Safe Journeys

Crash Records by Year

Source: NZTA Crash Analysis System (CAS). Updated every 4 hours.

About Safe Journeys

Safe Journeys is an interactive crash risk explorer for New Zealand roads. It predicts where crashes are most likely to occur and how severe they may be, using machine learning applied to real crash data.

Data Source

New crash data is automatically pulled from the CAS API every 4 hours.

How Predictions Are Calculated

The system uses two separate models that answer different questions, then combines them:

1. Severity Model — "If a crash happens here, how bad will it be?"

A LightGBM gradient-boosted decision tree trained on 789,270 historical crashes (2000–2021) with 80 features covering road attributes, weather, light, vehicle types, impact objects, and location history. It predicts the probability of Death or Serious Injury (DSI) for each location. Test accuracy: AUC 0.82 on unseen 2024–2025 data.

2. Frequency Model — "How often do crashes happen here?"

A statistical Poisson rate model that counts historical crashes per H3 hexagonal cell (~600m across) and computes an hourly crash rate. This is adjusted in real time based on current conditions:

• Weather — live data from Open-Meteo (rain intensity, temperature, wind, visibility)
• Light — calculated from the sun's position at the current time
• Holidays — NZ public holiday periods have historically elevated crash rates

3. Combined — ETNA (Expected Time to Next Accident)

The frequency rate, adjusted for current conditions, gives the expected time between crashes at each location. This is a statistical average, not a countdown — a cell showing "6 months" means crashes occur there roughly every 6 months on average.

What the Map Shows

Traffic Volume Data

Where available, traffic volume data (Annual Average Daily Traffic) is sourced from NZTA's Carriageway dataset covering state highways and some local roads. This allows crash rates to be expressed per 100 million vehicle-kilometres — separating genuinely dangerous roads from merely busy ones.

Key Findings from the Data

• Vulnerable road users (pedestrians, cyclists, motorcyclists) are involved in fewer than 10% of crashes but face dramatically higher severity — a motorcyclist is 7x more likely to die or be seriously injured than a car occupant
• Rural roads have double the DSI rate of urban roads (10.4% vs 4.8%)
• Speed matters — 100 km/h zones have an 11% DSI rate vs 4.5% in 50 km/h zones
• Most crashes happen in good conditions — fine weather, daylight, dry roads — because that's when most driving occurs

Limitations

• The model identifies statistical patterns, not causes — correlation, not causation
• CAS only includes police-reported crashes; minor unreported incidents are not captured
• The CAS API does not provide time-of-day or day-of-week, limiting temporal analysis
• AADT traffic volumes are only available for state highways, not all local roads
• ETNA is a statistical average, not a prediction of when the next crash will occur

Technology

Built with LightGBM (machine learning), H3 hexagonal spatial indexing (Uber), Flask, Leaflet.js, PostgreSQL, and Open-Meteo weather API. No GPU required — all inference runs on CPU in milliseconds.

Data: NZTA/Waka Kotahi Crash Analysis System (CAS), 2000–2025
Traffic: NZTA Carriageway ADT dataset
Weather: Open-Meteo API (open source, free)
Safe Journeys is a research tool and does not constitute official road safety advice.