Vector vs raster, coordinate systems, and the file formats — the fundamentals I picked up building the interactive geological map of India.
Dealing with location information is a big part of computer engineering — yet in twelve years as a software engineer, I’d never had to work with large amounts of geospatial data. Building the interactive geological map of India finally forced me to learn the fundamentals for the first time, so I’m writing them down here in case you’d like to read them too.
It turns out all geodata is one of two things.
Vector — shapes stored as coordinates
A point is one coordinate [lng, lat]; a line is a list of them; a polygon is a closed loop. A river is a line, a lake is a polygon, a city is a point. Vectors are exact, scale infinitely without blurring, and carry attributes — a polygon can say {name: "Thar", area: 200000}. This is what pins and outlines are.
Concretely, a vector feature is just geometry plus properties. Here’s the Thar Desert as a polygon, in GeoJSON:
{
"type": "Feature",
"properties": { "name": "Thar Desert", "area_km2": 200000 },
"geometry": {
"type": "Polygon",
"coordinates": [
[[70.1, 27.9], [72.4, 27.2], [71.8, 25.6], [69.9, 26.3], [70.1, 27.9]]
]
}
}
That’s the whole trick. A point is a single [lng, lat]; a line (a river) is a list of them; a polygon (a lake, a desert) is a list that closes back to its first coordinate. The geometry is literally arrays of longitude/latitude pairs, and the map library draws them.
You’ll meet vectors as these files:
| Format | What it is | Notes |
|---|---|---|
| GeoJSON | Plain JSON: geometry + properties, always WGS84 lat/lng | The web-native format. Human-readable, one file, works directly in Leaflet. What everything on my map ends up as. |
Shapefile (.shp) |
The old GIS workhorse, from Esri | Not one file — a bundle (.shp geometry + .shx index + .dbf attributes + .prj CRS). Miss one and it breaks. Binary. Convert to GeoJSON with mapshaper. |
| KML / KMZ | Google Earth’s format (XML) | Draw-and-export friendly. KMZ = a zipped KML (+ images). Carries styling. Convert to GeoJSON. |
| GPX | GPS tracks / waypoints | What a fitness watch or handheld GPS spits out. |
GeoPackage (.gpkg) / TopoJSON |
Modern one-file DB / topology-aware JSON | GeoPackage is a whole SQLite database of layers. TopoJSON stores shared borders once — smaller, and no slivers between neighbours. |
Why so many for the same idea? History and trade-offs. Shapefile is legacy-but-everywhere, KML is Google/consumer, GeoJSON won the web. They all encode points, lines, and polygons — converting between them is lossless-ish for the geometry; the differences are styling, CRS handling, and how many files it takes.
Raster — a grid pinned to the Earth
A raster is a grid of pixels, where the grid is pinned to the ground so each pixel covers a real patch of Earth. A satellite image, an elevation map, a scanned geological map. Rasters are great for continuous things (colour, height, temperature) and terrible for “what’s the exact boundary” — you have to trace it into vector first (exactly what georeference.ai + the trace script do in the map project).
Think of it as a spreadsheet laid over the land, where each cell holds a value — here, elevation in metres:
elevation.tif
┌─────────────────────┐
│ 210 214 220 231 │
│ 208 212 219 228 │
│ 205 209 215 224 │
└─────────────────────┘
header: top-left cell = [72.000, 25.000], cell size = 0.001°
The file does not store a coordinate for every cell. It stores one anchor (the top-left corner) plus a cell size, and every other cell’s position is computed from that. Ask a raster for the boundary of a forest and it can’t answer — it only knows “this cell is green.” To get an outline, you trace it into vector first.
You’ll meet rasters as these files:
| Format | What it is |
|---|---|
| GeoTIFF | A TIFF image + embedded georeferencing (an anchor coordinate + pixel size + CRS). This is what “each pixel knows its lat/lng” really means — the header stores a transform, not a tag per pixel. Satellite imagery, elevation (DEM), scanned maps. |
| Map tiles (XYZ / WMS) | Not a file — a service. The base map is millions of 256×256 PNG tiles named by zoom/x/y; the map fetches only the ones on screen. The satellite layer on my map is exactly this. |
| NetCDF / HDF | Scientific multi-dimensional rasters (climate, ocean — a value over lat × lng × time). You meet these in earth-science datasets. |
The thing that makes coordinates mean a place: CRS
A pair of numbers [77.2, 28.6] is meaningless until you know the Coordinate Reference System. The one you’ll see 95% of the time is WGS84 (aka EPSG:4326) — plain latitude/longitude on the globe, what GPS uses. GeoJSON is always this.
The wrinkle: the Earth is round, screens are flat. A projection flattens it, and every projection lies somehow — it distorts area, or shape, or distance. Web maps (Leaflet, Google) use Web Mercator (EPSG:3857), which is why Greenland looks enormous. You mostly don’t touch this — but when a shape lands in the wrong spot, a mismatched CRS is usually why.
The one idea to keep
Vector is “here are the exact edges, as coordinates.” Raster is “here’s a picture pinned to the globe.” Every file format above is just one of those two — differing in age, who invented it, and how much extra (styling, CRS, multiple layers) it bundles along. Once that clicked, the whole ecosystem stopped feeling like alphabet soup.