Skip to content
Multigrid

Geospatial data and models

Geocoding, spatial indexing, routing and satellite imagery — the coordinate systems and accuracy tiers that decide whether an answer is usable.

A geospatial pipeline almost never fails loudly. It returns a coordinate, a polygon, a route or a classified pixel, and the answer looks exactly like a correct answer would. The failure is in the provenance: a point that came from a postcode centroid rather than a building, a distance measured in a projection that stretches by a factor of two at high latitude, a change map that found the sun moving rather than the ground. Nothing downstream can tell the difference, which is why the accuracy tier has to travel with the data.

These pages work from the mechanism outward. Where a number matters it is either taken from the specification that defines it and attributed in the sentence, or derived here from inputs you can change. Where a method has a documented failure mode, the failure mode is the point of the page.

Geocoding Accuracy: What Rooftop, Interpolated and ZIP-Level Actually Mean

The three common match levels describe where a coordinate came from, not how wrong it is — so here is the arithmetic that turns each one into metres.

9 min read

Reverse Geocoding: Turning Coordinates Back Into an Address

What a reverse geocoder actually computes for a coordinate — a nearest-feature search plus a stack of point-in-polygon tests, not a lookup.

9 min read

H3 Hexagonal Spatial Indexing Explained

How H3 turns a coordinate into an integer, what each resolution buys you in metres, and the two places the hexagonal grid does not behave like a grid.

10 min read

Spatial Embeddings for Location Similarity

How POI density, demographics and mobility become one vector per location, and why the preprocessing decides the answer more than the model does.

10 min read

Building an Image Classification Pipeline for Satellite Imagery

A tiling-then-classify pipeline over a Sentinel-2 scene, with the tile arithmetic that tells you how much compute you have just committed to.

11 min read

Shortest-Path Routing Algorithms Explained

Dijkstra and A* on the same worked graph, with the settle order written out — so the difference between them is visible rather than described.

10 min read

Isochrone Maps: How "15 Minutes From Here" Gets Computed

An isochrone is a bounded one-to-many search plus a hull, and both halves have parameters that change the answer more than the travel model does.

9 min read

Geospatial Data Formats: GeoJSON, Shapefile and When Each One Breaks

What RFC 7946 actually mandates, what a shapefile silently does to your schema, and which of the two really loses coordinate precision.

10 min read

Address Matching and Normalization Across Data Sources

A parse-normalise-block-score-corroborate pipeline worked on two records for the same building, including the arithmetic that makes blocking non-optional.

10 min read

Detecting Building Footprints From Satellite Imagery

Segmentation gives you a mask, not a footprint — here is the resolution floor, the vectorisation step and the off-nadir geometry that displaces every roof.

10 min read

Change Detection in Satellite Imagery Over Time

Differencing two dated scenes is one line of arithmetic and four preconditions, and the preconditions are what decide whether the output is change or sunlight.

10 min read

Cloud Masking in Satellite Imagery Before Analysis

The spectral tests a cloud mask runs, worked on real reflectance values, plus the shadow geometry and the compositing bias that follow from masking.

10 min read

NDVI and What a Vegetation Index Actually Measures

Why the normalised difference of two reflectance bands tracks vegetation, what its negative values mean, and where it stops discriminating.

9 min read

Geofencing: How a Radius Trigger Actually Gets Computed

Build a radius trigger from a bounding-box prefilter, a haversine distance and a hysteresis band, and see why the naive version fires all night.

9 min read

Matching GPS Traces to Roads (Map Matching)

How a hidden Markov model turns a noisy sequence of fixes into a route, and why nearest-road snapping fails on exactly the traces you care about.

10 min read

Spatial Clustering for Delivery Zone Design

How to pick a DBSCAN epsilon from measured stop density rather than by trial and error, and why density clusters are not delivery zones.

9 min read

Nearest-Neighbor Search Over Millions of Coordinates

What a spatial index actually removes from a nearest-neighbour query, with the operation count derived at ten million points.

9 min read

What Geocoding a Million Addresses Actually Costs

A line-by-line derivation of a million-address geocoding bill from a labelled per-call price, a deduplication rate and a churn rate.

9 min read

Delivery Route Optimization: What the Algorithm Is Actually Solving

Why routing is a travelling-salesman variant, worked by hand at four stops, and what actually happens between four stops and four hundred.

10 min read

Deliveries Flagged at the Wrong Side of the Road

Why a delivery pin lands across the street from the house, traced to centreline interpolation, a degree-space offset and a flipped cross product.

10 min read

Estimating Population Inside a Custom Polygon

Areal interpolation from census units to an arbitrary catchment, worked line by line, then refined with building footprints.

10 min read

Detecting Duplicate Locations in a Points-of-Interest Dataset

A distance-plus-name-similarity score worked on a real pair, plus the two cases where the threshold that works is guaranteed to be wrong.

9 min read

Why a Distance Calculation From GPS Coordinates Was Wrong by 20%

The 20% is derived from a real coordinate pair, traced to Pythagoras on degrees, and separated from the Web Mercator and lat/lon-swap versions of the same bug.

10 min read

Snapping GPS Coordinates to the Correct Building Entrance

Why an address point that is geometrically correct still sends the driver to the wrong door, and how to build the entrance layer nobody sold you.

10 min read

Other topics