Examining the spatial distribution and temporal change of street trees using Google Street View and deep learning
As an important part of urban greenery, street trees help to improve air quality in urban street canyons that carry most of the human activities in cities. The street trees also act as a barrier to protect urban residents from noise pollution and visual intrusion. In hot summer, street trees help to provide shade and increase the thermal comfort for pedestrians. In addition, the existence of street trees usually increases people’s esthetic rating of streetscapes, which would further increase the resident’s physical activities and social cohesion in neighborhoods.
Planting trees has become a popular method to tackle urban environmental challenges, build climate resilience, and increase livability in cities. Many cities have launched urban greening projects aiming to increase the tree canopy cover. In 2007, New York City launched the MillionTreesNYC project aiming to plant one million new trees across the city. Following New York City, other cities such as Los Angeles, Denver, Boston, and many other cities around the world started similar campaigns to increase the tree canopies. However, planting and maintaining street trees in cities are usually difficult because of the harsh environment in the densely urban area for trees and the financial expense for the maintenance of trees.
In my new paper published on Environment and Planning B: Urban Analytics and City Science. I applied Google Street View and deep learning to monitor the spatial distribution and the temporal change of the street tree canopies in New York City from 2008 to 2018. This is a kind of extension of my previous Treepedia project by using more advanced deep learning algorithms and historical Google Street View images. This is also the first large-scale study using historical Google Street View images to examine the temporal change of street tree canopies.
Step 1. Collect Google Street View images
Because Google Street View images are distributed discretely along streets, therefore, the first thing to do is to create sample sites along streets. In my previous Treepedia repository, there is a description of how to create sample points along the streets. I have not maintained it a while. If you have any issues, let me know. When you get the sample points ready, then you can use those coordinates as the input to download the GSV images through Google Street View Image API. Fig. 2. show the pipeline to download the Google Street View images in New York City.
In order to examine the temporal change of the street trees canopies, we also need to check the distribution of time information of those GSV images in New York City. Fig. 3 shows the distribution of GSV capture-date in New York City.
The GSV images of any single year cannot fully cover the whole study area, therefore, I applied a multi-year rolling method that similar to the generation of American Census Survey 5-year aggregated data on GSV images to make sure the whole study area is fully covered. For each sample site, only the latest and oldest GSV images were used. All collected GSV images in the study area were recategorized into two time periods, 2008–2013 and 2014–2018 in order to match the census data and examine the temporal change of the street trees in New York City.
Step 2. Image segmentation using a deep learning algorithm
Extracting the street tree canopies from street-level images is a prerequisite for deriving quantitative information of street trees. With the advance of computer vision and deep convolutional neural network, it is possible to segment the street-level images accurately. Here I applied a repository created by the MIT CSAIL to do the image segmentation. Fig. 4. shows the image segmentation result for several static images in New York City.
Step 3. Calculate the Green View Index and map the spatial distribution
In order to quantify the amount of the street tree canopies, I used my previously developed green view index by focusing on the street trees. The new green view index for street trees can be calculated as,
where the Area_ri is the tree 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. The GVI hereby represents the visibility of the street tree canopies on the ground level from a pedestrian’s perspective.
Based on the collected GSV images and the image segmentation results, GVI values were calculated for all locations of the study area in different time periods (2008–2013 and 2014–2018). The point level GVI values can be further aggregated to census tract level by median values in order to be comparable with the census data. Fig. 5 shows the spatial distributions of GVI in the study area at the census tract level in the two time periods. It can be seen clearly that in both two time periods, the northern, eastern, and southern parts are greener than the central and western parts of the study area.
Fig. 6 shows the spatial distribution of the temporal change of the GVI from 2008–2013 to 2014–2018. Generally, the central area, where the GVI values are relatively low, gets more street tree canopy increase. The northern and southwestern parts of the study area have also experienced street tree canopy increase during that time.
I also compared the spatial distribution and temporal change of street trees with the socio-economic variables at the census tract level in New York City. For more details about the socio-environment analyses, please read my paper. Here is my conclusion, the street tree canopy is distributed unevenly across neighborhoods of different racial/ethnic groups in New York City and the Hispanics tend to live in neighborhoods with less street tree canopies. In the last ten years, the street tree canopies have increased slightly, while there is no sign that the Hispanics who tend to live in neighborhoods with lower street tree canopies have more street tree canopy increase. This study provides an efficient and scalable method to map and monitor the distribution and the temporal change of street tree canopies by combining deep learning algorithms and historical street-level images, which would provide an important tool for urban environmental planning.
Reference
Li, X. (2020). Examining the spatial distribution and temporal change of the green view index in New York City using Google Street View images and deep learning. Environment and Planning B: Urban Analytics and City Science, 2399808320962511.
