Incorporating tree diversity for a better understanding of urban form-air quality relationships through mobile monitoring

doi:10.21203/rs.3.rs-2245738/v1

Download PDF

Research Article

Incorporating tree diversity for a better understanding of urban form-air quality relationships through mobile monitoring

https://doi.org/10.21203/rs.3.rs-2245738/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Context: Air pollution is a major threat to landscape ecology and public health. The conventional LUR (Land Use Regression) method has been developed mostly based on 2-D urban form and emission source information. Besides, the effects of tree diversity on air quality have not been adequately addressed.

Objectives: This study explores the integration of 2-D and 3-D urban form metrics and examines the impact of tree diversity in urban form-air quality relationships.

Methods: We introduce these predictors into the LUR tools using unique NO₂ datasets collected through opportunistic mobile monitoring in the Bronx, New York, and Oakland, California, and further apply lacunarity to investigate the spatial scale sensitivity for regression predictors.

Results: The lacunarity-optimized model helps to reduce the computation burden by finding the upper limit of the spatial heterogeneity of predictors while keeping the model accuracy in both Bronx (R²≈0.50) and Oakland (R²≈0.79). Furthermore, although deciduous trees are surrounded by the highest NO₂ concentrations (9.73 ppb and 3.61 ppb in the Bronx and Oakland, respectively), the increase in tree diversity could facilitate the reduction of NO₂ concentration. However, according to the non-monotonic and marginal effects of tree diversity on NO₂ concentration, a higher level of tree diversity is not always better.

Conclusions: It is reasonable to seek a balance between the diversity and dominance of tree species to effectively improve air quality on the city scale. The findings are useful for both environmental scientists striving for better air quality and urban planners that care for the well-being of cities.

Air pollution

land use regression

urban form

urban forest

opportunistic mobile monitoring

lacunarity

Air pollution is the most significant environmental health risk worldwide. It is associated with increased mortality and a wide range of serious diseases (Haley et al., 2009; Eilstein, 2010; Kheirbek et al., 2016). A challenge in predicting urban air pollution is ascribed to the spatiotemporal heterogeneity of pollutants distribution and complex urban microenvironments (Apte et al., 2017). Due to their high cost, existing stationary regulatory monitoring networks are sparse and cannot capture granular variations of pollution (Hofman et al., 2018; Do et al., 2020). In recent years, mobile air quality monitoring has emerged as an effective practice to map hyperlocal air pollution in dense urban environments (Van den Bossche et al., 2015; DeSouza et al., 2020).

High-resolution mobile monitoring provides unique opportunities for environmental and urban scientists to understand the interactions between air quality, human activities, and urban environments from a holistic perspective of urban ecology at a fine scale (Cummings et al., 2021). Although LUR (Land Use Regression) is the most widely-used tool for air quality data interpretation and prediction due to the relatively lower level of demand for input data requirements compared to dispersion and chemical transport models (Hoek, 2017; Messier et al., 2018; Shi et al., 2016; Saha et al., 2019), the current LUR literature often neglects two essential aspects of urban environments. One is the impact of tree diversity. This is problematic because urban forests play a pivotal role in ameliorating air quality and improving urban ecosystem health (Roeland et al., 2019). The One Health initiative, which emphasizes the connection between the health of people, animals, and the environment, recognizes the key role that trees play in regulating urban ecology (CDC, 2022). Another is the impact of 3-D urban form, which has been shown to be more important in estimating intra-urban PM_2.5 than their 2-D counterparts (Tian et al., 2022).

The role of tree diversity is essential in supporting ecosystem services and air quality improvement, particularly for the abatement of airborne particulate pollution (Manes et al., 2014). For example, the UFORE (Urban FORest Effects) model, developed by the U.S. Forest Service to quantify multiple urban forest ecosystem services, identified tree diversity as one of the most effective methods of improving local air quality (Nowak et al., 2006; Saunders et al., 2011). To illustrate, Sicard et al. (2018) devised a novel Species Air Quality Index (S-AQI) of suitability to air quality improvement for tree diversity and suggested that city planners should select species with an S-AQI > 8. Manes et al. (2014) considered tree diversity in different climatic conditions to confirm the crucial role of trees in supporting significant ecosystem services to improve air quality. However, how and to what extent tree diversity can impact air pollution distribution at different spatial scales is still largely underexplored.

3-D urban form, consisting of both horizontal and vertical elements, can directly influence urban ventilation and the dispersion of pollutants (Peng et al., 2021; Yang et al., 2020), and their influence has also been verified in different urban scenarios, such as street canyons, neighborhoods, and communities, through idealized experiment models (Yuan et al., 2019; Eeftens et al., 2013; Hang et al., 2012). For example, SVF (Sky View Factor) can influence urban ventilation resistance, shape ventilation corridors, and determine the dispersion of air pollutants (Fang and Zhao, 2022). Tang et al. (2013) used building heights and geometry to enhance the estimation of land use-related variables and the pollution dispersion fields for long-term air pollutants. Edussuriya et al. (2011) found that air pollution concentrations were impacted by the AR (Aspect Ratio), building volume, and building height variation.

Furthermore, tree diversity and urban form also have synthetic effects on air quality (Pugh et al., 2012). For example, deposition and dispersion are tightly coupled to the 3-D urban form and the synoptic-scale flow. As the porosity of the barrier increases, the effective path-length decreases, and the opportunity for the removal of particles by deposition increases (Tong et al., 2016). The AR significantly affects pollutant dispersion because of alterations in airflow patterns (Zhong et al., 2016). When horizontal length scales and AR are small and residence times are short, there is little opportunity for deposition to become effective. Green walls in street canyons with a large AR may make appreciable differences in ground-level concentrations due to deeper or narrower street canyons (Pugh et al., 2012; Buccolieri et al., 2009). Notwithstanding, limited studies incorporated both tree diversity and 3-D urban form factors into the air quality research.

Additionally, there is a lack of widely recognized theoretical guidelines regarding the optimum buffer distances for air quality estimation. Most prior studies relied on arbitrary multiple buffer distances, which could not interpret the spatial scale sensitivity, and were prone to MAUP (Modifiable Areal Unit Problem) that biases the final results due to the different minimum analysis units (Dark and Bram, 2007), and be time-consuming for model construction. Therefore, researchers applied lacunarity, which measures the spatial heterogeneity of predictors depending on the image’s texture at varying scales (Gefen et al., 1983; Plotnick et al., 1996).

Our study targets bridging the current research gaps in urban air pollution prediction, including the underrepresentation of urban forestry information, the lack of 3-D urban form involvement, and the inefficiency of the traditional buffer sizing method. To illustrate, we integrate tree diversity and 3-D urban form into existing 2-D information-dependent LUR tools to identify the most influential predictors and to sheds light on the interactions between air pollution, urban form, and urban forest, which demonstrates our multi-disciplinary efforts to attain optimal health for the entire urban ecological system. Hyperlocal NO₂ levels were collected through an opportunistic mobile monitoring campaign in the Bronx, NY. To explore the impacts of urban form and urban forest in different cities, we transferred our air quality modeling methodology to Oakland, California, a city on the U.S. west coast with a distinct urban environment from the Bronx, to access the model’s general applicability. We also conducted the lacunarity analysis to determine the upper bounds of predictors’ buffer sizes while keeping the fixed buffer sizing as a benchmark. Models developed using lacunarity-optimized and conventional fixed buffers were contrasted to reveal the effects of spatial heterogeneity on air quality prediction. The proposed lacunarity method is highly scalable and transferrable to other air quality prediction applications. Our study is relevant to environmental scientists and ecologists striving for better air quality, and to urban planners and decision-makers to strategically manage urban forests.

2.1 Research areas

Figure 1 displays the mobile NO₂ (451,322 mobile measurements) in the Bronx, which has the highest asthma hospitalization rate in New York City, with a dense road network and other noxious land uses that are significant sources of air pollution (Gorai et al., 2014; Maciejczyk et al., 2004; Corburn et al., 2006). The NO₂ data has a mean value of 9.45 ppb (parts per billion) and a standard deviation of 3.96 ppb collected between September 10th to December 17th, 2021 as part of the City Scanner research initiative at the Massachusetts Institute of Technology (senseable.mit.edu/cityscanner). Five vehicles operated by NYC’s Citywide Administrative Services (NYC-DCAS), including four road maintenance trucks and one sanitation vehicle, were equipped with City Scanner low-cost sensing nodes with Alphasense NO₂-A4 electrochemical gas sensors, high-accuracy GPS, and temperature and humidity sensors (Mora et al., 2019). More details can be seen in SI (Supporting Information) Table S1: Mobile monitoring NO₂. The NO₂ measurements in Oakland refer to Apte et al. (2017).

2.2 Research design

For the data preparation, the mobile monitoring NO₂ data was first calibrated against a reference air quality station in the Bronx and was aggregated by the road segment to reduce the systematic noise and biases (Section 2.2.1). Then, we calculated the lacunarity value for different land use types and street trees. Tree diversity was characterized by SHI (Shannon Index), the richness and evenness degree of tree species, SII (Simpson Index), the dominance of tree species, and DBH (Diameters at Breast Height) (Section 2.2.2). Next, we calculated the 2-D (Section 2.2.3) and 3-D (Section 2.2.4) urban form factors to characterize the urban landscape comprehensively. The upper bound for predictor aggregation was calculated from lacunarity values (Section 2.2.5). Combined with other covariates, these variables were used to predict NO₂ through two modeling techniques: lacunarity-optimized and conventional LURs. After feature selection, we constructed the RF-LUR (Random Forest-Land Use Regression) model to analyze their correlation and interpret the final results through feature importance, influential directions, and marginal effects (Section 2.2.6). Figure S1 shows the research flowchart. Table S2 shows all the predictors according to the potential NO_x emission sources and factors affecting their distribution and dispersion (Messier et al., 2018; Apte et al., 2017).

2.2.1 NO₂ preprocessing

Before using the mobile monitoring NO₂ for further analysis, we calibrated the data using minute-by-minute reference EPA (Environmental Protection Agency) station data. Multiple calibration algorithms are tested, including linear and non-linear ones, to best reproduce the reference data. Calibration algorithm performances are stringently evaluated using 10% hold-out samples 100 times. Calibrated City Scanner NO₂ data is highly correlated with the reference signals using the best-performing algorithm (average external R² = 0.92), as shown in Figure S2. To further increase the transferability of the research, we also constructed the model using the NO₂ data in Oakland, CA, which was collected by Google Street View vehicles from May 28th, 2015 to December 21st, 2017 with 38,252 mobile measurements (Apte et al., 2017). All calibrations were performed “through the probe” by connecting the calibration system to the car at the sample inlet and maximizing the temporal correlations between the individual pollutants on a daily basis (Brantley et al., 2014). To remove the impacts of relative humidity on NO₂ measurements, values in both Bronx and Oakland were removed for relative humidity above 90% due to the growth of the hygroscopic particle, leading to mass overestimations (Goin et al., 2021). To keep compatible and alleviate the impact of the “Spike” signal from low-cost sensors (e.g. keep reporting high values for approximately 30s when a truck pass by), all NO₂ was processed and aggregated by the midpoint of each road segment and used median value to represent the NO₂ at 30m spacing to mitigate the measurement uncertainty at point locations in time as reported in previous research (Apte et al., 2017; DeSouza et al., 2020; Gani et al., 2021; Goin et al., 2021).

2.2.2 Tree diversity

Tree diversity is critical in determining pollutant removal in the built environment. We extracted data for 66,294 and 38,564 trees along the streets with 129 and 195 species in Bronx and Oakland, respectively, from OpenTrees.org (Figure S3: Spatial distribution of tree species in the Bronx and Oakland). To further explore how tree diversity may affect NO₂ concentration, we summarized the median NO₂ value around each tree species. Previous ecological studies strongly recommended using both SHI (Shannon Index) and SII (Simpson Index) to account for species richness and relative abundance, which could provide the most complete mathematical description of the data (Gorelick, 2006; Ricotta and Szeidl, 2006; Zahl, 1977; Galle et al., 2021). The SHI was initially proposed by Shannon (1948) to quantify the level of the mixture and can be extended to measure tree species entropy, which increases as the community’s richness and evenness increase. The SII was introduced by Simpson (1949) to measure the degree of dominance, such as the dominant distribution for specific tree species. Both of them can be used to measure tree diversity. More details about SHI and SII can be seen in Section S1: Tree diversity. The DBH (Diameter at Breast Height) is also aggregated based on the median value of diameters for all tree species.

2.2.3 2-D urban form metrics

To comprehensively characterize the urban context, we calculated 2-D urban form metrics in the study area through the integration of land use and landscape metrics. For example, we calculated six landscape metrics, including CA (Class Area), ED (Edge Density), PD (Patch Density), LPI (Largest Patch Index), LSI (Landscape Shape Index), and AI (Aggregation Index), for each land use type (i.e., building, roads, green space, and water, shown in Figure S4: land use in Bronx and Oakland). These metrics capture the spatial coverage, fragmentation, patch dominance, and shape complexity of land use patterns (Liu et al., 2017; Liu et al., 2018c; Buyantuyev et al., 2010) that could shape the spatiotemporal dispersion of air pollutants in urban landscapes (Borrego et al., 2006; Bechle et al., 2011; Martins, 2012; Cho and Choi, 2014; Rodríguez et al., 2016). Table S3 describes each metric in detail (McGarigal, 2015). Interactions between land use and landscape metrics also likely impact air pollution. Therefore, in addition to using each metric as an independent covariate, we also added their interaction terms as predictors (Tian et al., 2020; Tian and Yao, 2022). This gives us 24 urban form metrics (four land uses times six landscape metrics) in total. All landscape metrics were calculated in the open-source software Fragstats 4.2.

2.2.4 3-D urban form metrics

We used SVF (Sky View Factor), H_Var (Height Variation), and AR (Aspect Ratios) to capture 3-D urban form factors as their synthetic effects with tree diversity on air pollution distribution. More technical details and backgrounds of 3-D urban form metrics can be found in the Supporting Information Sections from S2 to S4.

The SVF quantifies the openness to the sky of a given location, which reflects the degree of urban ventilation affecting air pollution dispersion (Liu et al., 2016). In this study, we used the DEM as the background elevation and building height to calculate SVF. SVF ranges between 0 and 1 where close to 1 indicates that almost the entire hemisphere is visible (e.g., planes and peaks), and close to 0 means almost no sky is visible (e.g., deep sinks and valleys) (Zakšek et al., 2011). The H_Var describes the variation of the building height, which impacts the distribution of wind flow and the corresponding ventilation corridor (Liu et al., 2016). The AR represents the ratio of building height to street width, which gives an idea of the change of surface-air interfaces due to the vertical dimensions of buildings compared to the flat terrain (Grimmond and Oke, 1999), which plays an essential role in the interaction between tree diversity, urban form, and air quality. For example, an increase in AR can lead to decreased dilution of CO from ground-based traffic emission sources (Tang and Wang, 2007). Air quality improves with a decrease in AR due to a larger vertical exchange of air pollutants between street roofs suggesting that an open central street can result in better air quality (Shen et al., 2017).

2.2.5 Lacunarity

Compared with previous studies using a series of arbitrary buffer distances to construct the model predictors (Kerckhoffs et al., 2022; Zhang et al., 2022; Li et al., 2021), we apply lacunarity, as proposed by Mandelbrot and Mandelbrot (1982), to describe the texture of the landscape. To illustrate, objects with low lacunarity are homogeneous and translationally invariant because of the same gap sizes. In contrast, high-lacunarity objects have a wide range of gap sizes which are usually heterogeneous and translationally variant (Plotnick et al., 1993). As a result, for variable aggregation, increasing the buffer distance beyond a certain level would not result in any substantive impact on the statistical results, because the buffer distance surpasses the maximum level of the repeating patterns in the landscape (Myint and Lam, 2005; Roces-Diaz et al., 2015; Hoechstetter et al., 2011). Therefore, lacunarity could be used to define the upper limit of the spatial variability when quantifying the landscape patterns (Labib et al., 2020; Roces-Diaz et al., 2015; Hoechstetter et al., 2011). In this study, we calculate the lacunarity value for land uses and street trees that have proven to be correlated with NO₂ distribution (Apte et al., 2017) and select the flattened value of lacunarity as the upper bound of the buffer distance. This way, we can capture the maximum ranges of spatial heterogeneity of predictors and reduce the computation burden, particularly for predictors aggregation at large buffer distances. Consequently, it provides a comprehensive and less resource-intensive solution to address scale effects for model construction (Labib et al., 2020). Details about calculating lacunarity can be seen in Section S5: Lacunarity.

2.2.6 Model construction

We conducted two sets of models: the conventional LUR model, using a series of fixed buffer distances as in previous studies (Kerckhoffs et al., 2022; Zhang et al., 2022; Li et al., 2021), and lacunarity-optimized, using a series of the same buffer distances but setting the upper bounds according to the flattened point of lacunarity value. To lift linear assumptions, researchers have improved LUR by incorporating machine learning models as they can account for non-linear relationships, deal with multidimensional independent variables, and allow for complex interactions between the various predictors (Ren et al., 2020). Previous researchers have found that RF (Random Forest) models have been at least as accurate and, in many cases, more accurate than linear models (Tian et al., 2020). Compared to other non-parametric techniques, such as ANN (Artificial Neural Network), RF also provides metrics that capture the relative importance of the different independent covariates in predicting the outcome. Therefore, this study combined RF and LUR for model construction. More details can be seen in Figure S5: The procedures for constructing the RF model.

Considering potential multicollinearity between the covariates, we use the Greedy Stepwise algorithm to filter correlated covariates and select the subset of the most significant covariates as candidates in the RF-LUR model. The preferred subsets of features are highly correlated with the dependent variable while having low inter-correlations (Brownlee, 2019). To avoid overfitting, we apply the 10-fold cross-validation technique to tune the hyper-parameters of our model. To illustrate, the original sample is randomly partitioned into 10 equal subsamples. A single subsample is retained as the validation data for testing the model, and the remaining nine subsamples are used as training data.

Besides the R² and RMSE (Root Mean Squared Error), we plot feature importance, SHAP (SHapley Additive exPlanations) value, and PDP (Partial Dependence Plot) to further interpret the RF-LUR models. To illustrate, feature importance is computed as the decrease in node impurity weighted by the probability of reaching that node, namely MDI (Mean Decrease in Impurity). It counts the times a feature is used to split a node weighted by the number of samples it splits (Perrier, 2015). SHAP value was proposed by Lundberg and Lee (2017) as a united approach to explaining the output of any machine learning model because it can: 1). show how much each predictor contributes, positively or negatively, to the target variable; 2). get its own set of SHAP values for each observation; 3). be calculated for any tree- or non-tree-based models. The PDP shows the predictors’ marginal effects on the predicted outcome (Friedman, 2001), and it demonstrates whether the relationship between the target and a feature is linear, monotonic, or more complex and how the average prediction changes as the specific predictor changes.

3.1 Descriptive analysis

3.1.1 Lacunarity results

Figure 2 shows the lacunarity value for different land uses and street trees. To illustrate, the “flattened point” (where the value flattens out and reaches the maximum value of heterogeneity pattern) would be used as the bounded buffer sizes (labeled as vertical lines) for the lacunarity-optimized model. For example, in the Bronx, the flattened point for “Buildings” is 34, so the upper bound is 1020m (34\(\times\)30m, 30m represents the spatial resolution of box size). Therefore, we select 100m, 300m, 500m, and 1020m as the buffer distances for calculating urban form metrics for “Buildings”. If the product of lacunarity value and the box size is larger than 5000m, we set 5000m as the upper bound because it has been used as the maximum potential influencing range in previous air quality models (Naughton et al., 2018; Shi et al., 2017; Meng et al., 2015). The procedures for detecting the flattened value and selecting upper bounded buffer sizes keep the same in Oakland. Details can be seen in Table S4 (Selected features for lacunarity-optimized models in the Bronx and Oakland).

As can be seen from Fig. 2, the flattened value for green space (198) and water (200) is higher than that for roads (18) and buildings (34), which indicates significant gaps, uneven distribution, and a higher level of spatial heterogeneity for green space and water. The roads have the smallest flattened value compared to other land uses, suggesting that road networks have relatively homogeneous spatial distribution in the urban context. This reflects the relatively dense distribution of road networks and sparse distribution of water and green space in the Bronx. Likewise, the distribution of lacunarity values for land uses and street trees maintain a similar trend in Oakland, where buildings and roads are more homogeneously distributed than the greenspace and water.

(The “Overall landuse” means we combine different land use types into a single image. The moving window sizes for land use and street trees are 30m and 10m, respectively.)

3.1.2 Tree diversity and NO₂ distribution

In Figs. 3 (a) and (b), we summarized the median value of NO₂ concentration around deciduous, semi-deciduous, and evergreen trees aggregated based on the buffer distances of 310m (31\(\times\)10m) and 450m (45\(\times\)10m) (flattened value of street trees, in Fig. 2). As can be seen, deciduous trees are surrounded by the highest level of NO₂ concentration (9.73 ppb in the Bronx and 3.61 ppb in Oakland). Besides, the Bronx also has an overall higher level of NO₂ concentration (9.72 ppb) than Oakland (3.67 ppb) for all trees, which could be attributed to the reason that the Bronx has a higher percentage of deciduous trees (89.92%) than Oakland (38.97%). This accords with previous findings that evergreen trees were more effective in reducing annual NOx concentration than deciduous trees (Tiwari and Kumar, 2020).

In Figs. 3 (c) and (d), we summarized the specific tree species with NO₂ concentration levels ranging from the 5th to 95th percentile. In the Bronx, the “Callery pear”, a deciduous tree with oval leaves, has the highest level of NO₂ concentration (10.28 ppb). The “northern red oak”, regarded as “one of the most handsome, cleanest, and stateliest trees in North America” (Foundation, 2022), enjoys a relatively lower level of NO₂ concentration (9.45 ppb). In Oakland, the worst air quality is found around “Pistacia Chinensis” (11.92 ppb), with deciduous leaves and absent terminal leaflet, while the “Ligustrum lucidum”, an evergreen tree with opposite and glossy dark green leaves, is surrounded by the lowest level of NO₂ concentrations (1.98 ppb). In addition, the SII (Simpson Index) and SHI (Shannon Index) augment with the increase of the buffer distances in both cities. Details can be seen in Table S5 (Summary statistics of tree diversity in Bronx and Oakland). The findings suggest that it is necessary to incorporate tree diversity into air quality models as the NO₂ concentration varies around different tree species.

(Figures (a) and (b) show the median NO₂ concentration for all deciduous, semi-deciduous, and evergreen trees in Bronx and Oakland, respectively. Figures (c) and (d) indicate from the 5th to 95th percentile of NO₂ concentration around specific tree species in Bronx and Oakland, respectively.)

3.2 Model performance

Table 1 shows R² and RMSE (Root Mean Squared Error) of the lacunarity-optimized and conventional LUR models. It can be seen that the performance of the lacunarity-optimized models is comparable to that of the conventional LUR models in both Bronx (R² = 0.51 vs. R² = 0.52) and Oakland (R² = 0.79 for both models). However, the number of predictors has significantly reduced from 235 (conventional LUR models) to 158 (lacunarity-optimized models) in the Bronx and 161 in Oakland (Table S4 and Table S6: Summary statistics of predictors for conventional LURs in the Bronx). It demonstrates that the lacunarity analysis is advantageous in reducing or avoiding the repetitive aggregation for predictors, especially for calculating predictors with large buffer distances, which is time- and labor-intensive. Therefore, instead of applying a series of arbitrary buffer distances, we can calculate the lacunarity value first to find the upper bounds of the buffer distance for each predictor and leave out the distances that are larger than that bounds.

We also compared the performance through ANN (Artificial Neural Network) and MLR (Multiple Linear Regression) in Table 1. We find that RF overall outperforms ANN and MLR for both lacunarity-optimized and conventional LUR models in Bronx and Oakland with higher R² and lower RMSE. Besides, the accuracy of R² of LUR-RF has reduced to 0.47 and 0.48 without considering tree diversity in lacunarity-optimized and conventional LUR models, respectively, in the Bronx, and reduced to 0.77 for both models in Oakland (Table S7: Model performance without considering tree diversity), which confirms the necessity of incorporating tree diversity to improve model accuracy.

Table 1

Model performances
Buffer selection Model		Lacunarity- optimized	Conventional- LUR
Bronx
RF-LUR	R²	0.51	0.52
RF-LUR	RMSE	2.79	2.73
ANN-LUR	R²	0.09	0.10
ANN-LUR	RMSE	3.71	3.69
MLR-LUR	R²	0.10	0.11
MLR-LUR	RMSE	3.15	3.15
Oakland
RF-LUR	R²	0.79	0.79
RF-LUR	RMSE	3.67	3.71
ANN-LUR	R²	0.41	0.40
ANN-LUR	RMSE	6.55	6.70
MLR-LUR	R²	0.42	0.44
MLR-LUR	RMSE	6.34	6.22

3.2.1 Feature importance

Given the higher accuracy of RF models, we explored the top 10 selected feature importance (Fig. 4). The selected features from conventional LUR models strongly align with the lacunarity-optimized models, which confirms the premise that the lacunarity value could capture the maximum range of spatial heterogeneity. In other words, most selected features are captured within the upper bounded buffer sizes limited by lacunarity value (Figure S6: Feature importance for all selected features in lacunarity-optimized and conventional LUR models in Bronx and Oakland). For example, the flattened value of lacunarity for the building is 34 (refers to 1020m of upper limit) and for roads is 18 (refers to 540m). The final selected features in the conventional LUR model in the Bronx for buildings and roads are all less than 1020m (e.g., LPI_Buildings_300, Largest Patch Index of buildings calculated within 300m buffer size) and 540m (e.g., LSI_Roads_300, Largest Shape Index of roads calculated within 300m buffer sizes). In Oakland, the lacunarity value for the street trees is 45 (refers to 450m). Therefore, the maximum buffer distances in conventional LUR models for selected SII (Simpson Index) and SHI (Shannon Index) are 500m (e.g., SII_100, SII_500, and SHI_500), which is close to 450m. Those findings are also consistent with roads, water, and green space, suggesting that the lacunarity value could be an essential indicator to extract the upper limit of the spatial heterogeneity of predictors and facilitate reducing the computation burden.

For selected predictors, the aggregation (e.g. AI, Aggregation Index) and dominance (e.g. PD, Patch Density) of water and green space at large buffer distances (e.g., 1000m, 3000m, and 5000m) account for a higher level of feature importance than other features in both models in the Bronx, which could be attributed to the fact that urban green space has the potential to mitigate air pollution in high-density urban areas (Neft et al., 2016). In Oakland, the AR (Aspect Ratio), SII, and SHI show higher feature importance than the green space and water, suggesting the essential synthetic effects between urban form and tree diversity on the distribution of NO₂ concentration. Moreover, the fragmentation degree of buildings (e.g., LPI_Building_300 in the Bronx) and building-street organization (e.g., AR_100 in Oakland) are also associated with heat transfer, air circulation, and consequently the diffusion of pollutants in urban environments (Liu et al., 2018a).

(Each term has a combined meaning. For example, AI_Water_3000 means the aggregation index for water calculated within the 3,000m buffer size, AR_100 indicates the aspect ratio calculated within the 100m buffer size, and so forth)

3.2.2 SHAP (SHapley Additive exPlanations) value

To further interpret the potentially influential directions of the selected features on the prediction results, we plotted the SHAP value for the selected top 10 features (Fig. 5). In the Bronx, the aggregation and dominance of water (e.g., AI_Water_3000 and LSI_Water_5000) and the roads (e.g., LSI_Roads_300) in both lacunarity-optimized and conventional LUR models show relatively positive associations with NO₂ concentration (higher values distributed on the right side of the X-axis). In contrast, PD_Water_5000 in the lacunarity-optimized model shows negative associations, suggesting that urban watershed distributions at large buffer distances may have non-monotonic effects on NO₂ concentration. Likewise, the dominance of green space (e.g., LPI_Greenspace_5000 in the lacunarity-optimized model and LSI_Greenspace_1000 in the conventional LUR model) is also negatively related to NO₂ concentration. In Oakland, SII_450 (Lacunarity-optimized model) SII_500 (conventional LUR model), LSI_Water_1000, and Feel_temperature (both models) show positive correlations, whereas SHI_450 (Lacunarity-optimized model) and SHI_500 (conventional LUR model) displays a negative association, which reconfirms that the Lacunarity-optimized model could capture the same crucial information as conventional LUR models. In sum, the increase in the dominance of specific tree species (larger SII) is associated with a higher level of NO₂ concentration, while planting more diverse tree species to increase the richness (larger SHI) relates to better air quality. This manifests the multifaced effects of tree diversity on NO₂ concentration as in previous studies. For example, on the one hand, trees could facilitate the widespread deposition of various gases and particles and influence microclimate and air turbulence to improve air quality (Neft et al., 2016; Berardi et al., 2014; Nowak et al., 2013). On the other hand, trees could produce wind-dispersed pollen and emit a range of gaseous substances (e.g., NOx) that take part in photochemical reactions to deteriorating air quality (Grote et al., 2016). The SHAP value for all selected features can be seen in Figure S7.

3.2.3 PDP (Partial Dependence Plot) distribution

As tree diversity shows a non-monatomic correlation on NO₂ concentration, it is necessary to investigate how and to what extent tree diversity relates to NO₂ concentration on the city scale. Therefore, we plotted the PDP by selecting tree diversity indices with higher feature importance (SII_100 and SII_310 in Bronx and SII_450 and SHI_450 in Oakland based on lacunarity-optimized models) to explore the marginal effects between tree diversity and NO₂ concentration. According to Fig. 6, in the Bronx, the NO₂ concentration is relatively lower when SII_100 is lower than 0.80 and the SII_310 ranges between 0.82 to 0.92 (purple clusters). In addition, the increase of SII_100 associates with lower NO₂ concentration, whereas the SII_310 has complicated impacts. Too high (> 0.92) or too low (< 0.79) of SII_310 would both relate to the higher NO₂ concentration (> 9.91 ppb). In Oakland, the NO₂ concentration is relatively lower when SHI_450 is higher than 0.70 and SII_450 is lower than 0.95. The impacts of SHI_450 and SII_450 themselves on NO₂ concentration are almost monotonic negative (values change from high to low from a bottom-up perspective) and positive (values change from low to high from a left-right perspective), respectively, which accords with the findings in Fig. 5. Therefore, the level of tree diversity (both SII and SHI) needs to be controlled within limited ranges to improve the air quality effectively instead of going to extremes by planting more diverse tree species or simply one specific species on the city scale. We also find that SVF (Sky View Factor), AR (Aspect Ratio), and SII have evident synthetic effects on the NO₂ concentration in the Bronx according to the PDP between the tree diversity and 3-D urban form in Figure S8.

(Values in PDP figures indicate the level of NO₂ concentrations. Low to high NO₂ concentrations indicate the gradual change of colors from purple to green and yellow. To better explore the marginal effects between tree diversity indices on NO₂ concentration, we need to single out the tree diversity indices with higher feature importance. According to Figure S6, in the Bronx, only SII_100 (1.72%) has been selected in conventional LUR models. However, SII_100 and SII_310 account for 5.35% of feature importance in total in the lacunarity-optimized model. Therefore, we choose SII_100 and SII_310 in the Bronx through the lacunarity-optimized model to plot PDP. In Oakland, SII_450 and SHI_450 account for 10.17% and 10.58% in total for the conventional and lacunarity-optimized models, respectively. Therefore, we select SII_450 and SHI_450 in the lacunarity-optimized model to plot PDP in Oakland. The DBH is not considered here because no relevant DBH parameters have been selected in Oakland for both models.)

3.2.4 Prediction comparison

To compare the prediction results between lacunarity-optimized and conventional LUR models, we created the fishnet with the grid size of 510m (17\(\times\)30m) and 540m (18\(\times\)30m), the lacunarity value of “Overall landuse” in Fig. 2, in the Bronx and Oakland, respectively (Fig. 7). The meteorological factors were input as the average value during the collection period for Bronx (September 10th to December 17th, 2021) and Oakland (May 28th, 2015 to December 21st, 2017), respectively. In the Bronx, the spatial distribution of NO₂ concentration for the two models is similar with higher values located in the southern and eastern parts. Besides, the prediction results from conventional LUR models show an overall higher NO₂ concentration (mean value of 10.37 ppb) and larger variance (standard deviation of 0.34 ppb) than the lacunarity-optimized (mean value of 10.15 ppb and standard deviation of 0.14 ppb) model. In Oakland, higher NO₂ concentration is almost distributed in the southern part for the two models. Besides, the lacunarity-optimized model has a higher average value (22.10 ppb vs. 21.84 ppb) and lower variation (3.23 ppb vs. 2.60 ppb) than the conventional LUR model.

In addition, the locations with higher NO₂ concentrations have larger relative errors ( > = 50%) between the two models in both Bronx and Oakland, indicating that the model may not estimate the higher level of NO₂ concentration reliably partly due to the transported pollutants from nearby emission sources. Therefore, future models could incorporate more accurate and detailed NOx emission inventories to reduce the bias in these hotspot regions. Besides, the absolute prediction error is lower in the Bronx (from − 0.50 to 1.03 ppb) than that in Oakland (-8.99 to 12.71 ppb), which could be attributed to the fact that urban form-air quality relation also is modulated by other factors, such as seasonality (Liu et al., 2018b), city sizes (Fan et al., 2017), and development stages (Liang and Gong, 2020) among different cities.

In this study, we leveraged opportunistic mobile monitoring to examine intra-urban NO₂ concentration on hyperlocal scales and incorporated both urban forest and 3-D urban form factors into the model construction. We investigated the associations between a wide range of tree diversity, urban form, and NO₂ concentration distribution and proposed the objective lacunarity-optimized method to explore the spatial heterogeneity of predictors for feature extraction, as they do not rely on predefined landscape units or patches and thus can effectively deal with the complexity of urban context. Our lacunarity approach can efficiently determine land use feature-specific upper bounds of buffer sizes, therefore, saving computational resources for extracting information in large areas while capturing the same crucial information and keeping a similar level of accuracy as the conventional way simultaneously.

Even though deciduous trees are surrounded by a higher level of NO₂ concentration than semi-deciduous and evergreen trees in both Bronx and Oakland, the influential direction of tree diversity indicates the diversified tree species are associated with an overall lower NO₂ concentration. Notwithstanding, the level of tree diversity should not be unlimited high. According to the non-monotonic and marginal effects of tree diversity on NO₂ distribution, the diversity of tree species should be controlled in limited ranges to effectively control air quality. Therefore, it is promising to plant more semi-deciduous or evergreen trees to reduce air pollution regionally, but also necessary to strike a balance between the diversity and dominance of tree species to improve air quality on the city scale. The findings could help improve the urban air quality and urban ecological health for a better realization of the One Health initiative.

We acknowledge several limitations in our study. First of all, the current study reflects the association between tree diversity and NO₂ concentration, but the causality could be further analyzed through Rubin or Pearl’s structural casual models with modulators, such as traffic intensity and collection periods. Therefore, future studies could classify the mobile data by the TAZ (Traffic Analysis Zone) or the measurement time (e.g., weekdays and weekends or day-night division) (Montazeri et al., 2021). Secondly, the upper bounds of buffer distances obtained from lacunarity analysis could be further validated by analyzing the influential range of emission sources through deterministic models, such as Gaussian plume dispersion (Gibson et al., 2013). Thirdly, to represent air quality comprehensively, more types of air pollutants (e.g., PM_2.5, PM₁₀, and O₃) need to be added to represent the urban emission inventories. In addition, other predictors, such as FAI (Frontal Area Index) (Shi et al., 2016), LAI (Leaf Area Index) (Nowak et al., 2013), tree porosity (Tong et al., 2016), CVF (Crown Volume Fraction) (Gromke and Blocken, 2015), tree spacing (Ng and Chau, 2012), and phenology should be incorporated to better reflect the effects of tree diversity on air pollution distribution. Fourthly, more techniques, such as the semi-variogram (Zhou and Lin, 2019), can be experimented with to address the spatial dependence of mobile monitoring observation points to reduce the bias in regression analysis. Finally, during model construction, we applied a series of fixed buffer distances from 100m to 5000m. However, previous studies found that larger particles were more sensitive to wind factors (Tian et al., 2022). Thus, it deserves future research endeavors to create adaptive buffer distances for more fine-tuned models to explore the synthetic effects of local meteorology and land uses on NO₂ distributions.

Acknowledgment

The authors thank all other members of the MIT Senseable City Lab Consortium (FAE Technology, MipMap, Samoo Architects & Engineers, GoAigua, DAR Group, Ordinance Survey, RATP, Anas S.p.A., ENEL Foundation, Università di Pisa, KTH (Sweden), ITB (Indonesia), UTEC (Peru), Politecnico di Torino, SMART (Singapore), AMS Institute, and the cities of Laval, Curitiba, Stockholm, Amsterdam, Helsingborg) for funding this research, and special thanks to Joshua Apte and his group's work to provide NO₂ data for Oakland.

The authors thank the New York City Mayor’s Office of the Chief Technology Officer, the Department of Health and Mental Hygiene, the Department of Citywide Administrative Services, as well as the EDF - Environmental Defense Fund, for providing the sensing fleet and providing feedback on the research work.

Data availability

The data is available free of charge on request.

Competing interests

All authors declare no conflicts of interests.

Apte JS, Messier KP, Gani S, et al. (2017) High-resolution air pollution mapping with Google street view cars: exploiting big data. Environmental science & technology 51(12): 6999–7008.
Bechle MJ, Millet DB and Marshall JD (2011) Effects of income and urban form on urban NO2: Global evidence from satellites. Environmental science & technology 45(11): 4914–4919.
Berardi U, GhaffarianHoseini A and GhaffarianHoseini A (2014) State-of-the-art analysis of the environmental benefits of green roofs. Applied energy 115: 411–428.
Borrego C, Martins H, Tchepel O, et al. (2006) How urban structure can affect city sustainability from an air quality perspective. Environmental modelling & software 21(4): 461–467.
Brantley H, Hagler G, Kimbrough E, et al. (2014) Mobile air monitoring data-processing strategies and effects on spatial air pollution trends. Atmospheric measurement techniques 7(7): 2169–2183.
Brownlee J (2019) How to choose a feature selection method for machine learning. Machine Learning Mastery 10.
Buccolieri R, Gromke C, Di Sabatino S, et al. (2009) Aerodynamic effects of trees on pollutant concentration in street canyons. Science of the Total Environment 407(19): 5247–5256.
Buyantuyev A, Wu J and Gries C (2010) Multiscale analysis of the urbanization pattern of the Phoenix metropolitan landscape of USA: time, space and thematic resolution. Landscape and Urban Planning 94(3–4): 206–217.
CDC (2022) One Health. Available at: https://www.cdc.gov/onehealth/index.html.
Cho H-S and Choi M (2014) Effects of compact urban development on air pollution: Empirical evidence from Korea. Sustainability 6(9): 5968–5982.
Corburn J, Osleeb J and Porter M (2006) Urban asthma and the neighbourhood environment in New York City. Health & place 12(2): 167–179.
Cummings LE, Stewart JD, Reist R, et al. (2021) Mobile monitoring of air pollution reveals spatial and temporal variation in an urban landscape. Frontiers in Built Environment 7: 648620.
Dark SJ and Bram D (2007) The modifiable areal unit problem (MAUP) in physical geography. Progress in Physical Geography 31(5): 471–479.
DeSouza P, Anjomshoaa A, Duarte F, et al. (2020) Air quality monitoring using mobile low-cost sensors mounted on trash-trucks: Methods development and lessons learned. Sustainable Cities and Society 60: 102239.
Do TH, Tsiligianni E, Qin X, et al. (2020) Graph-Deep-Learning-Based Inference of Fine-Grained Air Quality from Mobile IoT Sensors. IEEE Internet of Things Journal 7(9): 8943–8955.
Edussuriya P, Chan A and Ye A (2011) Urban morphology and air quality in dense residential environments in Hong Kong. Part I: District-level analysis. Atmospheric Environment 45(27): 4789–4803.
Eeftens M, Beekhuizen J, Beelen R, et al. (2013) Quantifying urban street configuration for improvements in air pollution models. Atmospheric Environment 72: 1–9.
Eilstein D (2010) Exposition prolongée à la pollution atmosphérique et mortalité par pathologies respiratoires. Revue Française d'Allergologie 50(2): 51–61.
Fan C, Tian L and Li J (2017) Research Progress of Impacts of Urban Form on Air Quality. Urban Development. 12.
Fang Y and Zhao L (2022) Assessing the environmental benefits of urban ventilation corridors: A case study in Hefei, China. Building and Environment. 108810.
Foundation AD (2022) Northern Red Oak
Quercus rubra. Available at: https://www.arborday.org/trees/treeguide/TreeDetail.cfm?Itemid=2451.
Friedman JH (2001) Greedy function approximation: a gradient boosting machine. Annals of statistics. 1189–1232.
Galle NJ, Halpern D, Nitoslawski S, et al. (2021) Mapping the diversity of street tree inventories across eight cities internationally using open data. Urban forestry & urban greening 61: 127099.
Gani S, Chambliss SE, Messier KP, et al. (2021) Spatiotemporal profiles of ultrafine particles differ from other traffic-related air pollutants: lessons from long-term measurements at fixed sites and mobile monitoring.
Gefen Y, Meir Y, Mandelbrot BB, et al. (1983) Geometric implementation of hypercubic lattices with noninteger dimensionality by use of low lacunarity fractal lattices. Physical Review Letters 50(3): 145.
Gibson MD, Kundu S and Satish M (2013) Dispersion model evaluation of PM2. 5, NOx and SO2 from point and major line sources in Nova Scotia, Canada using AERMOD Gaussian plume air dispersion model. Atmospheric Pollution Research 4(2): 157–167.
Goin DE, Sudat S, Riddell C, et al. (2021) Hyperlocalized Measures of Air Pollution and Preeclampsia in Oakland, California. Environmental science & technology 55(21): 14710–14719.
Gorai AK, Tuluri F and Tchounwou PB (2014) A GIS based approach for assessing the association between air pollution and asthma in New York State, USA. International journal of environmental research and public health 11(5): 4845–4869.
Gorelick R (2006) Combining richness and abundance into a single diversity index using matrix analogues of Shannon's and Simpson's indices. Ecography 29(4): 525–530.
Grimmond C and Oke TR (1999) Aerodynamic properties of urban areas derived from analysis of surface form. Journal of Applied Meteorology and Climatology 38(9): 1262–1292.
Gromke C and Blocken B (2015) Influence of avenue-trees on air quality at the urban neighborhood scale. Part II: Traffic pollutant concentrations at pedestrian level. Environmental pollution 196: 176–184.
Grote R, Samson R, Alonso R, et al. (2016) Functional traits of urban trees: air pollution mitigation potential. Frontiers in Ecology and the Environment 14(10): 543–550.
Haley VB, Talbot TO and Felton HD (2009) Surveillance of the short-term impact of fine particle air pollution on cardiovascular disease hospitalizations in New York State. Environmental Health 8(1): 42.
Hang J, Li Y, Sandberg M, et al. (2012) The influence of building height variability on pollutant dispersion and pedestrian ventilation in idealized high-rise urban areas. Building and Environment 56: 346–360.
Hoechstetter S, Walz U and Thinh NX (2011) Adapting lacunarity techniques for gradient-based analyses of landscape surfaces. Ecological Complexity 8(3): 229–238.
Hoek G (2017) Methods for assessing long-term exposures to outdoor air pollutants. Current environmental health reports 4(4): 450–462.
Hofman J, Samson R, Joosen S, et al. (2018) Cyclist exposure to black carbon, ultrafine particles and heavy metals: An experimental study along two commuting routes near Antwerp, Belgium. Environmental research 164: 530–538.
Kerckhoffs J, Khan J, Hoek G, et al. (2022) Mixed-Effects Modeling Framework for Amsterdam and Copenhagen for Outdoor NO2 Concentrations Using Measurements Sampled with Google Street View Cars. Environmental science & technology.
Kheirbek I, Haney J, Douglas S, et al. (2016) The contribution of motor vehicle emissions to ambient fine particulate matter public health impacts in New York City: a health burden assessment. Environmental Health 15(1): 89.
Labib S, Lindley S and Huck JJ (2020) Scale effects in remotely sensed greenspace metrics and how to mitigate them for environmental health exposure assessment. Computers, Environment and Urban Systems 82: 101501.
Li F, Zhou T and Lan F (2021) Relationships between urban form and air quality at different spatial scales: A case study from northern China. Ecological Indicators 121: 107029.
Liang L and Gong P (2020) Urban and air pollution: a multi-city study of long-term effects of urban landscape patterns on air quality trends. Scientific Reports 10(1): 1–13.
Liu X, Hu G, Chen Y, et al. (2018a) High-resolution multi-temporal mapping of global urban land using Landsat images based on the Google Earth Engine Platform. Remote sensing of environment 209: 227–239.
Liu Y, Fang X, Cheng C, et al. (2016) Research and application of city ventilation assessments based on satellite data and GIS technology: a case study of the Yanqi Lake Eco-city in Huairou District, Beijing. Meteorological Applications 23(2): 320–327.
Liu Y, Wu J and Yu D (2017) Characterizing spatiotemporal patterns of air pollution in China: A multiscale landscape approach. Ecological Indicators 76: 344–356.
Liu Y, Wu J, Yu D, et al. (2018b) The relationship between urban form and air pollution depends on seasonality and city size. Environmental Science and Pollution Research 25(16): 15554–15567.
Liu Y, Wu J, Yu D, et al. (2018c) The relationship between urban form and air pollution depends on seasonality and city size. Environmental Science and Pollution Research. DOI: http://doi.org/10.1007/s11356-018-1743-6. 1–14.
Lundberg SM and Lee S-I (2017) A unified approach to interpreting model predictions. Proceedings of the 31st international conference on neural information processing systems. 4768–4777.
Maciejczyk PB, Offenberg JH, Clemente J, et al. (2004) Ambient pollutant concentrations measured by a mobile laboratory in South Bronx, NY. Atmospheric Environment 38(31): 5283–5294.
Mandelbrot BB and Mandelbrot BB (1982) The fractal geometry of nature. WH freeman New York.
Manes F, Silli V, Salvatori E, et al. (2014) Urban ecosystem services: Tree diversity and stability of PM10 removal in the metropolitan area of Rome. Annali di Botanica 4: 19–26.
Martins H (2012) Urban compaction or dispersion? An air quality modelling study. Atmospheric Environment 54: 60–72.
McGarigal K (2015) FRAGSTATS help.
Meng X, Chen L, Cai J, et al. (2015) A land use regression model for estimating the NO2 concentration in Shanghai, China. Environmental research 137: 308–315.
Messier KP, Chambliss SE, Gani S, et al. (2018) Mapping air pollution with Google Street View cars: Efficient approaches with mobile monitoring and land use regression. Environmental science & technology 52(21): 12563–12572.
Montazeri A, Lilienthal AJ and Albertson JD (2021) A spatial land use clustering framework for investigating the role of land use in mediating the effect of meteorology on urban air quality. Atmospheric Environment: X 12: 100126.
Mora S, Anjomshoaa A, Benson T, et al. (2019) Towards large-scale drive-by sensing with Multi-Purpose City scanner nodes. 2019 IEEE 5th World Forum on Internet of Things (WF-IoT). IEEE, 743–748.
Myint SW and Lam N (2005) A study of lacunarity-based texture analysis approaches to improve urban image classification. Computers, Environment and Urban Systems 29(5): 501–523.
Naughton O, Donnelly A, Nolan P, et al. (2018) A land use regression model for explaining spatial variation in air pollution levels using a wind sector based approach. Science of the Total Environment 630: 1324–1334.
Neft I, Scungio M, Culver N, et al. (2016) Simulations of aerosol filtration by vegetation: Validation of existing models with available lab data and application to near-roadway scenario. Aerosol Science and Technology 50(9): 937–946.
Ng W-Y and Chau C-K (2012) Evaluating the role of vegetation on the ventilation performance in isolated deep street canyons. International journal of environment and pollution 50(1–4): 98–110.
Nowak DJ, Crane DE and Stevens JC (2006) Air pollution removal by urban trees and shrubs in the United States. Urban forestry & urban greening 4(3–4): 115–123.
Nowak DJ, Hirabayashi S, Bodine A, et al. (2013) Modeled PM2. 5 removal by trees in ten US cities and associated health effects. Environmental pollution 178: 395–402.
Peng Y, Gao Z, Buccolieri R, et al. (2021) Urban ventilation of typical residential streets and impact of building form variation. Sustainable Cities and Society 67: 102735.
Perrier A (2015) Feature Importance in Random Forests. Available at: https://alexisperrier.com/datascience/2015/08/27/feature-importance-random-forests-gini-accuracy.html.
Plotnick RE, Gardner RH, Hargrove WW, et al. (1996) Lacunarity analysis: a general technique for the analysis of spatial patterns. Physical review E 53(5): 5461.
Plotnick RE, Gardner RH and O'Neill RV (1993) Lacunarity indices as measures of landscape texture. Landscape Ecology 8(3): 201–211.
Pugh TA, MacKenzie AR, Whyatt JD, et al. (2012) Effectiveness of green infrastructure for improvement of air quality in urban street canyons. Environmental science & technology 46(14): 7692–7699.
Ren X, Mi Z and Georgopoulos PG (2020) Comparison of Machine Learning and Land Use Regression for fine scale spatiotemporal estimation of ambient air pollution: Modeling ozone concentrations across the contiguous United States. Environment International 142: 105827.
Ricotta C and Szeidl L (2006) Towards a unifying approach to diversity measures: bridging the gap between the Shannon entropy and Rao's quadratic index. Theoretical population biology 70(3): 237–243.
Roces-Diaz JV, Diaz-Varela RA, Alvarez-Alvarez P, et al. (2015) A multiscale analysis of ecosystem services supply in the NW Iberian Peninsula from a functional perspective. Ecological Indicators 50: 24–34.
Rodríguez MC, Dupont-Courtade L and Oueslati W (2016) Air pollution and urban structure linkages: Evidence from European cities. Renewable and Sustainable Energy Reviews 53: 1–9.
Roeland S, Moretti M, Amorim JH, et al. (2019) Towards an integrative approach to evaluate the environmental ecosystem services provided by urban forest. Journal of Forestry Research 30(6): 1981–1996.
Saha PK, Li HZ, Apte JS, et al. (2019) Urban ultrafine particle exposure assessment with land-use regression: influence of sampling strategy. Environmental science & technology 53(13): 7326–7336.
Saunders S, Dade E and Van Niel K (2011) An Urban Forest Effects (UFORE) model study of the integrated effects of vegetation on local air pollution in the Western Suburbs of Perth, WA. 19th International congress on modelling and simulation (MODSIM2011). vol.-, Modelling and Simulation Society of Australia and New Zealand Inc., pp.-, 19th International congress on modelling and simulation (MODSIM2011), Perth, Australia.
Shannon CE (1948) A mathematical theory of communication. The Bell system technical journal 27(3): 379–423.
Shen J, Gao Z, Ding W, et al. (2017) An investigation on the effect of street morphology to ambient air quality using six real-world cases. Atmospheric Environment 164: 85–101.
Shi Y, Lau KK-L and Ng E (2016) Developing street-level PM2. 5 and PM10 land use regression models in high-density Hong Kong with urban morphological factors. Environmental science & technology 50(15): 8178–8187.
Shi Y, Lau KK-L and Ng E (2017) Incorporating wind availability into land use regression modelling of air quality in mountainous high-density urban environment. Environmental research 157: 17–29.
Sicard P, Agathokleous E, Araminiene V, et al. (2018) Should we see urban trees as effective solutions to reduce increasing ozone levels in cities? Environmental pollution 243: 163–176.
Simpson EH (1949) Measurement of diversity. nature 163(4148): 688–688.
Tang R, Blangiardo M and Gulliver J (2013) Using building heights and street configuration to enhance intraurban PM10, NOx, and NO2 Land use regression models. Environmental science & technology 47(20): 11643–11650.
Tang U and Wang Z (2007) Influences of urban forms on traffic-induced noise and air pollution: Results from a modelling system. Environmental modelling & software 22(12): 1750–1764.
Tian Y, deSouza P, Mora S, et al. (2022) Evaluating the Meteorological Effects on the Urban Form–Air Quality Relationship Using Mobile Monitoring. Environmental science & technology.
Tian Y and Yao X (2022) Urban form, traffic volume, and air quality: A spatiotemporal stratified approach. Environment and Planning B: Urban Analytics and City Science 49(1): 92–113.
Tian Y, Yao XA, Mu L, et al. (2020) Integrating meteorological factors for better understanding of the urban form-air quality relationship. Landscape Ecology 35(10): 2357–2373.
Tiwari A and Kumar P (2020) Integrated dispersion-deposition modelling for air pollutant reduction via green infrastructure at an urban scale. Science of the Total Environment 723: 138078.
Tong Z, Baldauf RW, Isakov V, et al. (2016) Roadside vegetation barrier designs to mitigate near-road air pollution impacts. Science of the Total Environment 541: 920–927.
Van den Bossche J, Peters J, Verwaeren J, et al. (2015) Mobile monitoring for mapping spatial variation in urban air quality: Development and validation of a methodology based on an extensive dataset. Atmospheric Environment 105: 148–161.
Yang J, Shi B, Shi Y, et al. (2020) Air pollution dispersal in high density urban areas: Research on the triadic relation of wind, air pollution, and urban form. Sustainable Cities and Society 54: 101941.
Yuan M, Song Y, Huang Y, et al. (2019) Exploring the association between the built environment and remotely sensed PM2. 5 concentrations in urban areas. Journal of cleaner production 220: 1014–1023.
Zahl S (1977) Jackknifing an index of diversity. Ecology 58(4): 907–913.
Zakšek K, Oštir K and Kokalj Ž (2011) Sky-view factor as a relief visualization technique. Remote sensing 3(2): 398–415.
Zhang A, Xia C and Li W (2022) Exploring the effects of 3D urban form on urban air quality: Evidence from fifteen megacities in China. Sustainable Cities and Society 78: 103649.
Zhong J, Cai X-M and Bloss WJ (2016) Coupling dynamics and chemistry in the air pollution modelling of street canyons: A review. Environmental pollution 214: 690–704.
Zhou S and Lin R (2019) Spatial-temporal heterogeneity of air pollution: The relationship between built environment and on-road PM2. 5 at micro scale. Transportation Research Part D: Transport and Environment 76: 305–322.

No competing interests reported.

SupportingInformation.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

Incorporating tree diversity for a better understanding of urban form-air quality relationships through mobile monitoring

Status:

Version 1

Abstract

Figures

1. Introduction

2. Methods

2.1 Research areas

2.2 Research design

2.2.1 NO₂ preprocessing

2.2.2 Tree diversity

2.2.3 2-D urban form metrics

2.2.4 3-D urban form metrics

2.2.5 Lacunarity

2.2.6 Model construction

3. Results And Discussions

3.1 Descriptive analysis

3.1.1 Lacunarity results

3.1.2 Tree diversity and NO₂ distribution

3.2 Model performance

3.2.1 Feature importance

3.2.2 SHAP (SHapley Additive exPlanations) value

3.2.3 PDP (Partial Dependence Plot) distribution

3.2.4 Prediction comparison

4. Conclusions

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1

Incorporating tree diversity for a better understanding of urban form-air quality relationships through mobile monitoring

Status:

Version 1

Abstract

Figures

1. Introduction

2. Methods

2.1 Research areas

2.2 Research design

2.2.1 NO2 preprocessing

2.2.2 Tree diversity

2.2.3 2-D urban form metrics

2.2.4 3-D urban form metrics

2.2.5 Lacunarity

2.2.6 Model construction

3. Results And Discussions

3.1 Descriptive analysis

3.1.1 Lacunarity results

3.1.2 Tree diversity and NO2 distribution

3.2 Model performance

3.2.1 Feature importance

3.2.2 SHAP (SHapley Additive exPlanations) value

3.2.3 PDP (Partial Dependence Plot) distribution

3.2.4 Prediction comparison

4. Conclusions

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1

2.2.1 NO₂ preprocessing

3.1.2 Tree diversity and NO₂ distribution