Machine Learning in the Service of Urban Studies


Any analysis should start with the initial data. In many sciences, there are lots of parsed and prepared datasets, but urban studies are less fortunate with this. Modern man has satisfied his hunger, recovered from diseases, learned the secrets of the atomic nucleus and the universe — it’s probably time to equip the urban space around him. But first, it should be studied.

Despite the coronavirus, everything is heading towards the fact that mental disorders will become the main epidemic of the 21st century. Already today, the leading cause of disability aged 15–45 in the United States is depression. One in five people aged 60+ are taking antidepressants, and more than half of the US population has taken antidepressants at least once in the past fifteen years.

Scientific research shows that people in “greener” areas are less likely to suffer from depression and commit fewer crimes. The findings of the researchers: if we want a healthy and safe society, then we need not only pills and the police, but also a green urban environment.

Modern technologies of aerial photography and neural networks make it possible to identify the greenery of the urban environment down to individual trees. The time has passed when you need to delve into the master plans of landscaping or personally visit the area in order to explore the city. Instead of text and boring numbers — visualizations. Instead of archivists and inspectors — drone operators and programmers. We decided that all of the above could be applied with interest and benefit to urban problems and trained the neural network to recognize tree crowns.

The most accurate and visual way to display the greenery will be its detailed map, compiled not from documents, but from real images. And trees are different, ranging from young seedlings and bare poplars to spreading oaks. For clarity and contrast, we made a map of trees in black-green color, where the crowns of trees are indicated in green, black, and everything else. To simplify your orientation, on the right you can see the usual map of the central part of Vilnius:

Dormitory areas are full of green pixels — good. Some of the black spots correspond to industrial zones — naturally. However, we would like to quantify this. To do this, we created a heat map of the trees in the capital. In fact, this will be a map of crown density:

The warmer the color, the greater the density of the trees. The historic center is almost completely built up and is displayed in purple — trees form only about 10% of the area. This trend is relevant for all old cities. The dormitory areas of Žirmūnai and Naujininkai look good — the share of tree area is 30–60%.

Despite the large area of ​​white-red areas, that is, a continuous forest, the north-western part of Vilnius is deprived of forest plantations. Some of the territories marked in black treeless color are simply grassy areas unoccupied by human activity. In our opinion, this is rather good than bad — there are prospects for it to be interesting to equip this area, in contrast to, for example, Manhattan, where it is impossible to demolish business skyscrapers.

It is interesting to look at some other European capital, for example, Paris. Administratively, the capital of France is small, so suburban dormitory areas have been added:

The city is contrasted. During Napoleon III in the middle of the 19th century, several large parks adorned Paris — these spots are visible on all maps. It is only one kilometer from the Arc de Triomphe to the Bois de Boulogne. The residential part of Paris and its suburbs would not attract dendrologists — the narrow streets, which are adjacent to the facades of houses, are not able to accommodate a sufficient amount of greenery and the bulk of the trees are located inside undeveloped courtyards. The west and southwest please with a high proportion of trees, but these are mostly non-refined forests. Further landscaping of the residential part of the city is complicated — there is no room left, c’est la vie.

The success of overseas colleagues in urban studies can be assessed in New York:

Manhattan is easily identified by the green elongated rectangle. In terms of “greenery,” the area looks empty, roughly like the industrial zone of Vilnius. From the south of Brooklyn, the crown density is small, 10–30%, with the exception of a few green neighborhoods and a park with a cemetery. Part of the reason is that Brooklyn has dense buildings and little space for greenery. The Bronx [north-center] and northeastern Queens look pretty green. The forested top left corner of the pictures is not a New York, but a suburb — New Jersey.

Moscow is also an interesting example:

The center of Moscow lacks landscaping, transport arteries are blackening. However, outside the third transport ring, residential areas are already buried in greenery. Numerous parks are 5–7 kilometers away from the center.

Closer to the outskirts, there are many red and white colors, and despite the names, these are not parks, but rather continuous forests with a crown density of 90–100%. It is not easy to be proud of this since significant areas of the outskirt “parks” are inaccessible or not equipped for walking.

A special dislike for landscaping is present in Asian megacities. Shanghai:

The city as a whole on the heat map looks “purple”, that is, the area of ​​trees is about 10%. There are parks in Shanghai, but they are small. There are also “green” neighborhoods [left center], but the houses themselves occupy more than half of the area there and the trees nestle in the stone jungle. One would think that the pentagon that stands out in green is a boulevard ring, but no — a tree-line enclosure of the ring highway.

What could be a truly “green” city? For contrast, we ran an aerial photo of one of the greenest cities in the United States through a neural network — Charlotte:

Charlotte is a forest in which people have built houses.

PS We also publish the trained neural network model and a simplified working example of the code. No programming skills are required to run.

Vyacheslav Laktyushkin, Polina Safronova, Acrux Cyber Services.