Abstract
Background Recent literature has highlighted the importance of visual accessibility to nature to reduce stress, anxiety, or depression amongst others. However, green visual accessibility is yet rarely considered in urban policy implementations. Reasons behind this are manifold, and include the challenges associated with the measurability of green views which require data-intensive pedestrian view computations, and assessment methods are yet to be agreed upon.
Methods Two methods, Street View Images (SVI) and semantic classification, and geospatial viewshed analysis, were used to compute street level tree views. All street views contained within 2 municipalities from the Brussels Capital Region (BCR) have been studied. Using the SVI method, 15 green view indicators have been proposed. Using the viewshed analysis, the tree view area ratio (TVar) from each SVI geo-location has been computed. The independence between the indicators was evaluated, and using a random forest model, the principal SVI indicators to describe the TVar have been studied.
Results The variability explained by the random forest model was approximately 60% to 70%. The SVI indicators related to the horizontality of green infrastructure and tree canopy explained most of TVar. The results also reveal the tree canopy differences between both municipalities.
Conclusions SVI tree view indicators provide acceptable predictions of the TVar which could be particularly useful for municipalities with no access to detailed geospatial data. The 30% to 40% of the unexplained variability, could be related to errors derived from the tree canopy geospatial layer, differences in the data collection dates, or geolocation errors of the SVIs.
Introduction
The benefits of urban green infrastructure accessibility for citizens’ physical and mental health has been widely reported in prior literature (Hartig 2008; Hartig and Kahn 2016). The role visual accessibility to green infrastructure can play for urban stress, anxiety, or depression mitigation amongst others has been acknowledged (Huang et al. 2021; Konijnendijk 2023). However, while physical accessibility to green areas is starting to be assessed for urban policy implementations (van den Bosch and Ode Sang 2017), visual accessibility is yet rarely considered. In recent years, strategic tree planting investments have been allocated to enhance the tree accessibility in several cities. For example, the city of Liege, Belgium, put forward a 10-year tree planting plan named Plan Canopee, with the ambition to reach 24,000 trees planted (Ville de Liege 2007). Similarly, between 2007 and 2017, the Million Trees initiative in New York City successfully planted one million trees by promoting citizens’ engagement to enhance the urban forest and reduce tree access inequality (Million Trees NYC 2017). Likewise, in Paris, a plan to plant 170,000 trees by 2026 was recently set (Masterson 2022). In 2018, the city of Frederiksberg, Denmark, introduced an ambitious policy that declared all inhabitants should be able to view at least one tree from their household, which builds on physical and visual tree accessibility principles (Frederiksberg Kommune 2018). However, urban tree planting strategies to account for tree view accessibility are still very limited. New approaches have been formulated, namely the 3-30-300 rule for urban forestry planning, which recommends a minimum of 3 trees visible from each building, no less than 30% tree canopy in every neighborhood, and a maximum distance of 300 m to the nearest green public space from every household (Konijnendijk 2023). However, measuring green views requires extensive participant surveys or data-intensive computations for which the assessment methods are yet to be agreed upon (Hu et al. 2023; Tabatabaie et al. 2023).
Tree view accessibility assessments utilizing participant questionnaires and surveys have proved to be very time consuming and resource intensive (de Vries et al. 2013; Triebner et al. 2019). On the other hand, given the increasing availability of Street View Imagery (SVI), various publications have made use of street level panoramic and fisheye photos to compute metrics such as the Green View Index (GVI) or Green Area Index (GAI) amongst others (Li et al. 2015; Ye et al. 2019). The possibilities to develop tools and strategies for geospatial tree viewshed computations are also starting to be discussed (Cimburova and Blumentrath 2022). Prior literature has also reported inconsistent results when comparing SVI and geospatial methods. For example, Urban Green Infrastructure (UGI) quantification results obtained through aerial imagery and geospatial sources have been reported to show distinct trends to those obtained through pedestrian level image analysis (Long and Liu 2017; Kumakoshi et al. 2020). On the other hand, a recent publication that focused on the assessment of results obtained from varied SVI sources to address the effect of panoramic or non-panoramic views based on image distortion aspect ratios or fields of view, demonstrated that the results using distinct SVI sources showed comparable results (Biljecki et al. 2023).
This research compares 2 methodologies used to assess the view accessibility of street trees and proposes potential next steps for tree view assessment. Method 1 uses SVIs and semantic classification to define green view indicators using 360° Google Street View (GSV) images. Method 2 uses two-dimensional geospatial urban viewshed analysis. Geospatial visual corridor, viewshed, or isovist analysis approaches have been used by architects and urban planners to develop urban navigation analysis, urban escape route planning, or studies focusing on urban fabric geometry on citizen stress levels amongst others (Sander and Manson 2007; Cai et al. 2023). However, for the study of street tree view accessibility assessments, their application remains limited (Labib et al. 2021; Cimburova and Blumentrath 2022).
Using the proposed methods, this research aimed at the following objectives:
Compare the tree view visibility indicators and assess the variable explanation between indicators;
Propose SVI based indicators to assess tree view accessibility; and,
Identify the advantages and challenges of viewshed analysis to compute tree visibility.
A case study focusing on 2 municipalities (Saint Gilles and Molenbeek) located in the Brussels Capital Region (BCR) is presented. Using the 2 methods, tree visibility was computed and results were compared. Given that access to geospatial data remains limited in many cities while SVIs are becoming increasingly available covering large regions, the potential use of SVI indicators to explain the geospatial tree area viewshed is also assessed.
Materials and methods
Datasets for the Studied Municipalities
Making use of geospatial data readily available for the Brussels Capital Region, the building boundary information, street axes, and tree canopy areas were identified for the 2 municipalities under study. SVIs were also collected using 360° GSV images, and the geotagging of each image was saved and stored as a geospatial point array. In total 5,816 points (i.e., same number as the geo-tagged panoramic images) were studied for Saint Gilles, and 10,682 for Molenbeek. For more information on the data sources, see Table 1.
Method 1: Google Street View Imagery
Through queries from Street View Download 360° (Orlita 2023), all available geo-tagged 360° SVIs were compiled for both municipalities. In average, the SVIs are approximately 10 m apart from each other and located at central positions of the street canyons. Following the image compilation, a semantic classification was developed with the Python library PixelLib—implemented with Deeplabv3+ framework (Olafenwa 2021). The training datasets based on Ade20k have been utilized with architecture based on the model Xception (Zhou et al. 2019). The images were segmented into 14 categories: sky, road, building, car, tree, sidewalk, streetlight, sign, grass, bike, plant, person, path, and fence. From the segmented images and using the pixels comprising the grass, plant, and tree categories, the Green View Index (GVI) was computed (see Equation 1) as the division of all green pixels by all image pixels, for all images within each square. The obtained ratios include total UGI, Green ratio (Gr) which is the GVI, and those distinguished by types for the Tree ratio (Tr) and Grass ratio (GRr)(Figure 1). 1
Using the segmented images, indicators that describe the location of the UGI relative to the image boundaries were derived. These indicators include the Green Vertical angle (GVa), the Tree Vertical angle (TVa), and the Grass Vertical angle (GRVa), and provide insights about the relative position and scale of the UGI. Aside from the total metrics, the Green Vertical top angle (GVta), the Tree Vertical top angle (TVta), the Grass Vertical top angle (GRVta), the Green Vertical bottom angle (GVba), the Tree Vertical bottom angle (TVba), and the Grass Vertical bottom angle (GRVba) have been computed. These angles enable capturing the relative size (e.g., crown width) and position of the UGI. The smaller the GVta and GVba, for example, the closer the UGI is to the user, and depending on the Green Horizontal angle (GHa) and GVa ratio, the type of UGI can be derived (see Figures 1 and 2 and Table 2).
Method 2: Viewshed Analysis
Two-dimensional viewsheds were computed using Esri’s ArcGIS Pro v2.7.3 Spatial Analyst package (Achilleos and Tsouchlaraki 2004; ArcGIS 2022) (Table 1). The tree canopy layer was made available by Brussels Environment. Details on the preparation of this dataset can be found in prior literature (van de Voorde 2017; Pelgrims et al. 2021). An iterative workflow to compute the viewshed for all points collected for Saint Gilles and Molenbeek was implemented using the ArcGIS visual programming ModelBuilder geoprocessing tool (Figure 3). The viewshed analysis was run for a spatial resolution of 1 m. Two different raster outputs were created to compute the ratio of visual area in presence of vegetation and without it from which the Tree View area ratio (TVar) was computed (Equation 2). The viewshed tool parameters are by default with a fixed outer radius of 300 m.
For the error assessment of the viewshed computation, from the 5,816 points studied for the case of Saint Gilles, approximately 1% were run for a second and third iteration to understand if the results varied following the differences reported in prior literature (Riggs and Dean 2007; Parent and Lei-Parent 2023). For the studied TVar computations, the differences between iterations remained < 0.11% and thus it was considered negligible. 2
Statistical Methods
The Pearson correlation was computed to assess the linear correlation between the studied SVI indicators and TVar. To find the group of variables that carry comparable information, the Principal Component Analysis (PCA) was studied (Pearson 1901). For dependencies that cannot be explained linearly, the random forest model as per Breiman (2001) was used to assess the nonlinear relationship between the TVar and remaining variables. The R 4.7-1.1 package for randomForest was used (Liaw and Wiener 2002). The random forest model involved randomly picking 5 variables—following the 1/3 selection of the continuous variables—from the 15 SVI indicators for each tree node which was limited to a total number of 500 trees. For validation purposes, the complete data (combing Molenbeek and Saint Gilles) was split (70% for training and 30% for testing), and the percentage of variance explained remained consistently approximately 66% on the validation sets. The assessment of the random forest model results is provided with the explained percentage of variability in TVar.
Results
The summary statistics obtained from the computed 16 indicators and for both municipalities under study—Saint Gilles and Molenbeek separately—are included in Table 3. Overall, the results for Molenbeek show higher mean values (for the case of the TVar, a higher ratio indicates a lower tree area). Amongst others, the representative GVa, THa, and Gr are higher for the case of Molenbeek. This is also the case for TVar obtained using the viewshed analysis.
By studying the density distribution and bar plots for all 16 indicators, the differences between the 2 municipalities become visible. For example, a higher mean THa and smaller TVta or TVba indicates the presence of a higher tree canopy. This is also the case for the lumped grass and tree metrics as per row 3. Similarly, for the TVar obtained following the viewshed analysis, a lower mean indicates a higher presence of trees for Molenbeek. On the other hand, overall a larger variability is observed for Saint Gilles for various indicators such as TVa or TVba which indicates a more uneven distribution of the tree canopy in this municipality. The density plots also show tree concentration peaks for the municipality of Saint Gilles, which is observable for example by studying the indicator THa, while for Molenbeek a more even distribution is observed. This is also the case for the TVar computed using the viewshed analysis (Figures 4 and 5). In the map visualizations of the results obtained following Method 2 viewshed analysis, it is also qualitatively visible that street tree distributions are different between the municipalities (Figure 6).
The PCA results for Saint Gilles show 4 main indicator groups: (i) the variables TVta, TV ba, GVta, and GVba, which are indicators of the height of the tree; (ii) the grass related indicators GRVa, GRHa, or GRr; (iii) those inversely correlated GRVta and GRVba to the vertical component, which highlight that these are independent to the tree metrics; and (iv) Tr, Gr, THa, or GHa, which reflect upon the tree crown width. In the case of Molenbeek, while the 4 indicator groups are also visible, they are more dispersed, particularly those related to the crown width (i.e., Tr or THa) which are more differentiated from the GHa or Gr (Figure 7). The first 2 dimensions covered approximately 75% of the variability in the data for both municipalities.
An initial study was developed to assess whether significant linear correlations were found between the studied SVI method indicators, using the viewshed TVar indicator as the dependent variable. Highest Pearson Correlation Coefficients (PCC) were observed for Gr, Tr, and THa with PCC approximately −0.6 (p < 0.05) (Table 4). However, in the data correlation, nonlinearity was observed (Figure 8), and thus, a nonlinear random forest model (Breiman 2001) was deployed, aiming to explain TVar using the indicators obtained following the SVI method. This model explained 67.32% of the variability in TVar for the case of Molenbeek, and 60.23% for the case of Saint Gilles. The 4 variables that better explained the TVar for both municipalities were the GHa, Gr, THa, and Tr (Figure 8).
Limitations
The limitations identified in the proposed methodologies are to be considered. Method 1 relies on the segmentation of 360° GSV imagery which was captured between the years 2009 to 2021. Along those years, seasonal changes, natural tree growth, tree removal, or trimming may have occurred. While the GSV images are generally located at < 15 m distances, it is also possible that their geo-tagging may also include a geo-location error of approximately 2 m as reported in prior literature, as well as temporal instability which involves an incomplete date-stamp (Curtis et al. 2013; Krylov et al. 2018). Likewise, for the viewshed calculation, the tree canopy shapefile provided by Brussels Environment was used, where the Normalized Difference Vegetation Index (NDVI) was only transformed into a binary variable (presence/absence of vegetation) which may also incur potential errors (van de Voorde 2017; Pelgrims et al. 2021). Furthermore, the tree canopy database dates back to the year 2020, and thus differences with the SVIs may be found given recently planted trees or geometry changes due to tree maintenance. That is, the 30% to 40% of the unexplained variability could be related to errors derived from these limitations which could only be overcome by collecting a more accurate and up-to-date tree data documentation.
Conclusions
Focusing on all streets located at 2 municipalities from the Brussels Capital Region, the pedestrian level tree view accessibility has been studied. Two methods to compute the tree view accessibility have been compared. Using all available SVIs for the chosen municipalities and through semantic classification, the quantification of various indicators for grass, tree, and total green, as well as indicators to address their relative position within the panoramic image were proposed. Two-dimensional geospatial viewshed analysis has been deployed for the same geo-locations SVIs were available for, and the Tree View area ratio (TVar) was computed. To assess whether the SVI green view indicators can explain the viewshed analysis results, an initial linear correlation was studied using TVar as the dependent variable. Significant correlations were observed particularly between the indicators related with the tree crown width such as THa or GHa, with PCC approximately −0.6 (note that a low ratio of TVar denotes a larger tree view area). Nonlinearity was observed in the data, thus subsequently, a random forest model was deployed which explained approximately 70% variability of the TVar for Molenbeek and approximately 60% for Saint Gilles. The remaining unexplained 30% to 40% variability could also be attributable to the limitations of the utilized datasets such as geo-location errors or date incongruencies, as well as to errors due to the remote sensing data processing, which can affect the accuracy of the geospatial tree canopy layer.
Overall, SVI tree view indicators provide acceptable predictions of the pedestrian level tree canopy viewshed, which could be particularly useful for municipalities with no access to detailed geospatial data. In all cases, the computational costs associated with the tree view access analysis won’t be negligible, thus the challenges for their implementation for urban planning practices are yet to be overcome.
Future work will involve a three-dimensional characterization of the tree canopy, using detailed digital elevation models and urban point clouds which are starting to be collected by several municipalities. Working with three-dimensional meshes can enable the computation of the viewshed analysis around a tree from various angles and better describe the tree trunk and crown geometries.
Conflicts of Interest
The authors reported no conflicts of interest.
Acknowledgements
This research was funded by the Louvain4City 2021 Grant from UCLouvain.
- © 2024 International Society of Arboriculture