An
experimentation project has demonstrated the capabilities of machine learning
in urban development. It used images as a starting point and came up with
interesting and useful applications.
“I read data
science papers on how machine vision algorithms can be used with satellite
imagery. I immediately saw a connection to what we had been doing,” Antti
Kauppi, architect at Arkkitehdit Sankari, explains. “Most people associate
image recognition with Google’s visual searches. Google can distinguish whether
a photo shows a cat or another animal, for example. We went a step further.”
An Experiment with Open Urban Imagery
Arkkitehdit
Sankari Oy, a Finnish architectural design firm began the experimentation
project CityCNN in May 2018. It received funding from KIRA-digi, the
Finnish government’s digitalization program for the built environment. CityCNN
explored the possibilities of using machine learning and open data for urban
development.
Kauppi
collected data from Espoo, Finland’s second-largest city on the outskirts of
Helsinki. He created a piece of Python software to retrieve data from the
city’s server.
Kauppi has
done programming for many years but does not consider himself an AI expert: “My
special skill is to apply the technology to our business. It’s much like using
Photoshop. You can use it successfully even if you don’t master its inner
workings.”
Using Competing Neural Networks
The
experiment used so-called generative adversarial networks (GANs). It is a
machine learning technique in which you match two neural networks against each
other.
The first
network, the generator, creates new images. The second network, the
discriminator, uses real images and takes in the newly generated images. It
evaluates whether an image is real or generated. In repeating this process over
and over again, the generator and the discriminator become more accurate and,
as a result, the generated images improve.
Conditional adversarial
networks used in CityCNN are an extension of the basic technique. They are
trained with image pairs. They can create photographic images from line
drawings or convert an impressionist painting into a photograph, for example.
The Experiments
“We use
Amazon’s cloud for computational power. I’ve trained the network with hundreds
of image pairs,” Kauppi explains. He shows examples of aerial photos that the
network has converted into city plans. By reversing the neural networks, the
system created aerial illustrations using city plans.
Another
CityCNN application marked buildings in a satellite image with a color. Kauppi
likens it to a five-year-old who’s given crayons and told to color all the
buildings. Reversing the action, the neural network can create satellite images
from building masses, automaticallly adding roads and streets and even parking
lots for larger buildings. It has learned what the landscape in Espoo looks
like and mimics it fairly accurately.
Kauppi gives
a live demonstration of the network’s capability. He draws rectangles in an
empty window and the machine creates a counterpart in another. When he draws a
small rectangle in the middle, the machine treats it as a house in the middle
of a tree-covered area. A long narrow line ends up being a street with houses
on both sides. When Kauppi draws large building masses, the network sees them
as an industrial complex or warehouse and creates adjoining large parking areas
automatically.
All the
visualizations are algorithmic and based on mathematics. The machine does not
“understand” the context. However, its behavior looks disturbingly human.
Practical Applications
An
application that may have a practical use right away is a tool that identifies
areas for potential supplementary development. Kauppi had taken aerial photos
of Espoo’s residential areas and used a paint program to mark spaces he deemed
suitable for infill. Using 500 image pairs as training material, he taught the
network to do the same to any satellite image. This way, the machine could
quickly spot all the potential areas for supplementary development.
Intelligently, it did not flag parks or woods.
“If we gave
5,000 image pairs to experts and had them mark meaningful things on the images,
the network would learn how to do the same on a national level,” Kauppi
envisions. “That sounds like a lot of work, but it would take about 10 days and
a few thousand euros for the computing, which is reasonable. After the initial
training, the network can generate new images in milliseconds.”
If you
combine image data with other types of urban data, new applications emerge.
Kauppi mentions Mapita’s Maptionnaire app, which allows citizens to give
locational feedback on their urban experiences. If people tag certain areas as
being unsafe or pleasant, a machine learning algorithm could automatically
locate other similar places to help with city plans of the future.
“Now that
we’ve completed this experiment satisfactorily, we’ll report the results and
share our experiences openly,” says Kauppi. “We’re happy to discuss how to
develop these ideas further.”
You can email Antti Kauppi at
.
Responses