GTA Urban Analytics

Why this exists

The five police services that cover the Greater Toronto Area — Toronto, York, Peel, Halton, and Durham — each publish crime data in their own format. Different schemas, different coordinate reference systems, different crime-type vocabularies, different update cadences. If you want to compare property crime rates in Mississauga against Markham, or see whether shootings cluster near transit hubs region-wide, you cannot do it with any single feed. The data exists, but it does not talk to itself.

This pipeline fixes that. It normalizes all five sources into one unified schema, then joins Statistics Canada 2021 census data — population and median household income by dissemination area — so that crime counts become per-capita rates. The questions it answers: crime rate per 1,000 residents and violent-crime rate by municipality; property-versus-person breakdowns; location-type distribution (business, outdoor, residence); shootings and firearm discharges; and case clearance rates across regions. A methodological refinement worth noting: anomaly filtering excludes incidents within 500 metres of known regional outliers — large malls, amusement parks, hospitals, GO transit stations — that would otherwise inflate per-capita rates for the small dissemination areas they fall in.

Engineering approach

The codebase is organized as a Python package following an Extract, Transform, Analyze, Visualize pipeline. The transform stage runs a 9-step pipeline that takes raw CSV/JSON from each police service and produces a single normalized dataset.

Schema validation is the backbone. Pandera schemas gate every boundary in the pipeline — typed, with coerce=True so that format inconsistencies are resolved deterministically. Invalid rows are not silently dropped; they are quarantined to a separate CSV with a validation_errors column that explains exactly why each row failed. This means I can audit data quality without re-running the pipeline.

Cross-region ID collisions are a real problem when merging five datasets. Each region prefixes its source_identifier on ingest, and a pytest test specifically guards the known Peel-versus-Halton OBJECTID=1 collision that would otherwise corrupt joins. The coordinate systems require real CRS reprojection — York Region publishes in UTM Zone 17N, which goes through pyproj to WGS84 before it can merge with the other sources.

Enrichment is idempotent and re-runnable. GeoPandas sjoin performs point-in-polygon assignment, attaching crime counts and per-capita rates to census dissemination areas. The entire workflow is reproducible through uv — uv run download, uv run transform, uv run analyze, uv run full-pipeline, uv run build-map, uv run pytest. Tests run manually rather than through CI.

The visualization

The map is a custom React and TypeScript single-page application built on kepler.gl, compiled with esbuild into a standalone HTML file. It renders four layers: a 3D GPU-aggregated crime hexbin that makes density immediately legible, a median-income choropleth, a crime-rate-per-1,000-residents choropleth, and shooting-to-municipality-centroid arcs that show firearm incident flow.

The standalone build embeds its entire dataset inline — gzip-compressed and base64-encoded, decoded in the browser via DecompressionStream — over a Carto basemap. No external data fetch is needed at load time; the page is fully self-contained.

The link above opens the live, interactive kepler.gl map in a new view. It is a large interactive page (~22 MB with embedded data) and needs a moment to load and initialize.