Green Spaces, Land Cover, Street Trees and Hypertension in the Megacity of São Paulo

Proximity to green spaces has been shown to be beneficial to several cardiovascular outcomes in urban spaces. Few studies, however, have analyzed the relationship between these outcomes and green space or land cover uses in low–medium income megacities, where the consequences of rapid and inordinate urbanization impose several health hazards. This study used a subgroup of the dataset from The Brazilian Longitudinal Study of Adult Health ELSA-BRASIL (n = 3418) to identify the correlation between the medical diagnosis of hypertension and green spaces in the megacity of São Paulo. Land cover classification was performed based on the random forest algorithm using geometrically corrected aerial photography (orthophoto). Three different indicators of exposure to green spaces were used: number of street trees, land cover and number of parks within 1 km. We used logistic regression models to obtain the association of the metrics exposure and health outcomes. The number of street trees in the regional governments (OR = 0.937 and number of parks within 1 km (OR = 0.876) were inversely associated with a diagnosis of hypertension. Sixty-three percent of the population had no parks within 1 km of their residence. Our data indicate the need to encourage large-scale street tree planting and increase the number of qualified parks in megacities.


Introduction
The rapid concentration of populations in large urban agglomerations has produced megacities, i.e., cities with more than 10 million inhabitants. A range of potential health hazards are present in megacities, especially in the developing world: traffic-related air pollution, substandard housing, urban violence, vector-borne diseases, inadequate sanitation, and insufficient or contaminated drinking water, among others [1]. São Paulo city, the 10th most populated megacity in the world, has more than 11 million habitants [2] and its metropolitan area is home to 21 million individuals.
The abovementioned adverse impacts on the health of a megacity can be mitigated by the availability of green spaces, which can offer different ecosystem services (ESSs) and create beneficial health effects. The ESSs associated with green spaces fall into different categories: provisioning (e.g., food, water, fuel and wood); regulating (e.g., regulation of temperature, water, disease); cultural activity); intermediate  min/week of moderate + vigorous activity); or poor (no moderate + vigorous physical activity); as adopted in other analyses [31].

Hypertension Definition
The information on medical and laboratory assessments of ELSA-BRASIL participants can be found in Aquino et al. (2012), Schmidt et al. (2013) and Bensenor et al. (2013) [28,33,34]. In the ELSA-BRASIL, hypertension was defined as a systolic BP ≥140 mmHg or diastolic BP ≥90 mmHg or the reported use of high blood pressure medication.

Study Area
This study was conducted in São Paulo city, which currently has 12.25 million inhabitants [35]. São Paulo's urban explosion, which created a megacity, occurred in the 1950s with the development of the vehicle industry and increasing industrialization along the highways, which attracted migrants from other parts of Brazil. The large influx of people caused a housing crisis, and the city fringe became populated without urban planning, good transportation or primary services. This irregular expansion affected important agricultural and wooded areas, with several irregular settlements in those areas [36]. Currently, there is high disparity in the distribution of green areas in the city, and only five regional governmental boundaries within the city have more than 15 m 2 of green area per habitant [37]. Traffic-related air pollution in São Paulo city, including PM 2.5 levels, consistently exceeds the levels recommended by the World Health Organization [38] in São Paulo.
The residential distribution of ELSA-BRASIL participants in São Paulo and the São Paulo city layer are shown in Figure 1.

Street Trees Map
A digital map of street tree locations in São Paulo was provided by the municipality of São Paulo. The map identified urban trees on sidewalks, street islands and roundabouts and excluded trees in squares, parks, reserves, and internal public and private areas. The map was generated by quantitative photos and local inventory analysis of trees (orthophoto of 2010). The images had a 2 m resolution on a 1:25,000 scale [39].

Land Cover Classification
A São Paulo city orthophoto ( Figure 2) from 2010, provided by the São Paulo Institute of Geography and Cartography, was used for supervised classification.
The orthophoto had a spatial resolution of 2 m on the pixel side with 3 spectral bands: near infrared (NIR), blue and red. The land cover classification was performed using the random forest algorithm (program QGIS2.18.11; Plugin Dtezaka). Random forest (RF) is a powerful learning classifier algorithm that is one of the most accurate methods of classifying land cover [40]. The RF is a general term for ensemble methods that use tree-type classifiers to train the random forest algorithm, which creates multiple CART-like trees [41]. For classification, each tree in the RF chooses a unit vote for the most popular class (pixel color) at the data input (trainer samples polygon = data input). The output of the classifier is determined by a majority vote of the class [42]. For the trainer samples, the classification of the classes was performed according to pixel color and spectral signature. One hundred and fifty training samples were prepared for each class of land cover. The images were classified within the following land cover classes: tree canopy, grass, bare soil, cement floor, swimming pool, shade, roof (white, gray, dark, ceramic), asphalt and river/lake (adapted from Myeong et al., 2003). For data analysis, we considered the sum of tree canopies and grass as green space and the sum of the different types of roof as constructed areas.  We used a false color composite scheme to allow for simple vegetation detection in the image. In this type of false color composite image, vegetation appears in different shades of red depending on vegetation type and condition due to its high reflectance in the NIR band [43]. Bare soil, roads and buildings may appear in various shades of blue, yellow or gray, depending on their material composition. The orthophoto false color composition used in this study was R (channel 1) = NIR band, G (channel 2) = red band, B (channel 3) = blue band.
The classification accuracy was determined using an error matrix, the Kappa index, which indicates the presence of misclassification. The thematic maps used in this study had Kappa values equal to or above 81%, which is considered an accurate classification according to [44].

Urban Parks
Using QGis 2.18.11, we drew a map of the São Paulo urban parks using the shapefile of the polygons of the municipal parks at the portal GeoSampa [39] and the state parks in the São Paulo orthophoto.

Land Cover Exposure Assessment
Three different green space exposure indicators were used: number of street trees, land cover and Euclidian distance ( Figure 3) from urban parks (reserves excluded) using the program QGIS2.18.11.
Street tree databases and land cover were evaluated within 300 m radius buffers ( Figure 3) around each residence of the ELSA-BRASIL participants. The WHO document "Green Spaces and Health" [45] recommended the 300 m buffer, as it corresponds to approximately 5 min walking distance along walkable roads or pathways. In addition, the number of street trees and amount of land cover were assessed within the 96 districts and 32 São Paulo regional government boundaries. (Figure 4). In São Paulo, regional governments are responsible for, among other tasks, the maintenance of the urban green areas within their limits. Each regional government boundary encompasses a certain number of districts that share similar geographic, urban, demographic and economic features [46].
From the residence of each participant, we measured the Euclidian distance to the 3 closest parks and counted the number of parks within a 1 km Euclidian distance (0-3 parks). The number of parks was also used as a categorical variable (0 park, 1 park, 2 parks and 3 parks).

Data Analysis
R 3.4.1 software for Windows was used for the statistical analysis. For each buffer, the district and regional government boundary data were presented as the mean (±SD), minimum and maximum percentage of the number of street trees, all land cover classes and the mean (±SD) distance of parks.
We performed logistic regression analyses to identify the association between street trees, land cover and the number of parks within 1 km of the participant's residence and the diagnosis of hypertension. Street trees and land cover data represented the 300 m buffer, district boundaries and regional government boundaries. The variance inflation factor (VIF) was used to check the multicollinearity of the variables, and binary logistic regression was used for dichotomous dependent variables. The significance level was set at p = 0.05.
The data are presented as odds ratios (95% confidence intervals). The models were adjusted for sociodemographic variables (age, gender, education level, and race) and cardiovascular risk factors (smoking habits, BMI, diabetes, dyslipidemia, excessive drinking, salt consumption and level of physical activity) from ELSA-BRASIL.
A sensitivity analysis using the individual incomes extracted from the ELSA-BRASIL database and the neighborhood Human Development Index (HDI) [47] was performed by a regression model to check whether the neighborhood socioeconomic status could have interfered in the results.

Results
The sample from ELSA-Brasil in this study (n = 3418) presented a mean age of 52.3 (SD = 9.2); 56% of the sample were women, 62.1% were white; and 32% were hypertensive.
The sociodemographic characteristics (Table 1) and the associated cardiovascular risk factors of the study population are presented ( Table 2). The risk factors for cardiovascular diseases considered in ELSA-BRASIL were smoking habits, BMI, physical activity, excessive drinking, hypertension, diabetes, dyslipidemia and salt consumption for the study population (n = 3418).  We did not find any significant correlation between the number of street trees and any land cover classes (tree canopy, grass, green areas, bare soil, river/lake, swimming pools and roofs) and hypertension diagnosis within the 300 m buffer.
The Descriptive statistic of land cover classes of 300 m buffers, districts and regional government boundaries (Table S1) and the descriptive statistic of proximity (meters) of tree closest park (n = 3418) (Table S2) can be found on the supplementary material.
The percentage of tree canopy and green space within the districts presented a tendency toward negative odds ratios (OR). There was a positive effect on the OR for hypertension diagnosis for each 1% of the constructed area within the districts; the associated OR was 1.011. There was no association between street trees and hypertension in the districts.
The number of street trees in the regional government boundaries was significantly associated with a diagnosis of hypertension in this population (Table 3). An increase of 10,000 street trees in the regional government boundaries was associated with an OR = 0.937. There was no association between all land cover classes and hypertension in the regional government boundaries.
The distribution of the study population according to the number of parks within 1 km was 63.5% (zero parks), 26.2% (one park), 6.4% (two parks) and 3.9% (three parks). No participant had more than three parks within a 1 km Euclidian distance. Having more than one park within 1 km of the participant residence was associated with an OR = 0.876. Additionally, the OR for the presence of two parks was 0.74 (p = 0.04; CI (95%): 0.532 to 0.986). No associations were found with the presence of three parks because of the low frequency of individuals with this condition (3.9%). Table 3. Odds ratios (and 95% confidence intervals) for the association between hypertension diagnosis, land cover and green space variables. The VIF did not show multicollinearity between the variables used in the regression models. The descriptive statistics and the prevalence of hypertension in the different districts are presented in the supplementary material.
The sensitivity analysis did not show significant effects of individual incomes or neighborhood socioeconomic status (assessed by HDI) on the diagnosis of hypertension for those parameters with significance in the previous regression models (Tables S3 and S4).

Discussion
In this study, we have shown that the number of street trees within the local regional government boundaries and the presence of parks within 1 km of the residence have a negative association with hypertension diagnosis in an adult population living in the megacity of São Paulo. In addition, our data showed a positive association with the proportion of constructed areas and hypertension diagnosis in this sample of ELSA-BRASIL.
Our data showed that an increase of 10,000 trees in the regional government boundaries was associated with a hypertension diagnosis OR of 0.93. Several studies have demonstrated positive associations between street trees and different health outcomes in larger cities [23,24,48]. Reid et al. (2017) found associations between better self-reported health and the number of street trees (but not grass) within a 1000 m buffer in New York [48]. Lovasi et al. (2008) showed that an increase in tree density was associated with a lower asthma prevalence in children in New York City [24]. Another study in the same city [23] correlated street trees and birth outcomes and described a significant inverse association between nearby street trees and the odds of preterm birth for all women. That study did not identify a consistent significant relationship between adverse birth outcomes and the normalized difference vegetation index (NDVI), access to major green spaces, or waterfront access adjusted for individual covariates.
The mechanisms by which street trees modulate the consistent positive associations with health are not clear. There is a known association between vehicular air pollution and cardiovascular morbidity and mortality [49]. Street trees are able to remove vehicular air pollution through dry deposition [50], thereby decreasing household exposure. On the other hand, green spaces also might be related to allergic respiratory diseases due to pollens and the release of volatile organic compounds that contribute to increasing air pollution [51][52][53]. In addition, the beneficial effects of green spaces on hypertension might be diminished by living close to major roads [54]. Interestingly, in the ELSA-BRASIL population, there was a positive association between commuting-related physical activity and hypertension among women, which could be explained by greater air pollution exposure [55]. According to Donovan (2017), a tree in areas with higher air pollution levels should have a greater impact on human health than the same tree in areas with lower levels of air pollution [56].
We observed a very irregular distribution of green areas and street trees within the urban fabric of São Paulo, resulting in clear disparities in access to green spaces, with extremes ranging from three to 899 trees within the 300 m buffer. Our data also showed that 63% of this population did not have any parks within 1 km of their residence, and 62% were considered poorly physically active. Distance to green spaces is considered one of the more important factors related to physical activity and green space use [57,58]. Some studies have demonstrated that beyond a 300-400 m distance to the individuals' residence, the use of green spaces begins to decline very quickly [59,60].
Our results showed that people living in areas with more than one park within a 1 km distance of their residence had a lower OR (0.87) for hypertension diagnosis, with an OR of 0.72 for the individuals who had two parks within a 1 km distance of the residence. Several ESSs provided by green areas could be related to this finding: a better local microclimate with a lower temperature and a reduction in air pollution provided by the park, increased physical activity and reduction in stress [61]. Green space exposure is associated with less blood cortisol, which in turn is associated with chronic stress and is a risk factor for cardiovascular diseases [62]. In our study, we lack information on the use of parks for the practice of physical activity, but previous data suggest a beneficial synergistic effect of park distance and use [63] in relation to the incidence of fatal cardiovascular disease.
The % of constructed areas (sum of all classes of roofs) had a weak positive association with hypertension diagnosis in this study, with an OR of 1.01. There are some studies suggesting an increased prevalence of hypertension in urban areas when compared to rural areas [64]. Few studies, however, have analyzed different levels of constructed areas in relation to cardiovascular outcomes. A study from Taiwan found that those living in areas with the highest level of urbanization had the highest prevalence of strokes [65]. The mechanisms involved in this correlation could be related to higher air pollution, temperature levels and stress/crowding in more urbanized areas.
Interestingly, we did not observe health associations using the 300 m buffer for all the land cover classes or street trees, but associations were found using larger scale areas, such as the regional government boundaries. In New York, Reid et al. (2017) did not find positive effects of green spaces on self-reported health using a 300 m buffer [48]. Given the complexities of megacity urbanization, with high density of constructions and several environmental negative pressures, it is possible that the health-associated ecosystem services provided by green spaces are not felt within a 300 m buffer.
We could not find significant associations between the proportion of total green areas (grass + tree canopy) and hypertension diagnosis in the 300 m buffer, districts and regional government boundaries.
There are a few possible explanations for these results. It is possible that green areas in São Paulo are not sufficient in their density, biodiversity or quality, such that the ESSs provided by these areas are not as expected. Fear of violence in nonqualified green areas may hinder their use by the population, showing that green spaces should not be considered equal in terms of the ESS offered [66]. Accordingly, a study using the dataset of the Amsterdam Health Survey (2004) linked the data of neighborhood stressors with cardiovascular outcomes in different races (divided into groups) and observed that high quality green space was associated with lower odds of hypertension, but only in the Moroccan group of the study [67]. It is possible that, if locally unwanted land uses or land uses perceived as unsafe were transformed in qualified green areas, we could have observed different results.
Our study is one of the few studies that investigates the relationships between multiple urban land cover exposure measures and health outcomes in a megacity. Previous studies addressing the correlation of green space and health outcomes used green space metrics such as the NDVI [68][69][70], green space [71] or tree cover [72] around residential locations. Given the urban complexity of a megacity such as São Paulo, we believe the methods we have used provide a comprehensive approach to capturing the different land covers of the city. Another strength of this work is the excellent quality of the health data obtained by the ELSA study.
Our study also has its limitations. We could not obtain information in the present study about the access, quality, usability, and maintenance status of the parks and green areas in this study, which could have improved our data. In addition, we do not have information about eventual disservices of green spaces such as pollen and volatile organic compound emissions in the city of São Paulo. Although there are large socioeconomic disparities in São Paulo that could have influenced our data, the studied population (civil servants) had a better socioeconomic status than the average for the Brazilian population. This fact is also a possible explanation for the lack of a significant associations between individual incomes and neighborhood HDI in the adjusted regression models. Another important limitation of this study is its cross-sectional nature, which does not allow us to determine causation. It would be interesting to perform similar studies in other megacities to compare the main findings.

Conclusions
Even in cities with low-density and irregularly distributed green areas, living in regions with more street trees and nearby parks has a beneficial impact on the diagnosis of hypertension. Our results encourage large-scale street tree planting and the creation of more qualified parks to make cities healthier, especially in megacities. This study aimed to sensitize decision makers by providing information on the number of benefits of green spaces and their associated positive impacts on human health.