A Study of Urban Sprawl and Commuting Modes in Canadian Metropolitan Areas


By Isabelle Bouchard, IVADO scholarship recipient and M.Sc.A., and Nael Shiab, computational journalist for CBC/Radio-Canada.

Contact: isabelle.bouchard77@gmail.com and nael.shiab@radio-canada.ca

The story based on this analysis was published on March 7, 2022. Click here to read it.

The code and the data can be downloaded here.

La version française de cette analyse peut être consultée ici.


Many experts agree that a reduction in urban sprawl and an increase in the densification of Canadian cities are important in the fight against climate change. This analysis uses satellite imagery and census data to quantify how Canadian cities have grown over the last two decades and the impact this growth has had on the number of cars on the road.

We compare the population characteristics of historically urbanized areas (before 2001) with the new neighbourhoods built in the last two decades. Since transportation is a significant producer of greenhouse gas, we mainly focus on commuting habits.

We narrowed this study to the top nine Statistics Canada census metropolitan areas (CMA). CMAs are formed by one or more adjacent municipalities centred on a population center (known as the core). A CMA must have a total population of at least 100,000, of which 50,000 or more must live in the core.

We also used the Proximity Measures Database from Statistic Canada to categorize neighbourhoods (high, medium or low density of services/amenities in close range) and compare CMAs.

Research questions

We want to answer these questions:

  • Q1: How much have urban areas expanded over the years, compared to the population, in each CMA? On average, how much space is occupied by a resident in a new urbanized area compared to a historical one?

  • Q2: According to census data, what are the differences between residents of new urbanized areas and historically urban areas regarding:

    • commuting mode of transportation?
    • type of housing and number of people per household?
    • individual incomes?

  • Q3: If the population density and the commuting modes in newly urbanized areas were the same as historical areas, how much space would have been saved, and how many cars wouldn’t have ended up on roads?

We will distinguish between neighbourhoods (based on the Proximity Measures Database) for the last two questions.

Summary of findings

Our algorithm classified 78.3 million pixels for each year between 2000 and 2021 (1.7 billion pixels in total), covering 47,615 square kilometres in the top ten CMAs (Kitchener-Cambridge-Waterloo was later excluded, as explained below).

In total, we found 4,994 square kilometres of historically urbanized areas (already built in 2001) and 1,700 square kilometres of newly urbanized areas (built between 2001 and 2020). It’s as if we had added two and a half other cities the size of the Greater Toronto Area (630 sq km) in Canada in two decades.

On average, the urbanized area of the top nine metropolitan areas expanded by 34 per cent, while their total population increased by 26 per cent (15.7 million in 2001 compared to 20 million in 2021). This gap caused a 6 per cent density loss (3,152 people per sq km in 2001 compared to 2,975 in 2021). In 2001, the average urban land area per resident was 317 square meters, increasing to 336 square meters in 2021.

According to our analysis, 89 per cent of newly urbanized land in the top nine metropolitan areas comprises neighbourhoods with a low density in services and amenities. In comparison, 73 per cent of historical urban land is low-density.

In low-density neighbourhoods, the average urban land per resident is 405 square meters. It’s 1.6 times more than medium-density neighbourhoods (254 m2) and 3.6 times more than high-density ones (110 m2).

Eight residents out of ten use their car to go to work in low-density neighbourhoods (80 per cent). That’s 25 per cent more than residents of medium-density neighbourhoods (64 per cent) and 122 per cent more than high-density neighbourhoods (36 per cent).


In six of the nine metropolitan areas, urban extent grew faster than the population between 2001 and 2020, resulting in a density loss. In the three other cities, density increased. Ottawa-Gatineau has by far the biggest decrease in population density (-19 per cent), followed by Winnipeg (-13 per cent), Hamilton (-11 per cent), Montreal (-10 per cent), Québec (-1.7 per cent) and Vancouver (-1.5 per cent). For the three other metropolises, Edmonton shows the greatest density gains (+5 per cent), followed by Calgary (+3 per cent) and Toronto (+2 per cent).

Regarding the three largest metropolitan areas (Toronto, Montréal and Vancouver), Montréal is the only city that failed to increase its density. One thing to keep in mind: these three major metropolitan areas were already the densest CMAs in the country (over 3,300 people per square kilometre).

Calgary and Edmonton had the greatest population growth (52 per cent for Calgary and 47 per cent for Edmonton). Their density increased as well.

The four smallest metropolitan areas, Ottawa-Gatineau, Québec, Winnipeg and Hamilton, lost density between 2001 and 2020.

In all cities, between 79 per cent and 95 per cent of newly urbanized lands are categorized as neighbourhoods with a low density in services and amenities. In these new neighbourhoods, the population density is lower than in historical low-density areas.


The latest available data to answer this question comes from the 2016 Canadian census. We use it to compare residents in six distinct areas: the historical and new low-density neighbourhoods, the historical and new medium-density neighbourhoods, and the historical and new high-density neighbourhoods.


Commuting mode

We see two trends:

  • the lower the density of services/amenities, the more likely residents are to drive their car to commute.

  • residents of new urban areas are more likely to commute by car than residents of historical areas.

These two trends add up, meaning that commuting car usage is largest in new low-density neighbourhoods.

We see the opposite trend for the use of active modes of transportation (bicycle, walking) and public transit.

Commuting destination

The lower the density of services/amenities is, the more likely residents are to work outside of their CSD. This trend is true for all cities.

Housing and households

All proportion kept, single-detached houses are more common in neighbourhoods with the lowest density in services and amenities. They are also more common in new areas compared to historical ones. Conversely, apartments are more widespread in core and historic neighbourhoods.

In new urban areas, a higher majority of residents are owners compared to historical locations. Similarly, residents of the lowest density neighbourhoods are more often owners.

All proportion kept, we find more people living alone inneighbourhoods with the lowest density in services and amenities. For a household size of two, it’s roughly the same in all neighbourhoods. But household sizes of three or more are more often found in low-density areas, especially in newly urbanized areas.


The latest available data to answer this question comes from the 2016 Canadian census.

We see that the population density tends to be lower in newly urbanized areas than in historical areas. If it were the same as historical neighbourhoods, 308 square km of land would have been saved. (We took into account the type of neighbourhoods for this calculation.)

We also see that commuting by car tends to be higher in newly urbanized areas than in historical areas. If it was the same as historical neighbourhoods, 129,092 less cars would be on the roads.


We retrieved the census metropolitan areas, census subdivision and dissemination areas geographical limits from the 2016 Census with the cancensus package. The population estimates and projections from 2001 to 2020 come from Statistics Canada. However, since these projections do not take into account the COVID-19 pandemic, we also used population data from the 2021 Census, after projecting them onto the 2016 geographic boundaries. To do this, we used the correspondence files between the 2016 and 2021 dissemination areas. To calculate the population in 2021, we used all dissemination areas fully or partially within the 2016 metropolitan boundaries.

The proximity index comes from the Proximity Measures database.

The satellite imagery comes from Landsat 7 (nominal scale of 30 meters per pixel) through Google Earth Engine. The satellite was launched in 1999 and is still active today, allowing us to study urban sprawl over more than 20 years with the same image source. We created a simple composite of images between May 1st and November 1st for each available year to get cloud-free images.

To classify each pixel as an urban or non-urban area, we trained Random Forest models with Canada’s 2015 Land Cover maps, produced by Natural Resources Canada and available on the Canadian Open Data Portal. Natural Resources Canada used the same kind of machine learning algorithm to produce its 2015 maps.

After training, the models classified each pixel in the CMAs areas between 1999 and 2021. This procedure was done in Google Earth Engine and the results were exported as .tif files. Google Earth Engine users can find the code here. We also added the code in the repository, available to anyone to download (link at the beginning of this analysis). We converted the .tif files into data frames in the script named data_preparation.R.

Once each pixel was classified, we did a spatial join between the pixels and the Census dissemination areas to infer information on the residents from the latest census year (2016).

In data_prepartion.R, we apply multiple noise reduction methods (described in the next section). We applied these algorithms to produce conservative results. In the worst-case scenario, we may underestimate the urban areas, but we did our best not to overestimate them.


According to Natural Resources Canada, Canada’s 2015 Land Cover maps have an accuracy of 79.9 per cent over the 18 classes it covers. We used only one class to train our algorithms: the urban area.

We experimented with two training strategies: one or multiple models. We trained one model on Canada’s 2015 Land Cover data and images in the first strategy. In the second, we focused on city cores and outskirts to train one model per city per year. The validation accuracy is 90.0 per cent for the first method (one_model) and 96.1 per cent for the second method (multiple_models). We have decided to use the second approach with multiple models.

We excluded Kitchener-Cambridge-Waterloo, given the unsatisfactory validation accuracy. This CMA geographical configuration is different from the other ones due to the fusion of three small city cores. It’s the tenth metropolitan area in population, so we narrowed this analysis to the nine largest cities.

Validation accuracy of the two model training strategies for each CMA.

Multiple models approach

We infer partial training data from the 2015 Land Cover maps to train one model per year. We assume that urban core and non-urban lands far from the urban core remained the same through time. We removed these areas to obtain partial maps from which we sampled 10,000 training data points. We visually validated the partial training data and masked some regions overrepresented in the non-urban class. For example, mountains in the Vancouver CMA were overrepresented compared to other non-urban lands. We also masked rivers and lakes according to Statistics Canada boundary files.