Geospatial Data Pipeline
Expert in processing, optimizing, and visualizing geospatial data at scale.
When to Use
✅ Use for:
- Drone imagery processing and annotation
- GPS track analysis and visualization
- Location-based search (find nearby X)
- Map tile generation for web/mobile
- Coordinate system transformations
- Geofencing and spatial queries
- GeoJSON optimization for web
❌ NOT for:
- Simple address validation (use address APIs)
- Basic distance calculations (use Haversine formula)
- Static map embeds (use Mapbox Static API)
- Geocoding (use Nominatim or Google Geocoding API)
Technology Selection
Database: PostGIS vs MongoDB Geospatial
| Feature | PostGIS | MongoDB |
|---|
| Spatial indexes | GiST, SP-GiST | 2dsphere |
| Query language | SQL + spatial functions | Aggregation pipeline |
| Geometry types | 20+ (full OGC support) | Basic (Point, Line, Polygon) |
| Coordinate systems | 6000+ via EPSG | WGS84 only |
| Performance (10M points) | <100ms | <200ms |
| Best for | Complex spatial analysis | Document-centric apps |
Timeline:
- 2005: PostGIS 1.0 released
- 2012: MongoDB adds geospatial indexes
- 2020: PostGIS 3.0 with improved performance
- 2024: PostGIS remains gold standard for GIS workloads
Common Anti-Patterns
Anti-Pattern 1: Storing Coordinates as Strings
Novice thinking: "I'll just store lat/lon as text, it's simple"
Problem: Can't use spatial indexes, queries are slow, no validation.
Wrong approach:
// ❌ String storage, no spatial features
interface Location {
id: string;
name: string;
latitude: string; // "37.7749"
longitude: string; // "-122.4194"
}
// Linear scan for "nearby" queries
async function findNearby(lat: string, lon: string): Promise<Location[]> {
const all = await db.locations.findAll();
return all.filter(loc => {
const distance = calculateDistance(
parseFloat(lat),
parseFloat(lon),
parseFloat(loc.latitude),
parseFloat(loc.longitude)
);
return distance < 5000; // 5km
});
}
Why wrong: O(N) linear scan, no spatial index, string parsing overhead.
Correct approach:
// ✅ PostGIS GEOGRAPHY type with spatial index
CREATE TABLE locations (
id SERIAL PRIMARY KEY,
name VARCHAR(255),
location GEOGRAPHY(POINT, 4326) -- WGS84 coordinates
);
-- Spatial index (GiST)
CREATE INDEX idx_locations_geography ON locations USING GIST(location);
-- TypeScript query
async function findNearby(lat: number, lon: number, radiusMeters: number): Promise<Location[]> {
const query = `
SELECT id, name, ST_AsGeoJSON(location) as geojson
FROM locations
WHERE ST_DWithin(
location,
ST_SetSRID(ST_MakePoint($1, $2), 4326)::geography,
$3
)
ORDER BY location <-> ST_SetSRID(ST_MakePoint($1, $2), 4326)::geography
LIMIT 100
`;
return db.query(query, [lon, lat, radiusMeters]); // <10ms with index
}
Timeline context:
- 2000s: Stored lat/lon as FLOAT columns, did math in app code
- 2010s: PostGIS adoption, spatial indexes
- 2024:
GEOGRAPHY type handles Earth curvature automatically
Anti-Pattern 2: Not Using Spatial Indexes
Problem: Proximity queries do full table scans.
Wrong approach:
-- ❌ No index, sequential scan
CREATE TABLE drone_images (
id SERIAL PRIMARY KEY,
image_url VARCHAR(255),
location GEOGRAPHY(POINT, 4326)
);
-- This query scans ALL rows
SELECT * FROM drone_images
WHERE ST_DWithin(
location,
ST_SetSRID(ST_MakePoint(-122.4194, 37.7749), 4326)::geography,
1000 -- 1km
);
EXPLAIN output: Seq Scan on drone_images (cost=0.00..1234.56 rows=1 width=123)
Correct approach:
-- ✅ GiST index for spatial queries
CREATE INDEX idx_drone_images_location ON drone_images USING GIST(location);
-- Same query, now uses index
SELECT * FROM drone_images
WHERE ST_DWithin(
location,
ST_SetSRID(ST_MakePoint(-122.4194, 37.7749), 4326)::geography,
1000
);
EXPLAIN output: Bitmap Index Scan on idx_drone_images_location (cost=4.30..78.30 rows=50 width=123)
Performance impact: 10M points, 5km radius query
- Without index: 3.2 seconds (full scan)
- With GiST index: 12ms (99.6% faster)
Anti-Pattern 3: Mixing Coordinate Systems
Novice thinking: "Coordinates are just numbers, I can mix them"
Problem: Incorrect distances, misaligned map features.
Wrong approach:
// ❌ Mixing EPSG:4326 (WGS84) and EPSG:3857 (Web Mercator)
const userLocation = {
lat: 37.7749, // WGS84
lon: -122.4194
};
const droneImage = {
x: -13634876, // Web Mercator (EPSG:3857)
y: 4545684
};
// Comparing apples to oranges!
const distance = Math.sqrt(
Math.pow(userLocation.lon - droneImage.x, 2) +
Math.pow(userLocation.lat - droneImage.y, 2)
);
Result: Wildly incorrect distance (millions of "units").
Correct approach:
-- ✅ Transform to common coordinate system
SELECT ST_Distance(
ST_Transform(
ST_SetSRID(ST_MakePoint(-122.4194, 37.7749), 4326), -- WGS84
3857 -- Transform to Web Mercator
),
ST_SetSRID(ST_MakePoint(-13634876, 4545684), 3857) -- Already Web Mercator
) AS distance_meters;
Or better: Always store in one system (WGS84), transform on display only.
Timeline:
- 2005: Web Mercator (EPSG:3857) introduced by Google Maps
- 2010: Confusion peaks as apps mix WGS84 data with Web Mercator tiles
- 2024: Best practice: Store WGS84, transform to 3857 only for tile rendering
Anti-Pattern 4: Loading Huge GeoJSON Files
Problem: 50MB GeoJSON file crashes browser.
Wrong approach:
// ❌ Load entire file into memory
const geoJson = await fetch('/drone-survey-data.geojson').then(r => r.json());
// 50MB of GeoJSON = browser freeze
map.addSource('drone-data', {
type: 'geojson',
data: geoJson // All 10,000 polygons loaded at once
});
Correct approach 1: Vector tiles (pre-chunked)
// ✅ Serve as vector tiles (MBTiles or PMTiles)
map.addSource('drone-data', {
type: 'vector',
tiles: ['https://api.example.com/tiles/{z}/{x}/{y}.pbf'],
minzoom: 10,
maxzoom: 18
});
// Browser only loads visible tiles
Correct approach 2: GeoJSON simplification + chunking
# Simplify geometry (reduce points)
npm install -g @mapbox/geojson-precision
geojson-precision -p 5 input.geojson output.geojson
# Split into tiles
npm install -g geojson-vt
# Generate tiles programmatically (see scripts/tile_generator.ts)
Correct approach 3: Server-side filtering
// ✅ Only fetch visible bounds
async function fetchVisibleFeatures(bounds: Bounds): Promise<GeoJSON> {
const response = await fetch(
`/api/features?bbox=${bounds.west},${bounds.south},${bounds.east},${bounds.north}`
);
return response.json();
}
map.on('moveend', async () => {
const bounds = map.getBounds();
const geojson = await fetchVisibleFeatures(bounds);
map.getSource('dynamic-data').setData(geojson);
});
Anti-Pattern 5: Euclidean Distance on Spherical Earth
Novice thinking: "Distance is just Pythagorean theorem"
Problem: Incorrect at scale, worse near poles.
Wrong approach:
// ❌ Flat Earth distance (wrong!)
function distanceKm(lat1: number, lon1: number, lat2: number, lon2: number): number {
const dx = lon2 - lon1;
const dy = lat2 - lat1;
return Math.sqrt(dx * dx + dy * dy) * 111.32; // 111.32 km/degree (WRONG)
}
// Example: San Francisco to New York
const distance = distanceKm(37.7749, -122.4194, 40.7128, -74.0060);
// Returns: ~55 km (WRONG! Actual: ~4,130 km)
Why wrong: Earth is a sphere, not a flat plane.
Correct approach 1: Haversine formula (great circle distance)
// ✅ Haversine formula (spherical Earth)
function haversineKm(lat1: number, lon1: number, lat2: number, lon2: number): number {
const R = 6371; // Earth radius in km
const dLat = toRadians(lat2 - lat1);
const dLon = toRadians(lon2 - lon1);
const a =
Math.sin(dLat / 2) * Math.sin(dLat / 2) +
Math.cos(toRadians(lat1)) * Math.cos(toRadians(lat2)) *
Math.sin(dLon / 2) * Math.sin(dLon / 2);
const c = 2 * Math.atan2(Math.sqrt(a), Math.sqrt(1 - a));
return R * c;
}
// San Francisco to New York
const distance = haversineKm(37.7749, -122.4194, 40.7128, -74.0060);
// Returns: ~4,130 km ✅
Correct approach 2: PostGIS (handles curvature automatically)
-- ✅ PostGIS ST_Distance with GEOGRAPHY
SELECT ST_Distance(
ST_SetSRID(ST_MakePoint(-122.4194, 37.7749), 4326)::geography,
ST_SetSRID(ST_MakePoint(-74.0060, 40.7128), 4326)::geography
) / 1000 AS distance_km;
-- Returns: 4130.137 km ✅
Accuracy comparison:
| Method | SF to NYC | Error |
|---|
| Euclidean (flat) | 55 km | 98.7% wrong |
| Haversine (sphere) | 4,130 km | ✅ Correct |
| PostGIS (ellipsoid) | 4,135 km | Most accurate |
Production Checklist
□ PostGIS extension installed and spatial indexes created
□ All coordinates stored in consistent SRID (recommend: 4326)
□ GeoJSON files optimized (<1MB) or served as vector tiles
□ Coordinate transformations use ST_Transform, not manual math
□ Distance calculations use ST_Distance with GEOGRAPHY type
□ Bounding box queries use ST_MakeEnvelope + ST_Intersects
□ Large geometries chunked (not >100KB per feature)
□ Map tiles pre-generated for common zoom levels
□ CORS configured for tile servers
□ Rate limiting on geocoding/reverse geocoding endpoints
When to Use vs Avoid
| Scenario | Appropriate? |
|---|
| Drone imagery annotation and search | ✅ Yes - process survey data |
| GPS track visualization | ✅ Yes - optimize paths |
| Find nearest coffee shops | ✅ Yes - spatial queries |
| Jurisdiction boundary lookups | ✅ Yes - point-in-polygon |
| Simple address autocomplete | ❌ No - use Mapbox/Google |
| Embed static map on page | ❌ No - use Static API |
| Geocode single address | ❌ No - use geocoding API |
References
/references/coordinate-systems.md - EPSG codes, transformations, Web Mercator vs WGS84/references/postgis-guide.md - PostGIS setup, spatial indexes, common queries/references/geojson-optimization.md - Simplification, chunking, vector tiles
Scripts
scripts/geospatial_processor.ts - Process drone imagery, GPS tracks, GeoJSON validationscripts/tile_generator.ts - Generate vector tiles (MBTiles/PMTiles) from GeoJSON
This skill guides: Geospatial data | PostGIS | GeoJSON | Map tiles | Coordinate systems | Drone data processing | Spatial queries