Introduction
In urban areas transportation and road networks have become a major source of concern for public health. Yet within a city, not all populations bear the same share of environmental pollution.
Many studies have been published, which relate various levels of social vulnerability with road pollution, mainly air quality assessments. In the UK a study \citep*{Namdeo_2008} has explored the impacts road user charging had on health, based on NO2 levels and the difference in exposure between populations. In another paper \citep{Carrier_2014} correlation between NO2 exposure and schools showing different deprivation levels was examined. In both cases the results showed a strong positive correlation between vulnerability among populations and health risks associated with road networks.
One of the elements associated with road networks often not considered is gas stations. These usually appear as a source of pollution, which also can have health impacts. One paper \citep{Morales_Terr_s_2010} investigated the influence gas stations had on air in their surroundings due to VOCs emanating from them. Air measurements were taken and a model based on concentration ratios was developed to conduct the study. The paper concluded stations had an influence on close surroundings with high concentrations of benzene and n-hexane being measured. This influence was shown to depend highly on the characteristic of the area close in the direct vicinity of the stations. When buildings were located close by the dispersion of VOCs was hindered, thus their concentration increased.
The present paper aims therefore to enlarge the scope of analysis described previously mixing both approaches. Indeed, the focus will be on the correlation, which exists between socioeconomically vulnerable populations and their proximity to gas station and areas polluted by those. The initial statement is that deprived populations are more inclined to be located closer to gas stations and their pollution related sites. Socioeconomic vulnerability was here considered in terms of housing assistance data available, which was then correlated with gas stations and their associated polluted areas in order to verify the above mentioned hypothesis. The area of study is the municipality of Vernier (Geneva, Switzerland).
Data
In order to perform the study several raster and vector layers were used.
Demographic data of Switzerland was collected from the Swiss Federal Office of Statistics (OFS). This demographic data contained point coordinates which enabled to build an hectometric raster grid.
In order to assess vulnerability, housing assistance data of the municipality of vernier was used. This data also came from the Swiss Federal Office of Statistics and presented itself as a table containing the detailed numbers of housing assistance.
Then, data collected from the open data collection from the Geneva Territory Information System (SITG) was also used. One point shapefile contained the location of gas stations within the entire canton of Geneva. Another shapefile, of polygons this time, was used to locate the registered polluted areas.
Finally, the boundaries of the municipality of Vernier stored in a polygon shapefile were also used. Those also came from the OFS.
All vector and raster layers used were projected according to the Swiss coordinates system : EPSG21781.
Methods
The results needed to realize the study were produced using both QGIS and GeoDa softwares. Those were used following their respective "QGIS User Guide"\citep{athan2017} and "GeoDa User Guide" \citep{anselin2003}.
The main analysis tool used was a 100x100m grid with its extent limited within the boundaries of Vernier. To create this grid the hectometric point coordinates contained in the Swiss demographic data file were necessary. These coordinates allowed to locate the center of the grid cells although a 50m both longitudinal and latitudinal shift was necessary in order to match the used data.
Using the "Spatial Query" tool available in QGIS enabled to confine the grid inside the municipality of Vernier. It is worth noticing that the grid did not cover the entirety of the municipality but only its populated areas.
After that, the vector "Analysis tool" in QGIS made possible to count the number of housing assistance provided in each hectometric cell. This proved useful to then categorize the grid according to a vulnerability graduation.
Both the service station point shapefile and the polluted areas polygon shapefile were also restricted to the municipality of Vernier thanks again to the "Spatial Query" tool. Further work was performed on the polluted areas shapefile restraining its elements to the ones having an activity type of either "service station" or "fuel related businesses".
Finally, minimal distances from the center of the grid cells to service stations and polluted areas were computed. To do this, layers of polygon centroids were created. The results were then calculated using the "Distance matrix" contained in the vector analysis tools.
In GeoDa the grid was exploited to extract box maps, residual maps and scatter plots. Those allowed to establish comparisons between the different values of vulnerability and distance. They were also of interest to verify the spatial correlation between the short distances to polluted areas and high housing assistance numbers, which made possible the verification of the hypothesis. In addition, two models of multivariate regression (with and without dependent spatially weighted variable) were performed in order to assess their performance with respect to real data.
Both methods are expressed as follows:
Ordinary Linear Regression (OLR):
\(y_i\ =\ \beta_0\ +\ \beta_1x_{1i\ }+\beta_2x_{2i}\ +\ \epsilon_i\)
Spatially weighted regression:
\(y_i\ =\ \beta_0\ +\ \beta_1x_{1i\ }+\beta_2x_{2i}\ +\rho\ \Sigma w_jy_j+\ \epsilon_i\)
In both equations \(y_i\) represents the dependent variable (housing assistance number), \(x_1\) (independent) is the minimum distance from the center of a grid cell towards the closest service station, \(x_2\) (independent) the minimum distance towards a polluted area and \(\epsilon_i\) the error. For the spatially weighted regression \(\rho\) is the spatial lag and \(w_j\) the weight of spatial unit \(j\) relative to the spatial unit \(i\).
Results