Examining the spatial distribution and temporal change of the green view index in New York City using Google Street View images and deep learning

Abstract

As an important part of the urban ecosystem, urban trees provide various benefits to urban residents. It is therefore important to examine the spatial distribution and the temporal change in urban tree canopies. Different from traditional overhead view remote sensing-based methods, street-level images, which present the most common view that people have of greenery, provide a more human-centric way to quantify street tree canopies. This study mapped and analyzed the spatial distribution and temporal change in the green view index, which represents the visibility of tree canopies along streets in New York City during the last 10 years using historical Google Street View images. Deep learning and computer vision algorithms were used to derive the quantitative information of street tree canopies from street-level images and map the spatial distribution of the green view index. This study further investigated the potential disparities in terms of green view index across different racial/ethnic groups by comparing with census data. Results show that non-Hispanic Whites tend to live in neighborhoods with higher green view index and Hispanics tend to live in neighborhoods with lower green view index. The green view index values in New York City have increased slightly in the last 10 years, and the change of green view index has no significant correlation with resident’s ethnic/racial status. This study proves the usability of historical Google Street View images for monitoring the temporal change of urban street tree canopy changes at large scale, and it also provides insights and a valuable reference for urban greening programs.

Residential Greenspace and Urban Adolescent Substance Use: Exploring Interactive Effects with Peer Network Health, Sex, and Executive Function

“…In the present study we collected the nearest geo-tagged GSV images that were within 50 feet of each adolescent’s home address using the GSV image API [ 52 ]. The deep convolutional neural network PSPNet, trained based on ADE20K, was used to extract street greenery from the street-level images, where accuracy for the identification of greenery typically approaches 95% [ 53 , 54 ]. Based on the resulting image segmentation, we generate the green view index (GVI), which is calculated as follows: where the Area ri is the green (i.e., tree, shrub) pixel number in one of the six pictures taken in six different directions and Area ti is the number of total pixels in one of the six images [ 55 ].…”

Correspondence Between Feelings Towards Neighbors and Appearance of Neighborhood: Analysis by Combining a Mail Survey and Google Street View

“…Those studies suggested that aesthetics and well-kept streets might encourage people to walk for health benefits, which seemed to be intuitive. Observations of greenery and walkability by automated methods through GSV could also reflect social inequality between areas (Ki & Lee, 2021;Li, 2021;Li, Zhang, Li, Kuzovkina, & Weiner, 2015), and be positively related with the vitality of neighborhood organizations (e.g., sports clubs) and strength of social networks in local communities (Wang & Vermeulen, 2021).…”

“…2 The API needed location coordinates to display the GSV imagery nearby. If the purpose of collecting GSV images is to capture comprehensive streetscapes of target regions regardless of the existence of houses, referring to coordinates spotted at a certain interval on the API would be a common procedure (e.g., Jiang et al, 2020;Ki & Lee, 2021;Li, 2021;Li et al, 2018;Li, Zhang, Li, Kuzovkina, & Weiner, 2015;Zhang et al, 2018). On the other hand, if the purpose of image collection is to observe streetscapes around certain buildings or facilities, using their addresses or coordinates would be a more reasonable procedure (e.g., Wang & Vermeulen, 2021).…”

“…Through those two image classification processes, we built two machine learning models to identify houses and plants in the extracted GSV images. While image segmentation has been employed to automatically classify objects from a massive number of GSV imagery by many previous studies (e.g., Ki & Lee, 2021;Li, 2021;Li, Zhang, Li, Ricard, et al, 2015;Nagata et al, 2020;Wang & Vermeulen, 2021;Zhang et al, 2018), a deep learning method termed as "chopped picture" (Ise et al, 2017) was applied instead in those models. The "chopped picture" method was originally developed to identify amorphous objects such as groups of various shapes of plants, unlike distinctive objects (e.g., human faces, dogs, and automobiles), which previous machine learning methods had focused on.…”

Near Real-Time Semantic View Analysis of 3D City Models in Web Browser

“…Panoramic images, such as those available from Google Street View (GSV), have become one of the most important sources for estimating the GVI, with several examples available [28,32,49,[51][52][53]. In these cases, the GVI is typically not estimated from four non-overlapping images (as in [44]), but rather from a set of segments of the 360 degree panoramic image, with slightly varying implementations.…”