18 October 2019
A satellite image of Damascus, Syria. Artificial intelligence is being used to analyze the bird's eye view of satellite pictures. Photo: Reuters
A satellite image of Damascus, Syria. Artificial intelligence is being used to analyze the bird's eye view of satellite pictures. Photo: Reuters

GANs open up enormous possibilities

“A falling leaf reveals the coming of fall,” says an old Chinese saying.

Humans always try to determine the overall picture on the basis of small details. However, some angles can only be achieved with the help of satellites, high-resolution video recorders, and drones.

In the past, under the guidance of feng shui masters, I scaled mountains and swam through the waters to learn more about mountain ranges and water flows.

All these skills turned out to be very helpful in adjusting my point of view.

Today, researchers are trying to tackle a question that runs in the opposite direction to what “a falling leaf reveals the coming of fall” is saying. They do this by means of artificial intelligence.

For example, how do we generate a filed view of a satellite picture? Is that achievable?

The University of California, Merced is using generative adversarial networks (GANs) to analyze the bird’s eye view of satellite pictures. GANs consist of two neural networks that are pitted against each other. One algorithm generates a ground graph from the satellite picture, while the other algorithm compares a real ground picture with the mock picture, thereby aiding machine learning.

The quality of the input data is critical. Researchers at the University of California selected 16,000 pairs of bird’s eye view graphs and ground images covering 71 square kilometers of London to train the algorithms. They would put 4,000 bird’s eye view images of specific locations in the system, and let the GAN work.

The result is very interesting. The system has generated ground images from bird’s eye view graphs, and the images were able to capture basic features such as roads, a city or the countryside. However, it lacks the details of a real ground picture.

How could this technology be applied to real life? One major task of city planning is to categorize different areas for various purposes, such as industrial zone, residential area or commercial district.

The computer-generated imagery (CGI) technology has an accuracy ratio of 73 percent when applied to land use planning, compared with the 65 percent accuracy rate of the traditional method.

GANs have offered a new and efficient solution to collecting data for land planning.

I believe if we can collect sufficient mobile data from the 71 square kilometer area, we might be able to generate images of human economic activities in the area. That’s very intriguing.

The full article appeared in the Hong Kong Economic Journal on July 10

Translation by Julie Zhu

[Chinese version 中文版]

– Contact us at [email protected]


Venture Partner of Sequoia Capital China, former head of the data committee and vice president at Alibaba Group.