Predicting urban sprawl

Can a model learn how cities grow? A neural network trained on 15 years of satellite images of Cairo, Dubai and Wuhan, then asked what comes next.

Role: Data processing and modeling, in a team of four
Product: Python model
Timeline: Fall 2016
Links: Code on GitHub (external link)

Satellite images of three cities over time with built and unbuilt areas and predictions — **Fig. 1**Cairo, Dubai and Wuhan: built and unbuilt land over time, and the model’s prediction.

01Context

Cairo, Dubai and Wuhan grew in very different ways. Running one model on all three tested whether urban growth follows patterns a machine can learn, regardless of the city.

02Approach

Satellite images from 2000, 2005, 2010 and 2015 were turned into figure-ground maps, cut into 32 × 32 pixel samples, and used to train a convolutional neural network to tell urban from non-urban land, then predict which pixels urbanize next.

Diagram of machine data sample preparation: satellite image to figure-ground to cropped samples — **Fig. 2**Machine data sample preparation: satellite image, figure-ground, then cropped samples.

Diagram of image processing and training — **Fig. 3**From satellite image to training samples.

Animated satellite view of a city over time — **Fig. 4**One city across the years.

03Outcome

Real images exist for every year, so each prediction could be checked against what actually happened. Made with Marwah Garib, Vrinda Sharma and Kun Qian for Danil Nagy’s Data Mining the City.

Next project

wastespace

Say hello.

Working on something for the built environment? I’d love to hear about it.

LinkedIn Resume (PDF)