Using satellite images to asses the effect of light intensity on the number of road accidents
Trafikdage 26-27 August 2019
Luca Furlanetto [lucfur@dtu.dk]
Kira Janstrup [kija@dtu.dk]
Thomas Kjær Rasmussen [tkra@dtu.dk]
Agenda
• Introduction
• Methodology
• Data
• Model formulation
• Results
• Discussion and Conclusions
Introduction
• Correlation between the risk of accident and road lighting is well-known
• The reduction of accidents with injuries during darkness is on avg 30% in areas with road lighting [Wanvik, 2009].
• Other studies analysed influence on safety of road lighting on highways [Frith et al., 2016]
or in intersections [e.g., Edwards, 2015]
• Typical approach is to compare # accidents in daytime to # accidents in nighttime
Introduction
• Satellite images are collected continuously
• Provide various information across large geographical area
– In traffic safety only used for evaluating road design [Najjar et al., 2017; Salman, 2016]
or traffic volume [Eslami & Faez, 2010]
• Nighttime satellite image data include information on light intensity, proxy for artificial light
• → How does light intensity influence the number of accidents in the dark hours?
– Not limited to a particular road type, consider most roads in Denmark – More disaggregate than simply light/dark
Methodology
Data Model development Results
Accidents
Light Intensity
Road type &
Traffic volume
Model type &
specification Estimates
Data - Accidents
• Police-reported accidents collected in Vejman (across all modes)
– Localisation, time, person(s), vehicle(s), light condition, road characteristics, etc.
– Severity (material damage, light injury, severe injury, death)
• 2012-2016 data
• 21,224 nighttime accidents included – 45,907 individuals
– 1,836/2,232/272 light injuries/severe injuries/deaths
– 23,221/1,262/1,272/3,371 cars/heavy vehicles/vans/vulnerable road users
Data - Accidents
Data – light intensity
• Visible Infrared Imaging Radiometer Suite sensor (VIIRS), launched 2011
• Average irradiation at nighttime, per month 2012-2016, 260x460m pixels
• Based on series of pictures each month. Some pictures dismissed due to cloud coverage
Data – light intensity
• Visible Infrared Imaging Radiometer Suite sensor (VIIRS), launched 2011
• Average irradiation at nighttime, per month 2012-2016, 260x460m pixels
• Based on series of pictures each month. Some pictures dismissed due to cloud coverage
• Filtering to include only months w. at least 2 cloud-free observations removed 15 month data
• Aggregation, average radiation across 5 years for 4 seasons
• Grouping into categories (<=5;5-10;10-30;>30)
Data – light intensity
Data – Road characteristics
• Vejman road network
– Highly disaggregate network representation (with many attributes) – Only main roads in some municipalities
– Grouped by road type (motorvej, motortrafikvej, øvrige)
Data – Road characteristics
• Vejman road network
– Highly disaggregate network representation – Only main roads in some municipalities
– Grouped by road type (motorvej, motortrafikvej, øvrige)
• Enriched by mapping Weekday traffic (HDT) from Danish National Transport Model when possible [proxy for night]
– <=2000 veh./day (default when no mapping possible) – 2000-5000
– 5000-10000 – 10000-25000 – >25000
Data – Road characteristics
• Vejman road network
– Highly disaggregate network representation – Only main roads in some municipalities
– Grouped by road type (motorvej, motortrafikvej, øvrige)
• Enriched by mapping Weekday traffic (HDT) from Danish National Transport Model when possible [proxy for night]
– <=2000 veh./day (default when no mapping possible) – 2000-5000
– 5000-10000 – 10000-25000 – >25000
Data – Road characteristics
• Vejman road network
– Highly disaggregate network representation – Only main roads in some municipalities
– Grouped by road type (motorvej, motortrafikvej, øvrige)
• Enriched by mapping Weekday traffic (HDT) from Danish National Transport Model when possible
– <=2000 veh./day (default when no mapping possible) – 2000-5000
– 5000-10000 – 10000-25000 – >25000
After joining datasets, the final dataset includes
- Urban areas: 123,337 roads sections; 11,614 accidents
- Non-urban areas: 48,991 road sections; 9,610 accidents
Model formulation
• No. accidents road-section level → Crash frequency model
• Negative Binomial Regression model
– Common approach for count data (non-negative and integer)
– Generalisation of Poisson regression (mean ne variance, over-dispersion) – Good at handling dataset with many 0-observations (zero-inflated variant)
Model formulation
– Probability of yi accidents on a given road i with attributes Xi
Model formulation
– Probability of yi accidents on a given road i with attributes Xi
Irradiance, road light intensity
Xi~ HDT, Annual Weekday Daily Traffic
Road type, motorway, motortrafikvej or other road
Results
Results
→
Results
Results
Discussion and Conclusions
• Road lighting does affect no. accidents
• No or little road lighting increases the number of accidents in rural areas
• Intense road lighting increases the number of accidents in urban areas – Do people drive faster in well-illuminated areas
– Complex environment, risk affected by multiple factors – Coarse image resolution
• For the future
– Higher resolution of satellite images, especially for urban areas – Does results/conclusions vary across severity levels?
– Before and after analysis