Accessing Eye-level Greenness Visibility from Open-Source Street View Images: A methodological development and implementation in multi-city and multi-country contexts

The urban natural environment provides numerous bene�ts, including augmenting the aesthetic appeal of urban landscapes and improving mental wellbeing. While diverse methods have been used to evaluate urban greenery, the assessment of eye-level greenness visibility using street-view level images is emerging due to its greater compatibility with human perception. Many existing studies predominantly rely on proprietary street view images provider such as Google Street View (GSV) data; the usage restrictions and lack of alignment with FAIR (Findability, Accessibility, Interoperability, and Reusability) principles present challenges in using proprietary images at scale. Therefore, incorporating Volunteered Street View Imagery (VSVI) platforms, such as Mapillary, is emerging as a promising alternative. In this study, we present a scalable and reproducible methodological framework for utilising Mapillary images for Green View Index (GVI) assessment using image segmentation approach and evaluate the completeness and usefulness of such data in diverse geographical contexts, including seven cities (i.e., Amsterdam, City of Melbourne, Dhaka, Kampala, Mexico City, Seattle, and Tel Aviv). We also evaluate the use of globally available satellite-based vegetation indices (e.g., Normalised Difference Vegetation Index-NDVI) to estimate GVI in locations where street-view images are unavailable. Our approach demonstrates the applicability of Mapillary data for GVI assessments, although revelling considerable disparities in image availability and usability between cities located in developed and developing countries. We also identi�ed that the NDVI could be used effectively to estimate GVI values in locations where direct street-level imagery is limited. Additionally, the analysis reveals notable differences in greenness visibility across cities, particularly in high-density, lower-income cities in Africa and South Asia, compared to low-density, high-income cities in the USA and Europe.


Introduction
Greenspaces play a crucial role in urban environments, offering a wide array of social, environmental, and ecological bene ts (Bain et al., 2012;Gómez-Baggethun & Barton, 2013;Ordóñez et al., 2023;Saw et al., 2015).As the urban population is projected to increase by 60% by 2050 (United Nations, 2018), sustainable development becomes imperative in addressing pressing issues, including climate change mitigation and improving the quality of life for urban residents (Savo et al., 2023).Urban green spaces provide multifunctionality to enhance the environmental quality within cities.They effectively mitigate the heat island effect (Keeley, 2011;Lafortezza et al., 2009), improve air quality by reducing pollutants (Akbari et al., 2001;Jim & Chen, 2008), reduce noise pollution (Attal et al., 2021;Wong et al., 2010), and minimise stormwater runoff (Armson et al., 2013;Berland et al., 2017).Moreover, the visual perception of street greenery is recognised as a vital sensory function that positively in uences individuals' experiences in urban settings (Lu et al., 2018;Wolf, 2005).Urban greenery serves as a protective barrier against visual intrusion (Li, Zhang, Li, Kuzovkina, et al., 2015).Furthermore, the presence of vegetation in urban landscape often leads to higher aesthetic perception of streetscapes (Lindemann-Matthies & Brieger, 2016), promoting physical activities (Zijlema et al., 2020) and fostering social cohesion among residents (Root et al., 2017).From a health perspective, visual contact with greenery also evokes positive emotions, reduces stress, and facilitates the recovery of mental fatigue (Kaplan, 1995;Ulrich, 1984;Wang et al., 2022).Different assessment methods have been employed to measure and evaluate urban greenery, including surveys, interviews, and audits, to gain insights into people's opinions and attitudes towards street-level views of urban greenery (Falfán et al., 2018;Li, Zhang, Li, Ricard, et al., 2015).However, each of these methods has its advantages and limitations.Surveys using questionnaires may be more re ective of people's personal experience on the ground but may be affected by response bias (Downs & Stea, 1977).
Additionally, audits require skilled evaluators and speci c criteria to assess visual aesthetic quality (Ellaway et al., 2005;Tang & Long, 2019).Such approaches may involve physically transporting evaluators to real locations for direct evaluation of environmental attributes, which can be time-consuming, expensive, and subjective to each participant (Gupta et al., 2012;Meitner, 2004;Yao et al., 2012).In contrast to interview approaches, recent advances in computational advances (e.g., Deep learning image analysis) along with high-resolution data from multiple sources such as geospatial data and street view image data allowed computational modelling of eye-level greenery at scales.One example is the geospatial viewshed model using high-resolution light detection and ranging (LiDAR) based data, which has been extensively employed to map and quantify vegetation cover in urban areas, bene ting from their advantages in terms of repeatability, synoptic view, and wide area coverage (Li, Zhang, Li, Ricard, et al., 2015.These methods involve three-dimensional data sets derived from highresolution remotely sensed images and LiDAR, enabling the detection of tree species and the vertical dimensions of urban trees (Alonzo et al., 2014;Edson & Wing, 2011;Shrestha & Wynne, 2012).According to Yu et al. (2016), these approaches offer several advantages, such as the capacity to adjust the relative height of the observer to simulate different vertical horizons and the potential to model arti cial buildings to simulate future landscapes.Nevertheless, the results depend on the original resolution and accuracy of the utilised datasets (Labib et al., 2021), and there is a risk of information loss when compressing point cloud data into a two-dimensional raster surface (Yang et al., 2020).Additionally, high-resolution LiDAR data and digital elevation raster surfaces are mostly unavailable in most cities or countries worldwide, making it di cult to use such models in diverse geographic contexts (Labib et al., 2021;2020).
In contrast to the viewshed based approach, Li et al. (2015) made signi cant advances in this eld by utilising automatically extracted Google Street View (GSV) images and employing image segmentation techniques for automating the greenery calculation process.This breakthrough has led to the emergence of several innovative computational approaches that have demonstrated a high level of agreement with human perception (Aikoh et al., 2023;Suppakittpaisarn et al., 2022;Torkko et al., 2023), offering several bene ts such as reduced research time and workloads, increased accessibility, and the ability to study urban greenery without the need for physical visits to eld sites (Lu et al., 2023;Rangel et al., 2022).
Another widely adopted approach is computational image segmentation, which is categorised into coloursegmentation and semantic-segmentation.Colour-segmentation method classi es images according to colour values of individual pixels, effectively identifying and delineating areas with green vegetation (J.  .On the other hand, semantic-segmentation is a machine learning-based approach that leverages contextual information to understand and segment images (Xia et al., 2021).By assigning semantic levels to individual pixels, this method allows to comprehend the image content, facilitating precise identi cation and mapping of green vegetation (Cai et 2023) investigated different eye-level greenery extraction methods and found the image segmentation approach better aligns with human perception than colour-based segmentation.The colour-based segmentation approach is sensitive to lighting conditions (Batlle et al., 2000;Pietikainen et al., 1996), and they also tend to misclassify green paint as greenery, leading to overestimating green areas (Larkin & Hystad, 2019;Li, Zhang, Li, Ricard, et al., 2015).
Numerous recent studies employing image segmentation for vegetation analysis heavily rely on data from GSV (Cai et  Zheng & Amemiya, 2023).In particular, GSV imposes restrictions on the use of imagery, including limitations on data analysis and extraction (Google, 2018b(Google, , 2020)).As stated by Rundle et al. (2022), the restriction on certain uses of Google Maps is not grounded in copyright law, which would potentially allow researchers to invoke the "Fair" use principles.Rather, these limitations are based on the contractual agreement that users must adhere to when accessing Google Maps (Google, 2018a).The enforceability of such contracts is currently subject to ongoing legal disputes, making it uncertain.Consequently, researchers engaging in this type of research and the journals publishing the resulting papers assume some legal risk as long as the legal status remains unsettled (Rundle et al., 2022;Stringam et al., 2023).In addition, using such data might hinder the adoption of FAIR (Findable, Accessible, Interoperable, and Reusable) data principles in scienti c research (Wilkinson et al., 2016).GSV is often easy to nd and interoperable with other data (e.g., provided by Google).However, they are not widely accessible due to restrictions on the number of free downloads; there is a limit to the quantity of data that can be obtained without incurring charges (Google, 2023;Zheng & Amemiya, 2023).Meaning researchers who require a considerable amount of GSV data may encounter nancial barriers to accessing them.This inaccessibility also results in a lack of reusability of such data by others.As whole GSV data primarily does not meet the FAIR data principles.
To overcome GSV limitations, integrating Volunteered Street View Imagery (VSVI) platforms, such as Mapillary and Open Street Cam, has been developed over the years (Alvarez Leon & Quinn, 2019; X. Yu et al., 2019;Zheng & Amemiya, 2023).Mapillary, in particular, stands out as the largest crowdsourcing-based street view platform, boasting a community of over 20,000 users who contribute street-level photos using GPS-enabled cameras or smartphones (Ma et al., 2020).Notably, Mapillary adopts an open data approach by realising imagery under the Creative Commons Attribution-ShareAlike 4.0 International license (Alvarez Leon & Quinn, 2019).This open data policy enables researchers to utilise the imagery more freely, including for commercial purposes, fostering innovation and collaboration in studying urban greenery and related phenomena (Alvarez Leon & Quinn, 2019).
While Mapillary provides an alternative solution through its open data approach, researchers are aware of its limitations, as these can affect the reliability and accuracy of the data (Antequera et al., 2020;D'Andrimont et al., 2018;Juhász & Hochmair, 2016;Krylov & Dahyot, 2019).Firstly, the quality of the imagery contributed by users on Mapillary may vary signi cantly due to differences in camera properties, including various camera models and settings used by contributors (Antequera et al., 2020).These variations can result in differences in resolution, blurring, restricted eld of view, and reduced visibility (D'Andrimont et al., 2018).Secondly, Mapillary primarily relies on forward-facing cameras mounted on vehicles driven on roads, resulting in linear motion without rotation, limiting the perspectives captured in the imagery and potentially limiting the comprehensive understanding of the surrounding (Antequera et al., 2020;Krylov & Dahyot, 2019).Finally, the level of engagement and participation from users can vary across different regions and communities, leading to disparities in data availability (D'Andrimont et al., 2018; Nielsen, 2006).As a result, researchers relying solely on Mapillary may encounter gaps in the coverage of speci c locations or road segments that can limit the ability to obtain a complete and representative sample of the desired study area (Juhász & Hochmair, 2016).
In contrast to missing street view image in certain locations, satellite-derived top-down measures of greenery, such as the Normalised Difference Vegetation Index (NDVI), is globally available at any place and often widely used in measuring greenness modelling (Martinez & Labib, 2023).It can be argued that for situations where there is insu cient street image data to determine the GVI, an alternative approach is to utilise NDVI to estimate GVI using statistical models.Although NDVI is a top-down measure and may not capture the actual visibility of greenery (Labib et al., 2021), however as NDVI represents available vegetation, thus it can be assumed that a model can learn from the relation between known GVI and NDVI within a certain distance zone to impute GVI for unknown locations.In this regard, Torkko et al. (2023) indicated that mean NDVI values within smaller distances (e.g., 50 m) from observation points might representation of perceived visible greenery.In a different research, O'Regan et al. (2022) demonstrated a methodology to estimate air pollution, incorporating GVI, NDVI, and other variables.Thus, using NDVI to estimate GVI can be a potential solution to ll the street view data gaps for modelling greenness visibility.
A few studies have demonstrated the potential of utilising Mapillary's images for evaluating and quantifying urban greenness visibility (Liang et al., 2022;Yap et al., 2022), and numerous studies have used NDVI to measure the presence of greenery (Martinez and Labib, 2023).Despite these promising ndings regarding the potential use of the data source, there remains an absence of systematic evaluation regarding the usability of Mapillary data in different urban contexts for estimating greenness visibility and the potential of using NDVI to estimate greenness visibility in missing locations.Additionally, there is a lack of complete methodological work ow to use Mapillary data in different places and reproduce these analyses, hindering the usage of such open-source and FAIR data in modelling greenness visibility in diverse geographic contexts.
To address the methodological limitations and exiting research gaps noted above, this study aims to provide a scalable, reproducible methodological framework for utilising Mapillary data for greenness visibility modelling as well as evaluate the completeness and usefulness of such data in diverse geographical contexts.The primary objective is to establish a robust and replicable methodology that can be applied for greenness visibility modelling at any locations.Additionally, the secondary objective of this study is to evaluate the use of NDVI to estimate GVI in locations where street view images are unavailable.To achieve these objectives, this study will examine the following research questions: RQ1) How does the image availability and usability of Mapillary data for green view assessments change across different cities worldwide?RQ2) To what extent is the NDVI suitable for lling in missing data points in the GVI assessment?RQ3) Are there differences in street level greenness visibility between cities located in different geographic contexts?
To answer these questions, we will conduct a comparative analysis of the GVI across different cities worldwide, allowing to assess the quality and completeness of the data available through Mapillary.This comparative approach will provide insights into the reliability and applicability of the platform for greenness visibility assessments in different parts of the world.The development of this methodology aims to democratise the analysis of urban greenness, making it FAIR for researchers, urban planners, and policymakers interested in assessing and monitoring the visibility of greenery in their cities.By establishing such a methodology, our project seeks to fostering a better understanding of the relationship between urban natural environments, human well-being, and urban sustainability.

Study areas
Our study examines a diverse range of cities across diverse geographic, demographic, and socioeconomic contexts and located in multiple countries and continents.The cities included in our research are City of Melbourne, Australia; Amsterdam, Netherlands; Kampala, Uganda; Tel Aviv, Israel; Seattle, USA; Mexico City, Mexico; and Dhaka, Bangladesh.City of Melbourne, Amsterdam, Seattle, and Tel Aviv are known for their advanced urban planning and commitment to environmental sustainability initiatives (Bush et al., 2021;Dierwechter, 2017;Herscovici et al., 2022;Mora & Bolici, 2017).These cities, situated in developed nations, have been able to invest signi cantly in their infrastructure, incorporating technology and innovation to enhance their urban environments (Bush et al., 2021;Dierwechter, 2017;Herscovici et al., 2022;Mora & Bolici, 2017).In contrast, Kampala, Dhaka, and Mexico City, located in developing regions, face unique challenges related to rapid urbanization (Abebe, 2013;Aguilar & Lopez, 2018;Mortoja & Yigitcanlar, 2022).These cities experience high population growth and are often faced with issues of limited resources in terms of infrastructure and public services (Abebe, 2013;Aguilar & Lopez, 2018;Mortoja & Yigitcanlar, 2022).Their inclusion in the study provides a contrasting perspective on urban development and the role of Volunteered Street View Imagery (VSVI), such as Mapillary, in dynamic and resource-constrained contexts.Since one of our aims is to evaluate the quality and completeness of the data available through Mapillary, we assume that these diverse cities will provide valuable insights into the reliability and applicability of the platform for assessing greenness visibility in different parts of the world using opensource images.It will also allow identifying variations or limitations in the data between cities.

Data Sources
The data sources for this study primarily consisted of open-source platforms, which provided access to a wide range of geospatial data.Three major data sets were utilised in this research, namely: the Street Network with OSMnx, the VSVI with Mapillary, and the Satellite-Derived NDVI.

Street Network with OSMnx
Street network data was retrieved using OSMnx, an open-source Python library speci cally developed for working with Open Street Map (Boeing, 2017).OSMnx provides a convenient set of functions and methods for downloading and processing street network graphs for speci c regions (Boeing, 2017).Our research employed this library to extract simpli ed street network data, with a speci c focus on obtaining street link information.

VSVI with Mapillary
The VSVI data was obtained by using the Mapillary API (Mapillary, 2021).The API provides access to vector tiles endpoints that contain information such as the position of images and sequences with their original geometries.Within the coverage tiles endpoint, three layers are available: overview, sequence, and image.For our research, we focused on the image layer, which contains details such as the compass angle of the image, the timestamp of image capture, the image ID, and a ag indicating whether it is a panoramic image.

Satellite-derived Normalised Difference Vegetation Index
The Satellite-Derived NDVI data were used in our study to determine the GVI on points in areas where images were unavailable.Speci cally, the NDVI data was derived from Sentinel-2 satellite imagery, which offers a resolution of 10 meters (Satellite Imaging Corporation, 2022).To ensure the quality and reliability of the NDVI data, certain criteria were applied to selecting satellite imagery.Firstly, the chosen imagery had a cloud cover of less than 10% to mitigate the potential distortion caused by clouds, which can obstruct the accurate assessment of vegetation indices (Meraner et al., 2020).Secondly, the selected satellite imagery was collected considering the best pixel over the course of the year 2020.This approach involved selecting the image with the most favourable conditions for vegetation analysis within the given time frame, maximising the accuracy and representativeness of the NDVI data for the analysis (Corbane et al., 2020).We used estimate NDVI Google Earth Engine (GEE) to identify images within the period spanning the rst and last day of 2020, from which a composite image was generated.The composite image consisted of pixels with the median value across all identi ed pixels based on the criteria noted above.The NDVI values range from − 1 to + 1, where a value of -1 represents water, values around 0.2 suggest vegetation coverage, and higher values indicate denser forest coverage.To ensure that only image pixels potentially containing vegetation were considered, values below 0 were removed from the analysis.

Modelling GVI using open-source data 2.3.1 Step 1. Retrieve Street Road Network and Generate Sample Points
Figure 1 shows an overview of our methodology.First, the simpli ed street network data, represented as a graph object with nodes and edges, was obtained from OSMnx by indicating the name of the city to be analysed (e.g., "Amsterdam, Netherlands").It is worth noting that bidirectional streets are represented as two separate overlapping roads, with the starting and ending nodes inverted.To avoid analysing the same road twice, we implemented a ltering strategy to retain only one of the duplicate roads and geographically projected the cleaned network.Once the street network data cleaned, we proceeded to conduct point sampling at regular intervals of 50 metres along the roads.These sampled points serve as reference locations for selecting the corresponding images from Mapillary, which were then used to analyse the GVI.

Step 2. Assign Images to Each Sample Point Based on Proximity
To access the VSVI data, requests need to be made to the Mapillary API using the tile coordinates and zoom level of 14 (Mapbox, 2023).Therefore, to implement the data retrieval, we divided the area covered by the obtained road network into tiles.For each tile, we accessed the metadata of all the available images within it.Subsequently, we assigned one image to each sample point in the road network based on their proximity using the cKDTree algorithm (SciPy, 2022).In our case, the algorithm was queried to identify the closest neighbour within a maximum distance of 25 metres from each sample point.This distance was chosen based on the spacing of the sample points, which are 50 meters apart.By using this distance, we can ensure that no two points will be assigned the same image.Once we had determined the image ID of the closest image to each sample point, we were able to access the corresponding image by using the URL provided by the image endpoint (Mapillary, 2021).

Step 3. Image cleaning and segmentation
After obtaining the image, the subsequent step involved semantic image segmentation and image ltering process to clean and segment the images for further processing.To begin, for panoramic images, we performed a cropping operation to remove the bottom 20% of the image.This area corresponds to a band captured by the camera that is present in all panoramic images.By removing this band, we ensured that only the relevant portions of the image were retained for further analysis.
In the next stage, we applied the Masked-attention Mask Transformer (Mask2Former) architecture based image segmentation model (Cheng et al., 2022) to segment the images in different objects.The Mask2Former (Fig. 2) is a powerful architecture designed for universal image segmentation, encompassing tasks such as panoptic segmentation, instance segmentation, and semantic segmentation (Cheng et al., 2022).To achieve its capabilities, the Mask2Former model incorporates several key components.One of the central elements is the use of masked attention, a mechanism that enables the extraction of localised features by constraining cross-attention within predicted mask regions.By doing so, the model can focus on relevant regions and capture ne-grained details, resulting in highly accurate segmentation results (Cheng et al., 2022).
The performance of Mask2Former was evaluated on four popular datasets: Common Objects in Context (COCO), Cityscapes, ADE20K, and Mapillary Vistas (Cheng et al., 2022).Notably, when tested on the Mapillary Vistas dataset, the model achieved impressive mean intersection over union (mIoU) scores of 57.4 for small objects and 59.0 for medium and large objects.These results underscore the competitive performance of Mask2Former compared to other state of-the-art methods on Mapillary Vistas (Cheng et al., 2022).In our study, Mask2Former enabled the identi cation and classi cation of different objects within the images, including vegetation and roads, that were used for subsequent analysis.
To determine the suitability of each image for analysis, we employed an algorithm designed to identify the road centres from the segmented image.If road centres were detected, then the image was considered suitable and retained for further calculation (Fig. 3a, b).Conversely, if no road centres were identi ed, the image was deemed unsuitable and subsequently discarded (Fig. 3c, d).
This technique was applied to ensure we extract most representative viewing angle from the images, as many Mapillary images often has poor orientation, and irregular camera angle (Fig. 3c, d).Therefore, excluding such images may improve the overall data quality and visibility estimation because irrelevant or less informative images were removed.In cases where an image was both suitable and panoramic, we performed an additional cropping operation (Fig. 3a1, a2).The image was divided into "N" parts based on the identi ed road centres.This division allowed for a more focused examination of speci c sections of the panoramic image and provide a holistic representation of the surrounding area.

Step 4. Calculate GVI
The GVI equation used in this study was based on Eq. 1 proposed by Yang et al. (2009) for evaluating the visibility of urban forests.Their GVI calculates the ratio of the total green area from four pictures taken at a street intersection to the total area of the four pictures.
Where Area green_pixels_i represents the count of green pixels in the image captured in the ith direction (north, east, south, or west) at a speci c intersection and Area total_pixels_i corresponds to the total number of pixels in the image captured in the ith direction.However, due to the unavailability of images captured in all four directions for all sample points in our dataset, we made a slight modi cation to the formula.The modi ed GVI we used can be expressed with Eq. 2.
Where Area green_pixels_i is the number of green pixels in the ith picture found in the sample point, Area total_pixels_i is the total pixel number of the ith picture found in the sample point and N is the number of analysed pictures in the sample point.For non-panoramic images, N is equal to 1 since there is only one image available at the sample point.However, for panoramic images, N corresponds to the number of road centres identi ed at the sample point, indicating the number of cropped images.

Evaluate Image Availability and Image Usability
In assessing image availability, we examined the presence of an assigned image for each sample point.Any sample point that did not have a corresponding image found within a 25-metre buffer was excluded.
The availability of images in the dataset is expressed in Eq. 3 as the ratio of the number of sample points with assigned images ( ) to the total number of sample points .
Similarly, in evaluating image usability, we focused on points that had assigned images and a known GVI value.This indicated that the image met the expected criteria for inclusion in the analysis.The quality of the data is represented in Eq. 4 as the ratio of the number of sample points with assigned images and known GVI values ( ) to the total number of sample points with assigned images ( ).
Then, to enable a more meaningful comparison between cities of varying sizes, an adjustment was made to both scores by incorporating the natural logarithm of the road length in kilometres, as shown in equations 5 and 6.By applying this logarithmic transformation, the adjustment places greater emphasis on the relative differences in road lengths between cities, rather than absolute values.
In addition to the availability and usability assessment, we conducted further analysis to provide additional insights.Firstly, we measured the proportion of panoramic images in each city.This assessment allows us to evaluate the availability and coverage of panoramic imagery, which can provide a more comprehensive view of the surrounding locations.Furthermore, we examined the types of roads (e.g., residential, primary, secondary roads) with the most missing images.By identifying the speci c road types where the image coverage is limited, we can gain insights into potential gaps in the dataset.This information is valuable for understanding the limitations and areas that require further attention in improving the availability of images for analysis.

Model GVI for missing image points
To estimate the GVI values for the missing data points, we employed two modelling approaches per city: Linear Regression and Linear Generalised Additive Models (GAM).In both cases, we utilised the NDVI raster le of each city.For each point in the dataset, we computed the average NDVI value within a xedsize buffer surrounding the point.After conducting tests with multiple buffers of 100, 50, and 25 metres, it was determined that the buffer of 25 metres exhibited the best model performance.Subsequently, we employed the points with known GVI values and their corresponding calculated NDVI values to train both the Linear Regression and the Linear GAM models.The performance of these models was evaluated using the Root Mean Square Error (RMSE) through cross-validation with 5 folds.Additionally, to facilitate a comparison between the two models, we computed the Akaike Information Criterion (AIC) score.Finally, both models allowed us to estimate GVI values based on the available GVI and NDVI data in each city, thereby completing the information for the missing points.

Computational resources
The data processing and analysis for this study were implemented using the Python programming language.Python provides a wide range of libraries and tools that facilitated the handing and manipulation of de data, as well as the implementation of various algorithms and models.To execute the code, we utilised Google Colab.The Google Colab environment provided several advantages, including the availability of a high amount of RAM and GPU acceleration, which signi cantly contributed to the e ciency of the image segmentation process.For this research, we employed a computational system with 25.5 GB of RAM and a T4 GPU accelerator with a 15 GB memory capacity.This con guration allowed us to handle large datasets and smoothly execute memory-intensive operations, resulting in a substantial reduction in processing time for each image.Additionally, we implemented parallel processing techniques to enhance the e ciency of the image segmentation process and maximise the utilisation of the available computational resources.Through the combination of these implemented techniques and computational resources described above, we achieved high processing capabilities that enabled us to analyse approximately 4,000 images within a span of 40 minutes.

Image availability and image usability of Mapillary data
After designing and implementing the methodology to calculate the GVI using Mapillary data in any city worldwide, we obtained the GVI information for seven different cities.In order to provide a comprehensive overview of our ndings, we will rst address Research Question 1, which involves comparing the GVI across different cities and evaluating the availability and usability of image data obtained through Mapillary.Figure 4 provides a visual representation of the calculated GVI values using exclusively Mapillary image data, highlighting disparities in image availability across the selected cities.As illustrated, Fig. 4a corresponding to Seattle, Fig. 4b to Amsterdam, and Fig. 4g to City of Melbourne, consistently exhibit the highest degree of image availability.This observation aligns with the high AIAS achieved by cities, as shown in Error!Reference source not found.. On the other hand, our ndings reveal a contrasting picture for Fig. 4c corresponding to Tel Aviv, Fig. 4e to Dhaka, and Fig. 4f to Kampala, which also exhibit signi cantly lower AIAS.Intriguingly, Fig. 4d shows a signi cant disparity in image availability within Mexico City, with its southern region displaying considerably fewer images compared to the central area.
Regarding the usability, as presented in Table 1, our ndings provide intriguing insights into the usability of street-view imagery, adjusted for city size, in our seven cities of interest.Notably, Mexico City stands out with the highest AIUS, suggesting that despite the uneven distribution of image availability as indicated in Table 1, the proportion of usable images remains high.Similarly, Tel Aviv, exhibits a relatively high IUS despite having the lowest availability in Table 1, indicating that although the volume of images in this city was lower, the proportion of usable images was high.Furthermore, Amsterdam once again demonstrates a high AIUS, consistent with its high Image Usability Score.This consistency indicates that the robustness of the usable image data in this city is largely maintained, even after accounting for city size.In terms of panoramic images, the data presented in Table 2 reveals a considerable difference in the availability of panoramic images.Amsterdam stands out prominently with the highest number, as more than 80% of its images are panoramic.In contrast, the remaining cities exhibit considerably lower proportions of panoramic images.For instance, Dhaka did not have any panoramic images in its dataset, while Mexico, City, City of Melbourne, Kampala, and Tel Aviv all have less than 2% of their total images classi ed as panoramic.Next, we analysed the presence of missing values in each street of the cities, identifying the top ve highway types with the highest number of missing images.Furthermore, we calculated the proportion of missing images corresponding to each category in every city.Our analysis revealed a consistent trend across all cities, with residential streets displaying the highest proportion of missing images, as depicted in Fig. 5. Figure 5 illustrates that not only do residential streets have the highest volume of missing images in each city, but this de cit is notably signi cant, with all cities surpassing a 30% mark.This consistent and substantial percentage underscores a signi cant data gap within residential areas.

Modelled GVI for missing points using NDVI
Regarding Research Question 2, the viability of using NDVI to ll in missing data points was assessed by utilising Linear Regression and Linear GAM.Our results, shown in Table 3, reveal some key insights.Across cities, Linear Regression consistently resulted in lower RMSE values compared to Linear GAM, indicating a more accurate model with smaller residuals.The AIC scores highlight a consistent preference towards the linear Regression models over the linear GAM models across all cities.For instance, in Seattle, the Linear Regression model resulted in an AIC of -161,384.4018,while the Linear GAM model yielded a less effective AIC of -136175.2627.This pattern is repeated across all examined cities, indicating a more favourable balance of t and simplicity for the Linear Regression models when predicting GVI with NDVI.Furthermore, Fig. 6 shows the trend line and R-squared values observed across the tested cities, showing a relatively low value for most of them.However, the comparatively lower RMSE values obtained in our study indicate that, despite the low R-squared values suggest a degree of unexplained variability, the RMSE values provide a more optimistic outlook on the utility of our models for predicting GVI using NDVI.
Figure 7 provides a visual comparison between missing GVI values and their predictions using Linear Regression models for two contrasting cities: Amsterdam and Kampala.The black squares in the zoomed images show that the predicted GVI values closely align the calculated GVI values using street-view imagery for the surrounding points in both cities.

Mean and median GVI comparison between cities
For Research Question 3, Fig. 8 provides a visualisation of the mean and median GVI value for each street in the studied cities indicating the differences in street-level greenness visibility between cities.The GVI values show clear variations between developed and developing cities across continents in terms of green visibility.Seattle in North America has a higher value of 30.05% (Fig. 8a), indicating abundant visible green spaces.In Europe, Amsterdam follows closely with a GVI of 22.38% (Fig. 8b).However, Asian, and African cities like Dhaka and Kampala have lower GVI values of 16.62% (Fig. 8e) and 18.52% (Fig. 8f), respectively.These variations underscore the disparities in green view visibility between developed and developing cities, with the former having a better supply.

Main ndings
In this study, we aimed to develop a methodological framework for utilising open-source image data for greenness visibility modelling and evaluate the completeness and usefulness of such data in multiple cities in different countries worldwide.To our best knowledge, no previous studies have developed such a methodology using open-source images (e.g., Mapillary) and implemented similar approaches in diverse geographic contexts.Our study revealed signi cant insights into using Mapillary data to calculate GVI across different cities worldwide, effectively addressing our rst research question.We found that image availability, represented by the AIAS, varied signi cantly among the selected cities. Seattle, Amsterdam, and the City of Melbourne consistently had high image availability, while cities like Tel Aviv, Dhaka, and Kampala had much lower scores, as indicated in Fig. 4 and Table 1.Interestingly, we found that the usability of images did not necessarily correlate with their volume.Table 1 shows that despite having fewer images, Mexico City and Tel Aviv had high AIUS indicating that a large proportion of the available imager was usable.This suggests that meaningful GVI assessments are possible for places with usable images even though the city's overall image availability is limited.The distribution of panoramic images also showed a remarkable contrast across cities (Table 2).The disparity could impact the depth and breadth of GVI analysis, as panoramic images typically provide a more comprehensive view of greenery in the surrounding area than a single image with a limited view angle.Furthermore, our study pinpointed a considerable data gap in residential areas.We observed that these areas consistently showed the highest volume of missing images, as illustrated in Fig. 5.This consistent de cit could pose potential challenges for comprehensive GVI assessments, highlighting the need for effective data supplementation methods.
Our results are consistent with the previous study by Juhász & Hochmair, (2016), who also demonstrated the availability of images varies considerably between countries and street types.
Addressing our second research question, our analysis also suggests that the NDVI can be used to calibrate a linear regression model for each city to ll in missing GVI data points with reasonable errors.In Fig. 7 we illustrated that even in cities like Kampala, where the number of missing images is signi cantly

Strengths and limitations
Our study exhibits multiple strengths that reinforce its relevance within the urban green visibility research.Firstly, it used open-source, free, and FAIR data from Mapillary, making a signi cant stride in democratising and promoting open science in GVI assessments.This method establishes an example for subsequent research while highlighting the effectiveness of open data for scienti c purposes.Secondly, the scalability of the designed methodology in our research is a noteworthy aspect.Our approach to calculating GVI is not con ned to the selected global cities but can be replicated and scaled for any city worldwide.This scalability is particularly signi cant considering the growing need for GVI assessments in urban spaces.
Thirdly, our innovative approach in modelling visibility by combining GVI and NDVI represents a step forward in the eld.By using NDVI and known GVI points we developed models that allows to ll in the missing GVI data points in a city, we provide a novel solution that holds the potential to expand and enhance the depth of GVI studies, even in contexts where street-view imagery data may be limited.Fourthly, it is important to emphasise that the code used in this study is reproducible.The implemented code is publicly available, thereby encouraging transparency, peer veri cation, and further innovation in the eld.Reproducibility not only reinforces the validity of our ndings but also empowers other researchers to build upon our work, advancing our collective understanding of urban green spaces and human well-being.
Nevertheless, our research had certain limitations that need to be acknowledge for a comprehensive understanding of our results.Despite the utility of Mapillary, we encountered poor image availability in Dhaka, Kampala, Mexico City, and Tel Aviv that can limit the assessment of GVI solely with street-view images.However, Amsterdam, City of Melbourne, and Seattle exhibited a more favourable situation regarding image availability.These disparities in image availability can be attributed to the socioeconomic backgrounds of the cities (Fry et al., 2020) that may potentially lead to a less engaged community contributing to crowd-sourced database.
Another signi cant limitation was the temporal nature of our study, given that the Mapillary data were sourced at a speci c point in time, and it may not be updated frequently or uniformly.The dynamic character of crowd-sourcing databases could mean that the image data might have been augmented, updated, or even rendered obsolete since our data collection.However, these issues are not exclusively to Mapillary; even GSV imagery exhibits signi cant variability in availability and quality (Biljecki & Ito, 2021;Hara et al., 2013).Further, the infrequent updating of images impedes temporal assessment, as certain regions may remain unchanged for years (Hara et al., 2013).Such limitations are particularly evident in lower-income regions, rural areas, and countries in Africa, South America, and Southeast Asia, where GSV imagery remains unavailable or limited (Biljecki & Ito, 2021;Rzotkiewicz et al., 2018).
Moreover, a few speci c limitations of using OSMnx for urban studies have been identi ed.One limitation is related to data availability and consistency, which is a recurrent issue in urban planning and street network analysis literature (Boeing, 2017).The data used by OSMnx, sourced from Open Street Map (Boeing, 2017), is provided by a community of contributors and thus, might not always be consistent or upto-date.This means that any changes to the street networks that have occurred since the last update would not be re ected in the analysis Furthermore, the reliance of OSMnx data quality and accuracy can vary considerably based on the location.The coverage of OSMnx is good across the United States and Europe, but developing countries might have less thorough street network coverage (Boeing, 2017).
Finally, while our study validated the use of NDVI to effectively ll in missing GVI data points, this approach is not devoid of challenges.NDVI primarily quanti es vegetation from bird's eye view and may not fully capture the nuances of human perception greenery (Labib et

Future research
While this study has provided valuable insights into the usability of Mapillary data for calculating GVI, there are several promising avenues for further investigation.One potential area of investigation is obtaining a weighted average GVI value using population density in determined areas, which can provide a more accurate re ection of green visibility within a city.This approach can provide a more precise depiction of green visibility within a city, considering areas where the presence of greenery might be less perceptible due to low or negligible population density.In addition to recognising the positive impact of green spaces on well-being, research has also highlighted the potential of blue spaces, such as bodies of water, in promoting human health.(Labib et al., 2020).Therefore, it would be valuable to explore the feasibility of integrating Mapillary data to calculate the Blue View Index for cities.This integration would enable researchers to investigate the in uence of both green and blue spaces on various health indicators, providing a more comprehensive understanding of the relationships between urban environments, natural features, and human well-being.Furthermore, to improve the estimation of GVI for missing points using mean NDVI values, it is worth considering the inclusion of more variables, such as building density, height, and orientation.However, such building data are mostly unavailable globally, but more and more data on building are becoming available such as OSM building, Microsoft Building Footprints.Exploring these future research directions will not only expand upon the ndings of this study but also contribute to the development of approaches for urban planning and public health interventions.

Conclusions
This research has developed and demonstrated a scalable, reproducible framework for utilising opensource and FAIR Mapillary image data to assess greenness visibility in diverse geographical contexts.
Importantly, this research effectively addressed its research questions, demonstrating that despite variations in image availability across different cities, a robust methodology for GVI assessment can be established using images from Mapillary.Our study identi ed disparities in Mapillary data availability and usability across the selected global cities, bringing to light an uneven global distribution of open-source and free image data.Nonetheless, these disparities did not diminish the potential of Mapillary for GVI evaluations.Additionally, our ndings accentuate Mapillary's adherence to FAIR principles, an asset in scienti c research in this domain.This feature makes Mapillary an advantageous free and FAIR alternative to Google Street View for GVI assessments despite the found discrepancies and temporal limitations that are not only inherent to crowdsourced data.Our methodology enables new avenues to make urban greenness analysis accessible to all.A key innovation of our study was using globally available NDVI data to estimate GVI values at missing locations.Although not without limitations, this approach proved to be a promising method for augmenting street-level GVI analyses in data-scarce areas.The successful application of NDVI in our study paves the way for future research to re ne and expand on this method, further enhancing the robustness of GVI analyses.Additionally, this study provided a comparative analysis between cities of varying socioeconomic contexts, nding a notably lesser greener visibility in densely populated, lower-income cities in South Asia and Africa as compared to their low-density, a uent counterparts in the USA and Europe.Overall, this study serves as a critical contribution to urban planning and environmental research, providing an accessible, scalable, and innovative approach to greenness visibility assessment.Our ndings and established framework, founded on open-source data and reproducibility, facilitate comparative studies, and establish guidelines for green view assessments across the globe using free images.Through this, we hope to foster a greater understanding of the relationships between urban natural environments and human well-being, driving toward healthier and more equitable, and sustainable cities.        Comparing Image Availability between Amsterdam and Kampala, with GVI Estimation for missing points using a linear regression model for each city.

Figure 3 Example
Figure 3

Table 2
Proportion of Panoramic Images for each analysed city.

Table 3
Linear Regression and Linear GAM model evaluation.
(Dierwechter, 2017;Mora & Bolici, 2017;Abebe, 2013;Mortoja & Yigitcanlar, 2022)et-view images.We argue that even though NDVI does not directly indicate the visibility of greenery, as they indicate the presence of vegetation, the modelled relations can reasonably estimate potential GVI values.Torkko et al. (2023)also observed that NDVI values might correspond to perceived visibility greenery.To further validate the reliability of these ndings, we compared our overall street-level Green View Index (GVI) values with data from MIT's Treepedia project (2023) conducted by the Senseable City Lab(Cai et al., 2018).The Treepedia project utilized Google Street View (GSV) data to model GVI in various cities, including the cities in our study (Amsterdam, Tel Aviv, and Seattle).Our analysis revealed that the median street-level GVI values for Amsterdam and Tel Aviv were similar to the estimates provided by Treepedia.However, in the case of Seattle, our approach slightly overestimated the overall GVI score.This discrepancy can be attributed to the limited availability of panoramic images in the Mapillary dataset (only 15% coverage in Seattle) compared to the predominantly panoramic GSV data used by Treepedia.Additionally, our study employed NDVI (Normalized Difference Vegetation Index) to estimate GVI in areas where data was missing, which was not implemented in the Treepedia project(Cai et al., 2018).Finally, for the third research question, our analysis underscores a notable disparity in GVI across cities, particularly in high-density cities in Africa and South Asia.As Fig.8illustrates, these densely populated, lower-income cities exhibit considerably less green visibility than low-density, high-income cities in the USA and Europe; such results are consistent with Treepedia data.This disparity may be attributed to various factors, including the differences in urban planning strategies and investment in public spaces, the socioeconomic development level, as well as the historical and cultural values assigned to green spaces in these regions(Dierwechter, 2017;Mora & Bolici, 2017;Abebe, 2013;Mortoja & Yigitcanlar, 2022).