Demographic analysis of land surface temperature in Canadian metropolitan areas
By and
This analysis shows that heat doesn't affect Canadians equally in metropolitan areas. Residents living in the hottest areas of cities have a lower median income (-${medianIncomeQuantileDifference} / {medianIncomeQuantilePercent}% less) than residents living in the coolest areas. The hottest areas are also more often home to immigrants (+{immigrantsQuantilePercentage}% more). Compared to the rest of the population, these groups are overexposed to heat waves and the rise in temperatures due to global warming.
The story based on this analysis was published on July 13, 2022, and can be read by clicking here. To download the code and the data related for this project, click here (≈ 6 Go).
Table of contents
Context
This analysis uses satellite imagery to estimate land surface temperature in Canada’s biggest metropolitan areas. The temperature is overlaid with Canadian census data. The goal is to compare the characteristics of residents in high- and low-temperature neighbourhoods.
This analysis also presents the historical and projected temperatures for 1950-2100 in the metropolitan areas studied.
Research questions
In the selected metropolitan areas, is there a correlation between average land surface temperature and:
- Residents’ income?
- Residents’ immigration status?
Data
The 10 most populated metropolitan areas in the country are included in this analysis. The two largest metropolitan areas in each province were also added to obtain a wider geographical distribution. As a result, this analysis includes 17 metropolitan areas, with a total population of {DAsTotalPopulation} representing {populationPercentage}% of the 2016 Canadian population. Also, note that some provinces and territories may not have any metropolitan areas (as defined by the Canadian census).
Land surface temperature
The land surface temperature in metropolitan areas is estimated from Landsat 8 imagery. The selected pictures are the least cloudy observations found from 2019 to 2021. On top of the manual selection, each pixel has a cloud score given by Landsat8 providers (U.S. Geological Survey). When the score is above 0.9, the pixels are excluded from the analysis.
Three images are manually selected between June 1, 2019 and Sept. 1, 2021 (summer days), for each metropolitan area to make up for uncontrolled meteorological events or acquisition errors. The land surface temperature of each pixel is calculated and then averaged. Images acquisition dates can be found in the table below, with UTC dates.
The land surface temperature is calculated from thermal infrared bands 10 and 11, as indicated in the Centre d'enseignement et de recherche en foresterie de Sainte-Foy methodology. After the calculations, each pixel has an estimated land surface temperature in Celsius.
First, the spectral radiance is calculated from the raw satellite digital numbers (DN) and luminance (L).
The black-body temperature (Tb) is calculated using the calibration constants K1 and K2.
Finally, the land surface temperature (LST) is calculated using the Split-Window quadratic algorithm.
The Normalized Difference Vegetation Index (NDVI) is also calculated from Landsat 8 imagery. The index is calculated as a ratio between the red (R: Band 4) and near-infrared bands (NIR: Band 5). The resulting value (between -1 and 1) quantifies the amount of vegetation. Since values below 0 usually indicate water, pixels with negative values were excluded from the analysis.
The data extraction, calculations and conversions from Kelvin degrees to Celsius degrees are run with the Google Earth Engine. GEE users can access the code here. The same code is available to download in the link provided at the top of this page.
The images below are cropped versions of Google Earth imagery, with the land surface temperature overlayed on urbanized areas (more about that in the next section).

Land cover data
Since this analysis focuses on demographics, it must exclude pixels without residents. The 2015 land cover data, produced by Natural Resources Canada, is the reference to identify pixels corresponding to urban areas.
Pixels in croplands, bodies of water, forests and other categories are excluded from the land surface temperature calculations in this analysis. Note that this procedure excludes urban parks (classified as forests) but does not exclude industrial areas (classified as urban).
The map below shows the urban areas (in purple) inside the metropolitan area boundaries (gray).
2016 census data
This analysis is based on the 2016 Canadian census, since the 2021 census data is not completely available yet. The Dissemination areas are used. It’s the smallest geographic unit of aggregation for variables like income and immigration status made publicly available. On average, 700 people live in each of these areas. The land surface temperature and normalized vegetation index of cloud-free urban pixels are averaged over each dissemination area.
Various factors (highly reflective surface material, residual cloud, measurement errors, etc.) can create very high or very low land surface temperature. These outliers are excluded for each metropolitan area. Overall, {outliersCount} outliers with a total population of {outliersPopulation} have been removed from our analysis. This represents {outliersPercent} % of all dissemination areas and {outliersPopulationPercent} % of the total population of the census metropolitan areas studied.
The outliers are the dissemination areas with a mean land surface temperature above the third quartile plus 1.5 times the interquartile range or below the first quartile minus 1.5 times the interquartile range. They are shown as circles in the box plot and the map below.
The table below shows the final selection of dissemination areas used for this analysis. There are {DAsCount} of them with a total population of {DAsTotalPopulation} (according to the 2016 Canadian census). The average land surface temperature (LST) and normalized vegetation index (NDVI) are taken from Google Earth Engine. For the quantiles variable, the explanation is further down. An R script (available to download in the link provided at the top of this page) merges this data with the 2016 Canadian census data using the cancensus package. This script also exports dissemination area centroids. [TODO: put code in repo and publish it]
This analysis uses abbreviations for the census variables. The table below shows the exact census definition for each variable. The abbreviations ending with “.total” are used to calculate percentages.
Quantiles weighted average
Since the geographical context of each metropolitan area affects its land surface temperature, a relative measure of cool and hot areas is needed instead of the absolute temperature itself. Consequently, the dissemination areas are grouped in deciles (quantile in the table below) within each metropolitan area. A quantile of 1 means that the dissemination area is among the coolest 10 per cent within its metropolitan area, while a quantile of 10 means it’s among the hottest 10 per cent.
As the chart below shows, the population living in each dissemination area can vary from a few hundred to several thousand. (The chart also allows you to see the distribution of all imported census variables.)
Census variables are aggregated over their corresponding metropolitan area and decile using a weighted average. The weight corresponds to the total number of respondents to the census question. For instance, to average the median income, the weight corresponds to the respondents aged 15 years and over.
The table below shows the results. This dataset is used to produce bar charts in the results section.
Limitations
Satellite imagery is the best available data to evaluate heat in an urban context across the country. Landsat 8 has a 30-metre resolution for any given location (100-metre resolution for infrared bands). In contrast, air temperature observations are limited to a few weather stations. These stations are often located in airports, which are not representative of an urban context. However, land surface temperature has its drawbacks.
Firstly, it’s essential to understand that land surface temperature is different from air temperature measured by weather stations. For example, the land surface temperature refers to an estimated temperature on the ground, rooftops and tree canopies. It’s influenced by many factors, such as the land surface type, and it’s usually higher than the air temperature. It’s not the temperature experienced by residents. However, studies (source 1, source 2, source 3) show that land surface temperature and air temperature are related: when one gets hotter, the other follows.
While weather stations record air temperature continuously, it takes time before a satellite can take a picture of the same location. Landsat 8 has a 16-day repeat cycle. During a three-month summer, Landsat provides five to six pictures of an area.
In the context of this analysis, the acquisition time of the day for each metropolitan area is different since the satellite travels around the Earth during its cycle. And because cloudy pictures are unusable and have to be excluded, acquisition dates are also different between metropolitan areas.
Land surface temperature is also influenced by the angle of the sun on the Earth's surface. This angle depends on the time of the day and the city configuration (low or tall buildings). This could influence the land surface temperature in a downtown setting especially. Precipitation in the hours before the acquisition or smog can also impact the observations. In the context of this analysis, no rain or only a small amount of rain was recorded before the acquisition of the pictures.
Significant bodies of water can also create specific conditions that land surface temperature may not be able to capture. For example, in Toronto, the breeze from Lake Ontario pushes the heat north in the morning (when the satellite takes its picture). But during the evening, when the breeze stops, the city's south is the hottest area.
Results
This section presents the correlation between census variables, namely median individual income and the immigration percentage, and the land surface temperature.
In five metropolitan areas (Saskatoon, Regina, St. John's, Moncton, Saint John), the total population is less than 300,000 people, and dissemination areas are below 400. In these regions, there is more variability in the observed trends.
Median individual income
The higher the median income is, the lower the land surface temperature. There’s a ${medianIncomeQuantileDifference} gap between the first decile and the last decile when looking at all metropolitan areas together. People living in the last decile earn {medianIncomeQuantilePercent} per cent less. The trend is quite clear across all metropolitan areas studied.
While calculating the weighted average of the median income per decile is not ideal, it’s the most detailed variable published by the Canadian census regarding income. The household income is also available, but the correlations are weaker. There are also several variables for the proportion of residents with a low income, but correlations vary between metropolitan areas (see Others at the end of this document).
First-generation immigrants
The higher the proportion of people born outside of Canada in a given area, the higher the land surface temperature. Again, the trend is quite clear across all metropolitan areas. There is {immigrantsQuantilePercentage} per cent more people born abroad in the last decile compared to the first decile.
There are also several variables for the proportion of immigrants born on a specific continent, but correlations vary between metropolitan areas (see Others at the end of this document).
Others
Other variables
Choose a 2016 Canadian census variable and see the correlation with the land surface temperature. The Normalized Vegetation Index (NDVI), the low-income measures and the immigrants' continent of birth show particularly interesting correlations.
Geographical distribution of dissemination areas
This map shows where dissemination areas' centroids are in a given metropolitan area. The colour represents the selected census variable.
Correlation
Choose two variables and see the correlation between them with all dissemination areas instead of deciles.
Predicted air temperature
The charts below show historical and projected air temperature according to three models. The data comes from ClimateData.ca. The summer season mean temperature for each metropolitan area was downloaded.
Conclusion
This analysis shows the correlation between the urban population socio-demographics and the land surface temperature in Canadian metropolitan areas.
If you spot any issue in the methodology, please do not hesitate to contact us.
Naël Shiab senior data producer, Isabelle Bouchard data analyst.