Environment International 126 (2019) 162–170 Contents lists available at ScienceDirect Environment International journal homepage: www.elsevier.com/locate/envint Associations of green space metrics with health and behavior outcomes at T different buffer sizes and remote sensing sensor resolutions Jason G. Sua, , Payam Dadvandb,c,d, Mark J. Nieuwenhuijsenb,c,d, Xavier Bartolle,f, ⁎ Michael Jerrettg a School of Public Health, University of California, Berkeley, USA b ISGlobal, Centre for Research in Environmental Epidemiology (CREAL), C/ Doctor Aiguader 88, 08003 Barcelona, Spain c Pompeu Fabra University, Doctor Aiguader 80, 08003 Barcelona, Catalonia, Spain d Ciber on Epidemiology and Public Health (CIBERESP), Melchor Fernández Almagro, 3-5, 28029 Madrid, Spain e Agència de Salut Pública de Barcelona, Barcelona, Spain f Institut de Recerca Biomèdica Sant Pau (IIB Sant Pau), Barcelona, Spain g Fielding School of Public Health, University of Los Angeles, Los Angeles, USA ARTICLE INFO ABSTRACT Handling Editor: Olga-Ioanna Kalantzi Satellite data is increasingly used to characterize green space for health outcome studies. Literature suggests that green space within 500 m of home is often used to represent neighborhood suitable for walking, air pollution and Keywords: noise reduction, and natural healing. In this paper, we used satellite data of different spatial resolutions to derive Normalized difference vegetation index WorldView2 normalized difference vegetation index (NDVI), an indicator of surface greenness, at buffer distances of 50, 100, RapidEye 250 and 500 m. Data included those of 2 m spatial resolution from WorldView2, 5 m resolution from RapidEye LandSat and 30 m resolution from Landsat. We found that, after radiometric calibrations, the RapidEye and WorldView2 Spatial resolution sensors had similar NDVI values, while Landsat imagery tended to have greater NDVI; however, these sensors Buffer distance showed similar vegetation distribution: locations high in vegetation cover being high in NDVI, and vice versa. Health outcomes We linked the green space estimates to a health survey, and identified that higher NDVI values were significantly Barcelona associated with better health outcomes. We further investigated the impacts of buffer size and sensor spatial resolution on identified associations between NDVI and health outcomes. Overall, the identified health outcomes were similar across sensors of different spatial resolutions, but a mean trend was identified in bigger buffer size being associated with greater health outcome. 1. Introduction et al., 2015b). Some studies have, however, found some of the above mentioned associations non-significant (Ali et al., 2017; Cohen-Cline Urban green space, including parks, forests, green roofs, streams, et al., 2015; Picavet et al., 2016; Potestio et al., 2009). Few studies and community gardens, provides critical ecosystem services (Wolch investigated the possible impacts of difference in green space char- et al., 2014). Access to green space has also been associated with var- acterization on health outcome estimates. One study using the global ious health benefits (Ulmer et al., 2016), such as higher levels of phy- positioning system on some 76 participants investigated the Modifiable sical activity (Akpinar, 2016; Almanza et al., 2012; Gomez et al., 2010; Area Unit Problem (MAUP) that affected estimated physical activity Gordon-Larsen et al., 2006; Kaczynski et al., 2008; Mytton et al., 2012; effects of green space and other built environmental variables Sugiyama et al., 2010; Villeneuve et al., 2018), improved children (Houston, 2014). That study suggests that both the size (scaling effect) cognitive skills development (Dadvand et al., 2015a), improved mental and shape (i.e., the zoning effect) of the spatial units used in an analysis health conditions (Dadvand et al., 2016; Gascon et al., 2015; Wood may influence resulting statistical inference. This study found both et al., 2017), reduced stress levels (Roe et al., 2017; Thompson et al., scaling and zoning effects, with smaller circular buffers generally re- 2016), and lower exposure to traffic-related air pollution (Dadvand lating to larger effect sizes when relating green space to objectively Abbreviations: NDVI, normalized difference vegetation index; TOA, Top of Atmosphere; MAUP, Modifiable Area Unit Problem; USGS, United States Geological Survey; BMI, body mass index; GHQ, General Health Questionnaire; IPAQ, International Physical Activity Questionnaire; SES, socioeconomic status ⁎ Corresponding author at: Division of Environmental Health Sciences, School of Public Health, 2121 Berkeley Way, University of California at Berkeley, Berkeley 94720-7360, USA. E-mail address: jasonsu@berkeley.edu (J.G. Su). https://doi.org/10.1016/j.envint.2019.02.008 Received 11 October 2018; Received in revised form 2 February 2019 Available online 21 February 2019 0160-4120/ © 2019 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/BY-NC-ND/4.0/). J.G. Su, et al. Environment International 126 (2019) 162–170 measured physical activity. The study, however, did not assess the in- atmospheric conditions and other factors that can influence the ob- fluence of resolution in the input data for characterizing green space. served energy on a sensor. Due to the fact that the data were all ac- To improve green space exposure estimates, researchers have in- quired in summer when vegetation was greenest and they were col- creasingly relied on the remote sensing images taken from satellites lected in close proximity in date, we believed that a surface, such as the (Almanza et al., 2012; Dadvand et al., 2012; Gradinaru et al., 2016; vigorousness of vegetation, remained unchanged and its corrected re- Vatseva et al., 2016). Most of the research using remote sensing sa- flectance remained the same across the three sensors. However, when tellites has used the normalized difference vegetation index (NDVI) the spatial resolution decreases, a pixel on a surface would have more with 30 m resolution from Landsat images or 250 m resolution from the mixed ground information due to increased pixel size. This would result MODIS (Casey et al., 2017; Crouse et al., 2017). To be able to accurately in over- or underestimating the degree of vegetation. At the same time, identify green space but at the same time limit potential financial with increasing buffer size, the over- or underestimation in degree of burden, we investigated whether it was necessary to purchase high vegetation would be further exaggerated. spatial resolution (e.g., meter or sub-meter) remote sensing data for the Radiometric calibration and correction on our remote sensing data identification of health outcomes. Literature suggests that green space included the following connected steps: We first converted the original within 500 m of home such as those of 50, 100, 250 and 500 m are often sensor acquired images in digital numbers to radiance based on the used to represent the immediate neighborhood of residence suitable for rescaling factors provided in the respective sensor metadata files. Based physical activity like walking (Su et al., 2011; Wolch et al., 2011), and on Earth-Sun distance, solar zenith angel and exoatmospheric irra- in the presence of green space suitable for reducing noise and air pol- diance at the time of data acquisition, the radiance was further con- lution (Dadvand et al., 2015b; Davies et al., 2009). The smaller buffers verted to Top of Atmosphere (TOA) reflectance. also may account for immediate green space that could be in view of the home. In an ongoing study, we identified the associations of green space 2.1.3. Derivation of vegetation index — NDVI with mental health status, social support and physical activity for the The NDVI is calculated as follows: City of Barcelona, Spain using the NDVI data derived from Landsat8 NDVI = (NIR RED )/(NIR + RED) (3) acquisition for buffers of 100, 250 and 500 m of participants' home addresses (Dadvand et al., 2016). In this paper, we extended those where RED and NIR stand for the spectral reflectance measurements green space analyses with three different satellite sensors in respective acquired in the visible red band and near-infrared regions, respectively spatial resolutions of 2, 5 and 30 m, for buffer distances from home (Kriegler et al., 1969). Table 1 lists the bandwidths used for deriving address of 50, 100, 250 and 500 m, to take into consideration of po- NDVI surfaces for the city. NDVI ranges between −1 and 1 with higher tential existence of unmapped trails to and visual effects from green numbers indicating more green vegetation. “−1” represents very high space. We then compared whether these different estimates of green reflectance in the visible red band but with little near-infrared scat- space influence the health outcomes observed in epidemiological stu- tering such as from snow or cloud cover. dies. We hypothesize that NDVI derived from higher spatial resolution We derived NDVI surfaces for the entire city, maintaining the initial remote sensing data or greater buffer size would lead to larger health spatial resolutions of the WorldView2, RapidEye and Landsat imagery. outcomes detection from green space exposure. Mean NDVI values of circular buffer distances of 50, 100, 250 and 500 m were calculated for all the image pixels in the study region. A 2. Material and methods regular grid of 10,141 points of 100 m apart was also generated to compare NDVI values from the three sensors across the four buffer sizes 2.1. Remote sensing data analysis and their associated distributions in the study region. The agreements in NDVI values between the three sensors across the four buffer sizes 2.1.1. Remote sensing data sources were conducted through paired sample t-tests on the regular grid of The City of Barcelona is located on the northeast part of Spain 10,141 points. We also ranked NDVI values from high to low using the (Fig. 1). It has about 1.6 million people (Figueras et al., 2008). To WorldView2 imagery as a reference and identified deviations from the identify the impacts of spatial resolution on green space characteriza- reference line when remote sensing sensor changed. tion, we applied three different resolution remote sensing data for the For a sensitivity analysis, we also generated NDVI surfaces and region (Table 1), all collected in July and early August when vegetation buffer statistics using the primitive remote sensing data that only had shows greatest greenness: (1) the 2 m resolution WorldView2 imagery digital numbers [counts]. The purpose of this was to identify whether (DigitalGlobal, Colorado, USA) collected on August 3rd, 2012, in- radiometric calibrations were necessary to create proper vegetation cluding three visible bands (red, green and blue) and a near infrared index. band; (2) the 5 m resolution RapidEye imagery (RapidEye AG, Berlin, Germany) collected by satellites on July 23rd, 2012 for three visible 2.2. Health outcome analysis (red, green and blue) and two infrared bands; and (3) the 30 m re- solution Landsat8 data (USGS, Reston, Virginia, USA) acquired from the 2.2.1. Study population United States Geological Survey (USGS) for data collected on July 5th, The health analysis was based on a cross-sectional study of data 2012, including 11 bands. We could not acquire the remote sensing obtained from a population-based randomized sample of adults residing data for the same day due to difference in day of sensor data acquisition in Barcelona. The data was collected in the context of the 2011 and difference in time of acquiring high quality images, e.g., cloud free, Barcelona Health Survey aimed to study the health status, life-styles for the city. For each satellite imagery retrieval, we assessed green and use of health services among Barcelona residents. Detailed de- space at buffer sizes of 50, 100, 250 and 500 m around the participants' scription of this survey has already been published (Dadvand et al., home addresses. 2016) and some descriptive statistics are presented in Table 2. Briefly, 4000 people residing across the 10 districts of Barcelona (400 from 2.1.2. Processing remote sensing data each district) were randomly selected from the Barcelona municipal When remote sensing data are acquired, they are presented as pixel register of residents to represent the age and sex structure of that dis- values or digital numbers. The value recorded for a given pixel includes trict. An invitation letter was sent to selected subjects, informing them not only the reflected or emitted radiation from the surface, but also the about the objectives of the survey and asking them to participate. The radiation scattered and emitted by the atmosphere. Given that the data non-responders were substituted by randomly-selected persons of same acquired from the three sensors were not captured at the same date and district, with the same age and sex. For this study, we limited the time, radiometric calibrations were applied to remove impacts from participants to those with age ≥ 18 years resulting in 3461 adults in the 163 J.G. Su, et al. Environment International 126 (2019) 162–170 Fig. 1. The geographic location of Barcelona in Spain, Europe. Table 1 Table 2 The remote sensing imagery and bandwidths for creating NDVI indices for The descriptive statistics of the participants in demographic characteristics and Barcelona. measured health outcomes. Data source Acquisition date Resolution (m) Red band Near-infrared Variable Description band Age WordView 2 08/03/2012 2 630–690 nm 770–895 nm 18–45 years 1571 (45.4%) RapidEye 07/23/2012 5 630–685 nm 760–850 nm 46–65 years 1039 (30.0) Landsat8 07/05/2012 30 640–670 nm 850–880 nm ≥ 65 years 851 (24.6%) Sex Male 1657 (47.9%) analytical data set. A participant's main home address in which the Female 1804 (52.1%) interview was conducted was geocoded and the green space metrics Education described above were assigned to the geocoded location through the No or primary school 666 (19.2%) ArcGIS interpolation function (ESRI, Redlands, California). We used the Secondary school 1618 (46.8%) 2012 rather than the 2011 NDVI data to match the 2011 health out- University 1148 (33.2%) Missing 29 (0.8%) come data due to the best available remote sensing data in 2012 on cloud free status in the summer months for all the three sensor acqui- Subjective general health Excellent/very good/good 2702 (78.2%) sitions. Fair/bad 755 (21.8%) Mental health status 2.2.2. Questionnaire data At risk 480 (14.4%) The data on the perceived general health, mental health, physical Not at risk 2856 (85.6%) activity, and relevant socio-demographic covariates were obtained Physical activity through a face-to-face interview survey conducted at the residential Moderate/high 236 (7.5%) place of the study participants. For perceived general health, partici- Low 2907 (92.5%) pants answered a question obtained from the Short-Form 36 (Ware and 164 J.G. Su, et al. Environment International 126 (2019) 162–170 Table 3 The correlation matrix of NDVI indices derived from WorldView2, RapidEye and Landsat NDVI indices at buffer distances of 50, 100, 250 and 500 m through a regular grid of 10,141 points of 100 m apart in the study region. WV50 WV100 WV250 WV500 RE50 RE100 RE250 RE500 LS50 LS100 LS250 WorldView (100 m) 0.97 WorldView (250 m) 0.92 0.97 WorldView (500 m) 0.89 0.94 0.98 RapidEye (50 m) 0.99 RapidEye (100 m) 0.99 0.98 RapidEye (250 m) 1.00 0.93 0.97 RapidEye (500 m) 1.00 0.91 0.95 0.99 Landsat (50 m) 0.97 0.97 Landsat (100 m) 0.99 0.98 0.98 Landsat (250 m) 0.99 0.99 0.93 0.97 Landsat (500 m) 0.99 0.99 0.90 0.94 0.98 Sherbourne, 1992): “In general, would you say that your health is...” health estimates through linear regression models. This was done for with possible responses being excellent/very good/good/fair/bad. The individual mean health estimates, but also for the three health out- Short-Form 36 has been reported to be a reliable and valid tool to assess comes combined. For example, when estimating the trends of associa- perceived general health in the Spanish population (Vilagut et al., tion of spatial resolution with perceived health, the predictor was 2005). The answers were dichotomized with cut-off at “less than good”, “sensor resolution” with values 2, 5 and 30 m and the response was following the same methodology used in previous studies (Maas et al., “perceived health” with three corresponding mean estimates on per- 2006; Triguero-Mas et al., 2015). We considered “less than good” an- ceived health. The linear regression model had 3-paired values. When swers as the reference category, therefore a positive association be- the three health outcomes were combined, the predictor “sensor re- tween greenness exposure and this variable could be interpreted as solution” had 9 values with values 2, 5 and 30 m all occurring 3 times better perceived general health. For mental health, participants an- for the three different health outcomes, creating 9-paired values be- swered the twelve questions of the General Health Questionnaire (GHQ- tween sensor resolution and modelled mean health estimates. Second, 12) (Goldberg, 1978). GHQ-12 has been reported to be a valid and we regressed buffer size against the modelled mean health estimates. reliable tool for screening non-psychotic mental problems (i.e. psy- The predictor “buffer size” included values of 50, 100, 250 and 500 m. chological distress) in the Spanish population (Sánchez-López and Similarly, the analyses were done for both individual mean health es- Dresch, 2008). The general score was dichotomized with those having a timates and the three health outcomes combined. score ≥ 3 being classified as being at risk of psychological distress (Goldberg, 1978). We considered being at risk of psychological distress 3. Results as the reference category; therefore, an association between greenness exposure and this variable could indicate lower risk of psychological 3.1. Differences in characterizing green space distress (i.e. better mental health status). For physical activity, parti- cipants answered the seven questions of the International Physical The correlation coefficients (R) between the four buffers of a remote Activity Questionnaire (IPAQ-Short version) (Craig et al., 2003). The sensing imagery on the regular grid of 10,141 points (highlighted areas IPAQ has been reported to have an acceptable validity and reliability in Table 3) showed that the degree of agreement in green space was for the Spanish population (Roman-Viñas et al., 2010). We developed a greater when the defined scopes of impact were closer to each other. binary variable indicating whether the participant could be considered The degrees of correlation for WorldView2, RapidEye and Landsat8, as having moderate or high physical activity levels based on the IPAQ ranged 0.89–0.98, 0.91–0.99 and 0.90–0.98, respectively, when com- guidelines (i.e. moderate/high vs. low levels of physical activity) (IPAQ pared within individual sensors across the four buffer sizes. When Research Committee, 2005). We considered low physical activity level compared between two different sensors, we found the correlations as the reference category; therefore a positive association between were slightly stronger between WorldView2 and RapidEye than be- greenness exposure and this variable could be interpreted as greater tween Landsat and these other two sensors. In addition, correlations likelihood of achieving moderate to high levels of physical activity. were slightly stronger with a larger buffer size. Though relative high in correlation coefficients, all the paired two sample t-tests showed sig- 2.2.3. Statistical analyses nificant differences in NDVI values derived from two types of sensors. We developed logistic regression models with perceived general When comparing NDVI values derived from different sensors health, mental health, and physical activity as outcome (one at a time) (Fig. 2), we found that NDVI values derived from Landsat8 data were on and measures of greenness exposure (one at a time) as predictor. The average greatest (mean = 0.27–0.28), while these values decreased for models were further adjusted for potential confounders identified a WorldView2 (mean = 0.18) and were closely followed by RapidEye priori: age (18–45, 46–65, or < 65), sex, and indicators of socio- imagery (mean = 0.16). NDVI indices derived from WorldView2 were economic status (SES) at both individual and area levels. Educational slightly higher than those derived from RapidEye for the buffer sizes of attainment (no or primary/secondary/university) was used as the in- 50, 100, 250 and 500 m. By comparison, Landsat8 consistently created dicator of individual-level SES and tertiles of 2010 household income higher NDVI indices by about 0.1. Even though there were these dif- by neighborhood (Generalitat de Catalunya, 2010) was applied as the ferences, the NDVI values derived from the three types of sensors indicator of area-level SES. showed a similar trend: higher NDVI values in one sensor was asso- In extending the previous work (Dadvand et al., 2016), we first ciated higher NDVI values in another sensor and vice versa, indicating modelled the associations of green space metrics with individual health that NDVI values were high when vegetation cover was high and de- outcomes separately for data acquired from sensors of three spatial creased when degree of vegetation decreased. The spatial patterns resolutions (2, 5 and 30 m) for four buffer sizes (50, 100, 250 and largely remained unchanged for NDVI indices derived from the three 500 m). The individual health outcomes in mean estimates were then different sensors. used to conduct trend analysis. Two main trends were analysed: First, We also compared the differences in NDVI for two specific neigh- we regressed sensor spatial resolution against the modelled mean borhoods of interest: the one at Estadi Olímpic with high in vegetation 165 J.G. Su, et al. Environment International 126 (2019) 162–170 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0.0 0.0 RapidEye_50m (m=0.16) RapidEye_100m (m=0.16) -0.2 Landsat_50m (m=0.28) -0.2 Landsat_100m (m=0.28) WorldView_50m (m=0.18) WorldView_100m (m=0.18) -0.4 -0.4 0.8 0.8 0.6 0.6 0.4 0.4 0.2 0.2 0.0 0.0 -0.2 RapidEye_250m (m=0.16) RapidEye_500m (m=0.16) -0.2 Landsat_250m (m=0.27) Landsat_500m (m=0.27) -0.4 WorldView_250m (m=0.18) WorldView_500m (m=0.18) -0.4 Fig. 2. The distribution of the NDVI values for WorldView2, RapidEye and Landsat data defined by buffer sizes of 50, 100, 250 and 500 m for the City of Barcelona. A regular grid of 10,141 points of 100 m apart was used for the analysis and WorldView2 imagery was used to rank the NDVI values from the highest to the lowest. Numbers in the parentheses are the mean NDVI values derived from a sensor. a1 a2 a3 b1 b3 b2 Fig. 3. NDVI surfaces and mean values in buffer sizes of 50, 100, 250 and 500 m (red circles) derived from WorldView2 (a1 for 0.59, 0.21, 0.20 and 0.25; and b1 for 0.10, 0.08, 0.07 and 0.07), RapidEye (a2 for 0.48, 0.17, 0.17 and 0.21;and b2 for 0.12, 0.09, 0.09 and 0.08) and Landsat8 (a3 for 0.60, 0.28, 0.29 and 0.34; and b3 for 0.18, 0.17, 0.18 and 0.17) for points of Estadi Olímpic (a) and Jardins del Mestre Balcells (b) vegetation cover areas in the immediate neighborhoods. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) 166 J.G. Su, et al. Environment International 126 (2019) 162–170 0.8 0.7 RapidEye_50m (m=-0.04) RapidEye_100m (m=0.01) Landsat_50m (m=0.27) Landsat_100m (m=0.27) 0.6 0.5 WorldView_50m (m=0.21) WorldView_100m (m=0.20) 0.4 0.3 0.1 0.2 -0.1 0.0 -0.3 -0.2 -0.4 0.7 0.8 RapidEye_250m (m=-0.04) RapidEye_500m (m=0.01) 0.6 0.6 Landsat_500m (m=0.26) Landsat_250m (m=0.27) 0.5 WorldView_500m (m=0.20) WorldView_250m (m=0.20) 0.4 0.4 0.3 0.2 0.2 0.1 0.0 0.0 -0.1 -0.2 -0.2 -0.4 -0.3 Fig. 4. The distribution of the NDVI values for WorldView2, RapidEye and Landsat data in buffers of 50, 100, 250 and 500 m with the imagery data not being corrected for atmospheric corrections. A regular grid of 10,141 points of 100 m apart was used for the analysis and WorldView2 imagery was used to rank the NDVI values from the highest to the lowest. Numbers in the parentheses are the mean NDVI values derived from a sensor. cover and the second at Jardins del Mestre Balcells with intermediate different patterns. Though their 95% confidence intervals overlap, we vegetation cover (Fig. 3). The high vegetation cover area was centered found that, for perceived health, health associations were relatively on a soccer field (a1, a2 and a3 in Fig. 3), and we found that for the larger with the buffer size of 250 m. For mental health, the peak was buffer defined within 50 m (the smallest circle), the NDVI indices de- associated with the buffer size of 100 m. For physical activity, the buffer rived from Landsat8 had the greatest value, these values decreased size of 500 m was seen having relatively larger health effect. slightly for WorldView2 and then a much bigger reduction by Rapi- For the satellite resolution (using TOA corrected imagery), we also dEye. When the bigger buffer sizes were used, some non-vegetated found that linear trends existed, with higher spatial resolution asso- areas were included. The NDVI values decreased but still maintained ciated with greater health outcomes; however, these trends were not that green space was greatest for Landsat8. RapidEye imagery, by statistically significant, with p values being 0.58, 0.39, 0.56, and 0.34 contrast, maintained a tendency of lower estimates of NDVI values for perceived health, mental health, physical activity and the combined compared to Landsat8. At the area defined by intermediate vegetation health. When limiting comparison within each health outcome to the cover, we found that both WorldView2 and RapidEye were very similar same buffer size (either 50, 100, 250 or 500 m) (i.e., the association to each other in estimating NDVI (b1, b2 and b3 in Fig. 3). For Landsat8 between sensor resolution and a health outcome given a fixed buffer data, the estimation of higher NDVI was still seen for all the four buffer size), we found that the green space characterized by finer spatial re- sizes. solution had larger health associations. The trends in association were In addition, we compared the NDVI values derived from the three largely not statistically significant, except for physical activity when sensors with buffer sizes of 50, 100, 250 and 500 m using the un- finer spatial resolution was associated with significant larger health corrected imagery (Fig. 4). For Landsat data, the corrections made outcomes for both buffer sizes 50 m (p = 0.05) and 500 m (p < 0.001). slight increases in NDVI values compared to those uncorrected for all the buffer sizes. For the WorldView2 imagery, the corrected imagery had little change to the uncorrected imagery across four buffer sizes. 4. Discussion For the RapidEye data, the uncorrected imagery showed significant underestimation of NDVI values than the corrected imagery, with an In this paper, three satellite sensors, including WorldView2, average difference of 0.15. RapidEye and Landsat, with respective imagery of spatial resolution in 2 m, 5 m and 30 m, were used to identify possible difference in char- acterizing green space in the City of Barcelona. Our research indicates 3.2. Differences in health outcomes that the vegetation index NDVI derived from three sensors are com- parable to each other, with NDVI derived from WorldView2 and Overall, our research found that higher vegetation index was asso- RapidEye imagery showing greater similarity and the Landsat imagery ciated with better perceived health, better mental health and greater having a tendency to predict higher NDVI values compared to the other physical activity identified through the three sensors across the four two sensors. The higher degree of NDVI identified from Landsat was buffer sizes. For trend analysis, we found that some significant linear probably (1) due to the use of vegetation as a representation for pixels trend existed in explaining health outcomes with varying buffer sizes with mixed vegetation and other land cover information and (2) due to (p = 0.01). The trend was significant and positive for perceived health the differences in sensor spectrum design. The latter scenario could be (p = 0.01) and physical activity (p < 0.001) with increasing buffer identified from Table 1 that the Landsat's NIR band had relatively size, but only marginally insignificant for mental health (p = 0.09). higher spectrum wavelength, resulting in greater vegetation re- When examining individual health outcomes trends in Fig. 5, we found flectance. The difference in the red band wavelength might also 167 J.G. Su, et al. Environment International 126 (2019) 162–170 Fig. 5. The modelled health outcomes (adjusted odds ratio & 95% confidence interval) in perceived health, mental health and physical activity using NDVI index identified through buffer sizes of 50, 250, 250 and 500 m for WorldView2, RapidEye and Landsat8 data. The modelled health outcomes were adjusted by age, sex, education and neighborhood level socioeconomic status. contribute to the NDVI difference and further studies should be con- tracking system that will help us resolve the MAUP issue. If GPS data ducted to confirm our assumption. are not available, the next option is to explore multiple buffer sizes of If the imagery data were uncorrected for atmospheric impact, the impact to identify whether the associations between green space and NDVI values derived from Landsat and WorldView2 imagery showed health are sensitive to the size of the buffer used to specify the potential minor differences compared to the corrected imagery; however, the relationship. Our study showed that perceived health had stronger as- NDVI values derived from RapidEye showed the necessity of making sociation with relatively smaller buffer size in trend analysis, possibly radiometric calibration before the data could be used. In applications of due to the visual impact of green space seen from home on health. The satellite remote sensing data for land observations, some studies did not physical activity instead, 500 m showed the greatest impact, possibly apply for radiometric calibrations. This might have minor impacts on indicating that people might travel further than perceived health par- results if data from only one sensor at one time period was used, such as ticipants to participate in neighborhood activities. These results con- the NDVI values derived from WorldView2 and Landsat sensors. If trast with Houston (2014) who found larger effects on physical activity possible, however, the land observation data from satellite sensors from smaller buffers. Our results may be different because people in should be corrected for impacts from atmosphere and other factors even Barcelona take > 47% of their trips by foot (Ajuntament de Barcelona, though the data are acquired from one sensor for one time period. The 2016) and possibly have slightly larger non-motorized activity space NDVI values derived from the RapidEye imagery showed that if the than in Los Angeles where among the five major counties walking trips imagery was not radiometrically calibrated, the values could be un- range from 8.09 to 14.67% of the total trips taken (Joh et al., 2015). derestimated by 0.15. When data are acquired from multiple themes in Our study also found that the green space characterized using dif- multiple time periods or from multiple sensors, radiometric calibrations ferent spatial resolution satellite data detected similar associations. must be done to adjust for differences created by atmospheric condi- Though Landsat8 data tended to overestimate vegetation index, it did tions or by sensors themselves. not significantly improve the identified associations. Given the higher Our research found that higher vegetation index was largely asso- costs associated with acquiring (including purchasing) finer spatial re- ciated with better perceived health, better mental health and greater solution data, an alternative option is to obtain a relatively coarser physical activity identified through the three sensors across the four spatial resolution remote sensing data, especially when an analysis area buffer sizes. This is similar to our literature search finding that green is too large to be cost bearable for a project. However, higher spatial space was largely associated with better health outcomes; however, resolution imagery like those from WorldView2 allows the identifica- some studies showed non-significant associations (Ali et al., 2017; tion of very small green areas such as those in backyard and front yard. Cohen-Cline et al., 2015; Picavet et al., 2016; Potestio et al., 2009). If a task is to identify objects or green space at that spatial resolution or Those non-significant associations could be due to the difference in to be linked to GPS-enabled activity space, we would suggest using green space characterization. It is typical in built environment studies those fine spatial resolution data for analysis. Though relatively coarser that researchers define the neighborhood as a distance from a re- spatial resolution in Landsat8 data, they still have a spatial resolution of sidence < 500 m. Given the small distance in defining a neighborhood, 30 m, which is considered very high spatial resolution compared to we still found that difference in buffer size had an impact on identified most satellite data which are > 1 or 10 km in spatial resolution. We do strength of associations with health outcomes and that mental health, not suggest using a remote sensing data that is too coarse with sub- perceived health and physical activity had different buffer sizes in stantial mixed pixel information so that the underlying phenomenon is charactering respective strength of association with health outcome. In greatly distorted or cannot be correctly identified. identifying the associations of green cover with increased physical ac- We used images from the three sensors for year 2012, rather than tivity, reduced air pollution and noise, or natural healing from stress for year 2011 when the survey was conducted. This was due to the reduction in the immediate neighborhood, we suggest that the first availability of the best images for the entire city of Barcelona: we fo- option is to identify activity space of an individual through GPS-enabled cused on date closeness from the three sensors for best available images 168 J.G. Su, et al. Environment International 126 (2019) 162–170 in summer when vegetation showed greatest greenness. All the remote (Positive Health Effects of the Natural Outdoor Environment in Typical sensing data were collected in July or early August in 2012, which Populations in Different Regions in Europe) (www.phenotype.eu). made it possible for us to get similar surface reflectance across the three sensors after radiometric calibration. We do not expect significant References changes in vegetation greenness in Barcelona between the two neigh- boring years in summer when vegetation is greenest. The identified Akpinar, A., 2016. How is quality of urban green spaces associated with physical activity impact would remain largely the same if the year 2011 remote sensing and health? Urban For. Urban Green. 16, 76–83. Ali, O., Di Nardo, F., Harrison, A., Verma, A., 2017. The link between perceived char- data were available. acteristics of neighbourhood green spaces and adults' physical activity in UK cities: Our study relied on cross-sectional data, and we could not estimate analysis of the EURO-URHIS 2 study. Eur. J. Pub. Health 27, 761–765. temporal relationships between the NDVI exposure and health out- Almanza, E., Jerrett, M., Dunton, G., Seto, E., Pentz, M.A., 2012. A study of community design, greenness, and physical activity in children using satellite, GPS and accel- comes. In addition, we could not identify whether self-selection had any erometer data. Health Place 18, 46–54. impact on the identified health outcomes. Further the identified health Ajuntament de Barcelona, 2016. Dades bàsiques de mobilitat informe. from. http:// outcomes were obtained from a survey, which could have bias issues ajuntament.barcelona.cat/ecologiaurbana/sites/default/files/DADES-BASIQUES- MOBILITAT_2015_Resum.pdf, Accessed date: 26 September 2018. (Dal Grande et al., 2016; Regber et al., 2013). The health outcomes Casey, J.A., James, P., Cushing, L., Jesdale, B.M., Morello-Frosch, R., 2017. Race, eth- across the 4000 participants were not identified by physicians, although nicity, income concentration and 10-year change in urban greenness in the United the self-assessed and mental health outcomes have been well validated States. Int. J. Environ. Res. Public Health 14. Cohen-Cline, H., Turkheimer, E., Duncan, G.E., 2015. Access to green space, physical in other studies. We also recognize that without objective ascertain- activity and mental health: a twin study. J. Epidemiol. Community Health 69, ment of physical activity, we cannot with definitive certainty ascribe 523–529. health or behavioral changes to actual use or visual contact with the Craig, C.L., Marshall, A.L., Sjöström, M., Bauman, A.E., Booth, M.L., Ainsworth, B.E., green space. In this analysis, we used home address as single location of Pratt, M., Ekelund, U.L.F., Yngve, A., Sallis, J.F., Oja, P., 2003. International physical activity questionnaire: 12-country reliability and validity. Med. Sci. Sports Exerc. 35, green space exposure; however, we understand that exposure could 1381–1395. occur in the community, at work, at home, at school and elsewhere. Crouse, D.L., Pinault, L., Balram, A., Hystad, P., Peters, P.A., Chen, H., van Donkelaar, A., Significant limitations exist in this study and the majority of the current Martin, R.V., Ménard, R., Robichaud, A., Villeneuve, P.J., 2017. Urban greenness and mortality in Canada's largest cities: a national cohort study. Lancet Planet. Health 1, studies with the ability to identify personal activity space and exposure e289–e297. (Guarnieri and Balmes, 2014). Future work is needed to assess time- Dadvand, P., de Nazelle, A., Figueras, F., Basagana, X., Su, J., Amoly, E., Jerrett, M., activity based exposure to green space rather than home-location based Vrijheid, M., Sunyer, J., Nieuwenhuijsen, M.J., 2012. Green space, health inequality and pregnancy. Environ. Int. 40, 110–115. exposure to reduce exposure misclassification. The availability of Dadvand, P., Nieuwenhuijsen, M.J., Esnaola, M., Forns, J., Basagaña, X., Alvarez- Google Street View (GSV) as a street-level urban greenery assessment Pedrerol, M., Rivas, I., López-Vicente, M., De Castro Pascual, M., Su, J., Jerrett, M., tool (Li et al., 2015) could help us assess time-activity based green Querol, X., Sunyer, J., 2015a. Green spaces and cognitive development in primary schoolchildren. Proc. Natl. Acad. Sci. 112, 7937–7942. exposure. Activity space-based exposure models could help us under- Dadvand, P., Rivas, I., Basagaña, X., Alvarez-Pedrerol, M., Su, J., De Castro Pascual, M., stand the optimal buffer distances in activity space of various groups of Amato, F., Jerret, M., Querol, X., Sunyer, J., Nieuwenhuijsen, M.J., 2015b. The as- participants and whether the resolution of remote sensing sensors might sociation between greenness and traffic-related air pollution at schools. Sci. Total Environ. 523, 59–63. have impact on identified health outcomes. Dadvand, P., Bartoll, X., Basagana, X., Dalmau-Bueno, A., Martinez, D., Ambros, A., Cirach, M., Triguero-Mas, M., Gascon, M., Borrell, C., Nieuwenhuijsen, M.J., 2016. 5. Conclusion Green spaces and general health: roles of mental health status, social support, and physical activity. Environ. Int. 91, 161–167. Dal Grande, E., Chittleborough, C.R., Campostrini, S., Taylor, A.W., 2016. Bias of health Though the advantage of using satellite data for green space char- estimates obtained from chronic disease and risk factor surveillance systems using acterization, cautions should be taken in selecting a reasonable buffer telephone population surveys in Australia: results from a representative face-to-face size of impact. We suggest investigating underlying mechanics for a survey in Australia from 2010 to 2013. BMC Med. Res. Methodol. 16, 44. Davies, H.W., Vlaanderen, J., Henderson, S., Brauer, M., 2009. Correlation between co- health outcome in order to identify an optimal distance of impact. Our exposures to noise and air pollution from traffic sources. Occup. Environ. Med. 66, study also suggests that sensor spatial resolution did not significantly 347–350. improve or reduce a health outcome estimate. Given financial strains Figueras, F., Meler, E., Iraola, A., Eixarch, E., Coll, O., Figueras, J., Francis, A., Gratacos, E., Gardosi, J., 2008. Customized birthweight standards for a Spanish population. and other limits, data from a coarser spatial resolution sensor might be Eur. J. Obstet. Gynecol. Reprod. Biol. 136, 20–24. a feasible alternative to the data derived from finer spatial resolution Gascon, M., Triguero-Mas, M., Martinez, D., Dadvand, P., Forns, J., Plasencia, A., sensors for green space characterization. However, more highly re- Nieuwenhuijsen, M.J., 2015. Mental health benefits of long-term exposure to re- sidential green and blue spaces: a systematic review. Int. J. Environ. Res. Public solved satellite data will reduce the problem of mixed pixels (i.e., Health 12, 4354–4379. misclassification of green space exposure due to pixels with nearly Generalitat de Catalunya, 2010. Renta familiar disponible bruta. In: Generalitat de equal areas that are green versus those that are not). Reducing the Catalunya. Goldberg, D., 1978. Manual of General Health Questionnaire. NFER publishing, mixed pixel problem should therefore reduce one potential component Windsor, UK. of measurement error that could bias epidemiological inference. Gomez, L.F., Sarmiento, O.L., Parra, D.C., Schmid, T.L., Pratt, M., Jacoby, E., Neiman, A., We also suggest that in order to best characterize green space in- Cervero, R., Mosquera, J., Rutt, C., Ardila, M., Pinzon, J.D., 2010. Characteristics of cluding its accessibility and quality, satellite data should be combined the built environment associated with leisure-time physical activity among adults in Bogota, Colombia: a multilevel study. J. Phys. Act. Health 7, S196–S203. with other evaluations of quality and accessibility. In this way, re- Gordon-Larsen, P., Nelson, M.C., Page, P., Popkin, B.M., 2006. Inequality in the built searchers will be able to assess the degree of vegetation, its accessi- environment underlies key health disparities in physical activity and obesity. bility, and quality so the potential health effects of exposure to green Pediatrics 117, 417–424. Gradinaru, C., Herban, I.S., Gabor, C., 2016. Development of urban green space mon- space can be more accurately identified. Such characterization may also itoring technique with remote sensing and its application. In: Proceedings of the help to illuminate the likely pathways from green space to various International Conference on Numerical Analysis and Applied Mathematics 2015 health outcomes, such as stress reduction, reduction of noise and air (Icnaam-2015), pp. 1738. Guarnieri, M., Balmes, J.R., 2014. Outdoor air pollution and asthma. Lancet 383, pollution, or physical activity. 1581–1592. Houston, D., 2014. Implications of the modifiable areal unit problem for assessing built Acknowledgements environment correlates of moderate and vigorous physical activity. Appl. Geogr. 50, 40–47. IPAQ Research committee, 2005. Guidelines for Data Processing and Analysis of the We are grateful for the contribution of the Barcelona Public Health International Physical Activity Questionnaire (IPAQ). Agency to providing health outcome data for this analysis. This study Joh, K., Chakrabarti, S., Boarnet, M.G., Woo, A., 2015. The walking renaissance: a longitudinal analysis of walking travel in the greater Los Angeles area, USA. was funded by the European Community's Seventh Framework Sustainability Basel 7, 8985–9011. Programme (Grant No. 282996) as part of the PHENOTYPE project 169 J.G. Su, et al. Environment International 126 (2019) 162–170 Kaczynski, A.T., Potwarka, L.R., Saelens, B.E., 2008. Association of park size, distance, Sugiyama, T., Francis, J., Middleton, N.J., Owen, N., Giles-Corti, B., 2010. Associations and features with physical activity in neighborhood parks. Am. J. Public Health 98, between recreational walking and attractiveness, size, and proximity of neighbor- 1451–1456. hood open spaces. Am. J. Public Health 100, 1752–1757. Kriegler, F.J., Malila, W.A., Nalepka, R.F., Richardson, W., 1969. Preprocessing trans- Thompson, C.W., Aspinall, P., Roe, J., Robertson, L., Miller, D., 2016. Mitigating stress formations and their effects on multispectral recognition. In: Proceedings of the Sixth and supporting health in deprived urban communities: the importance of green space International Symposium on Remote Sensing of Environment, pp. 97–131. and the social environment. Int. J. Environ. Res. Public Health 13. Li, X., Zhang, C., Li, W., Ricard, R., Meng, Q., Zhang, W., 2015. Assessing street-level Triguero-Mas, M., Dadvand, P., Cirach, M., Martinez, D., Medina, A., Mompart, A., urban greenery using Google Street View and a modified green view index. Urban Basagana, X., Grazuleviciene, R., Nieuwenhuijsen, M.J., 2015. Natural outdoor en- For. Urban Green. 14, 675–685. vironments and mental and physical health: relationships and mechanisms. Environ. Maas, J., Verheij, R.A., Groenewegen, P.P., de Vries, S., Spreeuwenberg, P., 2006. Green Int. 77, 35–41. space, urbanity, and health: how strong is the relation? J. Epidemiol. Community Ulmer, J.M., Wolf, K.L., Backman, D.R., Tretheway, R.L., Blain, C.J.A., O'Neil-Dunne, Health 60, 587–592. J.P.M., Frank, L.D., 2016. Multiple health benefits of urban tree canopy: the Mytton, O.T., Townsend, N., Rutter, H., Foster, C., 2012. Green space and physical ac- mounting evidence for a green prescription. Health Place 42, 54–62. tivity: an observational study using health survey for England data. Health Place 18, Vatseva, R., Kopecka, M., Otahel, J., Rosina, K., Kitev, A., Genchev, S., 2016. Mapping 1034–1041. urban green spaces based on remote sensing data: case studies in Bulgaria and Picavet, H.S.J., Milder, I., Kruize, H., de Vries, S., Hermans, T., Wendel-Vos, W., 2016. Slovakia. In: Int Conf Cartogr Gis, pp. 569–578. Greener living environment healthier people? Exploring green space, physical ac- Vilagut, G., Ferrer, M., Rajmil, L., Rebollo, P., Permanyer-Miralda, G., Quintana, J.M., tivity and health in the Doetinchem cohort study. Prev. Med. 89, 7–14. Santed, R., Valderas, J.M., Domingo-Salvany, A., Alonso, J., 2005. The Spanish ver- Potestio, M.L., Patel, A.B., Powell, C.D., McNeil, D.A., Jacobson, R.D., McLaren, L., 2009. sion of the short form 36 health survey: a decade of experience and new develop- Is there an association between spatial access to parks/green space and childhood ments. Gac. Sanit. 19, 135–150. overweight/obesity in Calgary, Canada? Int. J. Behav. Nutr. Phys. Act. 6, 77. Villeneuve, P.J., Jerrett, M., Su, J.G., Weichenthal, S., Sandler, D.P., 2018. Association of Regber, S., Novak, M., Eiben, G., Lissner, L., Hense, S., Sandstrom, T.Z., Ahrens, W., residential greenness with obesity and physical activity in a US cohort of women. Marild, S., 2013. Assessment of selection bias in a health survey of children and Environ. Res. 160, 372–384. families - the IDEFICS Sweden-study. BMC Public Health 13, 418. Ware, J.E., Sherbourne, C.D., 1992. The MOS 36-item short-form health survey (SF-36): I. Roe, J.J., Aspinall, P.A., Thompson, C.W., 2017. Coping with stress in deprived urban Conceptual framework and item selection. Med. Care 30, 473–483. neighborhoods: what is the role of green space according to life stage? Front. Wolch, J., Jerrett, M., Reynolds, K., McConnell, R., Chang, R., Dahmann, N., Brady, K., Psychol. 8. Gilliland, F., Su, J.G., Berhane, K., 2011. Childhood obesity and proximity to urban Roman-Viñas, B., Serra-Majem, L., Hagströmer, M., Ribas-Barba, L., Sjöström, M., Segura- parks and recreational resources: a longitudinal cohort study. Health Place 17, Cardona, R., 2010. International physical activity questionnaire: reliability and va- 207–214. lidity in a Spanish population. Eur J Sport Sci 10, 297–304. Wolch, J.R., Byrne, J., Newell, J.P., 2014. Urban green space, public health, and en- Sánchez-López, M., Dresch, V., 2008. The 12-item General Health Questionnaire (GHQ- vironmental justice: the challenge of making cities ‘just green enough’. Landsc. Urban 12): reliability, external validity and factor structure in the Spanish population. Plan. 125, 234–244. Psicothema 20, 839–843. Wood, L., Hooper, P., Foster, S., Bull, F., 2017. Public green spaces and positive mental Su, J.G., Jerrett, M., de Nazelle, A., Wolch, J., 2011. Does exposure to air pollution in health - investigating the relationship between access, quantity and types of parks urban parks have socioeconomic, racial or ethnic gradients? Environ. Res. 111, and mental wellbeing. Health Place 48, 63–71. 319–328. 170