Quantifying greenness of cities with satellite imagery and AI

Urban green spaces in Amsterdam segmented with 10m resolution Sentinel-2 images

An ever-growing scarcity of natural resources, climate change with increasing temperatures, droughts and floods, and a growing world population put cities under pressure. With cities growing exponentially larger and more people living, working and commuting within those cities green spaces are often the first to go.

With the changing climate, we face more frequent and more severe heatwaves. Green spaces can help to cool the city. The difference between a street with trees and a street without trees is very noticeable. Urban green spaces also help to capture stormwater and preserve biodiversity. The health and social benefits of urban green spaces have been discussed widely as well.

Benefits of Urban Trees (Source: The Nature Conservancy)

To help raise awareness and to help policymakers data is needed to improve decision making. To do this at a global scale, Husqvarna and 20tree.ai worked closely together to launch HUGSI. HUGSI is short of Husqvarna Urban Green Space Index and unveils insights about the size, distribution and health of green spaces in urban areas of 98 cities across 51 countries.

Within 3 months, from kick-off to launch, a dedicated team from both companies worked side-by-side to achieve this result. The teams consisted of a complementary set of talented people. Husqvarna’s expertise in green space management and 20tree.ai’s expertise in applying machine learning to satellite imagery proved to be a perfect match. The scale that we wanted to roll out the solution was unmatched. Green spaces in different cities have been compared in different studies, but often traditional remote sensing techniques have been used, for example, a simple NDVI threshold or simpeler methods like Random Forests. These show well-known shortcomings in terms of generalisability and accuracy.

HUGSI.green ranking

That’s were 20tree.ai’s expertise in deep learning comes in. From the beginning, it was clear that a thorough machine learning set-up and validation scheme was needed to make sure that the results were reliable and scalable with 10m resolution Copernicus Sentinel-2imagery. Ordering higher resolution satellite imagery, for example, 30cm or 1.5m resolution, was not scalable in terms of costs and availability.

With a limited training set, consisting of high-resolution labels of trees, low vegetation and grass, water and other, we had to train a deep learning segmentation algorithm to classify land-use at a global scale with 10m resolution. We needed to make sure that vegetation in different geographical areas, from Gothenburg to Dubai and Hong Kong is classified correctly. Therefore, we decided to keep the most challenging cities as a validation set, ensuring the generalisability of the algorithms.

Access to green spaces in and around São Paulo (Brazil) is not equally distributed

During the process, we’ve experimented with different segmentation algorithms. The beauty of satellite imagery is that it contains more bands than the visible RGB. Especially, Sentinel-2 with 12 bands, from Ultra blue to Visible and Near Infrared (VNIR) and Short Wave Infrared (SWIR) is extremely useful for analysing vegetation.

HUGSI is just the beginning. Husqvarna and 20tree.ai are committed to continuously extend and evolve our joint offerings. For municipalities, we provide more detailed analysis, that can include insights on tree level. For example, to assess biodiversity and changes over time.

Land-use classification Austin (USA)

How does your city rank on HUGSI.green? Do the numbers match your expectations?

Thanks to David de Meij.