fruitpunch

Guide · AI & data

AI for Greener Cities: Urban Trees, Heat and Green Space

By the FruitPunch research deskUpdated November 15, 20269 min read

In short

AI helps cities green up by mapping every tree, measuring street-level greenery and finding the hottest blocks from aerial, satellite and Street View imagery. Google's Tree Canopy tool grew from 14 to nearly 350 cities in 2023, and a Lancet study of 93 European cities estimated that raising tree cover to 30% would cool them by 0.4°C on average and could prevent 2,644 premature summer deaths. Satellite data from 293 European cities show tree-covered areas run 8 to 12 K cooler than dense urban fabric in Central Europe.

This address once hosted FruitPunch AI, an Eindhoven AI for good community. FruitPunch is a new, independent publication and is not that organization. This guide was written from public sources; the history page explains the change.

AI makes cities greener mainly by doing the mapping work that used to take crews of surveyors: it finds every tree in aerial photos, scores how green each street looks in Street View panoramas, and combines that with satellite heat data to show which neighborhoods need shade first. The stakes are measurable. A Lancet health impact assessment of 93 European cities found that urban heat islands raised summer temperatures by 1.5°C on average and estimated that 6,700 premature deaths in summer 2015 could be attributed to them, while raising tree cover to 30% "could prevent" 2,644 deaths. This guide explains how the methods work, which tools and open datasets exist, and what they can and cannot tell a city.

Key facts

  • 1 to 7°F: how much warmer daytime temperatures are in US urban areas than in outlying areas because of the heat island effect, with nights 2 to 5°F warmer (US EPA).
  • 8 to 12 K: how much lower land surface temperatures are under urban trees than over continuous urban fabric in Central Europe, across 293 European cities (Schwaab et al., Nature Communications, 2021).
  • 2,644 deaths: premature summer deaths that 30% tree cover could prevent across 93 European cities (Iungman et al., The Lancet, 2023).
  • Nearly 350 cities: coverage of Google's AI Tree Canopy tool after its 2023 expansion from 14 cities (Google).
  • 2.5 million trees, 23 cities: size of the Auto Arborist dataset for tree species recognition (Beery et al., CVPR 2022).
  • 68% by 2050: share of the world population projected to live in urban areas, up from 55% in 2018 (UN DESA).

Why do cities need to map their trees and heat?

More people live in cities every year. The UN Department of Economic and Social Affairs reported in 2018 that 55% of the world's population lived in urban areas, a share expected to reach 68% by 2050. Cities trap heat because pavement, roofs and walls absorb sunlight and release it slowly, and they have fewer plants to cool the air through shade and evaporation. The US Environmental Protection Agency summarizes research showing daytime temperatures in US urban areas about 1 to 7°F higher than in outlying areas, with humid and densely populated cities showing the biggest gaps.

Trees are one of the few fixes that work at street scale. The EPA describes vegetation as a simple and effective way to reduce heat islands. The hard part is knowing where trees already are, where they are missing and who lives in the hottest places. Tree inventories are expensive: the authors of the Auto Arborist paper note that city tree censuses built by human experts cost millions of dollars per census and are therefore collected infrequently. That gap is where machine learning helps.

How much do trees actually cool a city?

The best evidence comes from satellite thermal data. A 2021 study in Nature Communications used high-resolution land surface temperatures and land cover data from 293 European cities. It found that land surface temperatures under urban trees were on average 0 to 4 K lower than continuous urban fabric in Southern Europe and 8 to 12 K lower in Central Europe. Treeless green spaces such as lawns were about 2 to 4 times less effective at cooling than trees.

Land surface temperature is not the same as the air temperature people feel, so health studies model air temperature separately. The Lancet assessment by Iungman and colleagues worked on a 250 by 250 m grid across 93 European cities and estimated that bringing every city to 30% tree cover would cool them by a mean of 0.4°C, with a range of 0.0 to 1.3°C. That modest air cooling still translated into 2,644 prevented premature deaths, or 1.84% of all summer deaths in the study.

Study or source Scope Key figure
US EPA US cities, review of studies Days 1 to 7°F warmer, nights 2 to 5°F warmer than outlying areas
Schwaab et al., Nature Communications 2021 293 European cities, satellite LST Trees 8 to 12 K cooler than urban fabric in Central Europe, 0 to 4 K in the south
Iungman et al., The Lancet 2023 93 European cities, summer 2015 Heat islands added 1.5°C on average; 6,700 deaths attributable
Iungman et al., The Lancet 2023 Same cities, 30% tree cover scenario 0.4°C mean cooling; 2,644 deaths potentially prevented
Google Tree Canopy Lab Los Angeles, 2020 Over 50% of residents live where canopy is under 10%

How does AI map trees from aerial and satellite images?

Most tree mapping starts with overhead imagery. A model, usually a convolutional neural network trained on labeled examples, looks at each pixel or patch and decides whether it is tree canopy, grass, roof, road or water. Adding near-infrared bands helps because healthy leaves reflect strongly in near-infrared, and stereo images or lidar add height so a tree can be told apart from a lawn.

Google described this pipeline when it launched Tree Canopy Lab in Los Angeles in 2020: aerial photos taken from planes in spring, summer and fall, near-infrared images, and height maps built by comparing images from different angles, fed to a tree-detection AI that scans the images and produces a canopy density map. The analysis found that more than 50% of Angelenos live in areas with less than 10% tree canopy coverage and 44% live in areas with extreme heat risk, and that the coolest areas also had the most canopy.

At global scale the resolution is coarser but free. The European Space Agency's WorldCover maps land cover for 2020 and 2021 at 10 m resolution from Sentinel-1 and Sentinel-2. Dynamic World, described in Scientific Data in 2022, applies deep learning to 10 m Sentinel-2 imagery to provide an open, continuous feed of near real-time land cover predictions. Meta and WRI released a global map of tree canopy height at 1-meter resolution in 2024, which can detect single trees, with the data and models free to use.

What is Treepedia and how does street-level imagery help?

Overhead images show the tops of trees. Pedestrians see something else: the trees, hedges and shade along a street at eye level. The MIT Senseable City Lab built Treepedia to measure that view. According to its public code repository, Treepedia applies computer vision to Google Street View panoramas to compute a Green View Index, a 0 to 100 score for the share of canopy cover visible at a location. It focuses on street trees and does not map parks, because Street View cars rarely drive through them. The method traces back to a 2015 paper in Urban Forestry and Urban Greening, Assessing street-level urban greenery using Google Street View and a modified green view index.

Street-level images also make species identification possible, which overhead images struggle with. The Auto Arborist dataset, presented at CVPR 2022, joins public tree censuses from 23 cities with street level and aerial images. It contains over 2.5 million trees and 344 genera and was built to test how well models trained in one city work in another, a common failure point.

Vehicles that already drive every street are a natural source of this imagery. A past FruitPunch AI challenge, according to its archived challenge page, described working with the parcel company DPD and the Jheronimus Academy of Data Science on driving video data from delivery vehicles, with ideas such as spotting street waste or neighborhoods that could use more vegetation. This site is a separate publication and not connected to that project.

Can AI map air quality street by street?

Air pollution is the other side of green, livable streets. The World Health Organization reports that in 2019, 99% of the world's population lived in places where its air quality guideline levels were not met, and that ambient air pollution caused an estimated 4.2 million premature deaths that year.

Fixed monitors are sparse, so cities often have one station for a whole district. A 2017 study in Environmental Science and Technology equipped Google Street View cars with fast-response sensors and repeatedly drove every street in a 30 km² area of Oakland, California. The resulting 30 m maps of nitrogen oxides and black carbon showed pollution varying by 5 to 8 times within single city blocks, a pattern central monitors cannot see. Machine learning models trained on data like this can then estimate pollution on streets the cars did not cover, using traffic, land use and greenery as inputs.

Which tools and open datasets can people use?

These are real, documented resources that a city team, researcher or student can start with today.

Tool or dataset Who runs it What it offers
Tree Canopy in Environmental Insights Explorer Google AI canopy maps and heat context for nearly 350 cities as of 2023
Treepedia and Treepedia_Public code MIT Senseable City Lab Green View Index from Street View, open source under a BSD license
Auto Arborist Beery et al., CVPR 2022 2.5 million trees, 344 genera, 23 cities, street and aerial views
ESA WorldCover European Space Agency Global 10 m land cover for 2020 and 2021
Dynamic World Brown et al., Scientific Data 2022 Near real-time global 10 m land cover from Sentinel-2
Global canopy height map Meta and World Resources Institute 1 m canopy height, models and data free
ECOSTRESS NASA JPL Thermal instrument launched to the International Space Station in 2018
i-Tree Built on USDA Forest Service science Tools that estimate tree benefits; the site notes a Forest Service funding halt

A practical workflow for a mid-sized city looks like this: start with a free canopy layer (Tree Canopy if the city is covered, otherwise WorldCover or Dynamic World), overlay a summer land surface temperature map from satellite thermal data, add population and vulnerability data from the census, and rank blocks by heat, low canopy and number of residents. Street-level Green View scores then help pick the exact streets where new planting will be visible and shade sidewalks.

What are the limits of AI tree and heat maps?

  • Resolution and timing. A 10 m pixel can mix a tree, a roof and a road, and many global products are annual. Small street trees can vanish at that scale.
  • Surface heat is not air heat. Satellite land surface temperature shows how hot materials get, which is related to but different from the air temperature and humidity that affect health. The Lancet study needed a separate air temperature model.
  • Domain shift. The Auto Arborist authors built their dataset because models that do well in one city often fail in another with different species, lighting and street design.
  • Street View coverage. Treepedia skips parks and private yards, and image dates vary by street.
  • Planting is the slow part. Maps tell a city where to plant; trees still need water, space and years to grow before they cast real shade.

Frequently asked questions

What is an urban heat island?

An urban heat island is a city area that is warmer than its surroundings because buildings and pavement absorb and hold heat and there is less vegetation. In the United States, the EPA reports daytime temperatures in urban areas about 1 to 7°F higher than in outlying areas.

How does Google Tree Canopy work?

Google's Tree Canopy, part of its Environmental Insights Explorer, runs a tree-detection AI over aerial photos, including near-infrared and height data, to map canopy density. Google said in 2023 that it had expanded from 14 cities to nearly 350, and that Austin used it to prioritize tree planting and place bus shelters for shade.

What is the Green View Index?

The Green View Index is a 0 to 100 score for how much tree canopy is visible from a point on the street, computed from Google Street View panoramas. MIT Senseable City Lab's Treepedia uses it to compare street greenery between and within cities.

How many trees does a city need to reduce heat deaths?

There is no single number, but a Lancet study of 93 European cities modeled 30% tree cover and estimated a mean cooling of 0.4°C and 2,644 premature summer deaths prevented, about 1.84% of summer deaths.

Are urban trees better than parks with grass?

For cooling, yes. Across 293 European cities, treeless green spaces were about 2 to 4 times less effective at lowering land surface temperatures than urban trees.

Is this site the FruitPunch AI organization?

No. FruitPunch AI was an AI for Good community that ran challenges such as the one described on its archived AI for Greener Cities page. This site is an independent guides and rankings publication with no connection to that organization.

Spotted an error in this guide? Email editorial@fruitpunch.ai

Keep reading