places signs from imagery, ties each one to a road and a direction, and resolves them into zones that end where the traffic rules say they end.

  • Sign detection with classes mapped to QCVN 41:2019 codes
  • Localization that tolerates phone GPS noise and improves with each pass
  • Review queue showing the photo evidence beside each sign
  • Speed-limit zones per road and direction, exported as OSM-style tags
  • Supplementary plates (vehicle type, time window) on the roadmap
QCVN 41AutoTSReviewmaxspeed

Tech stack

PostGISpgSTACTiTilerMartinPMTilesMapLibreRapidH3FastAPISentinel-2MinIOHelm

Related projects