Attempting to Detect Construction from Satellite Imagery

Attempting to Detect Construction from Satellite Imagery

Satellite

Landsat 8 Satellite Image of the Washington, DC area courtesy of USGS EarthExplorer.

In my time as a research assistant, Code for DC'er, graduate student in Public Policy, and newly minted Oakland Brigade member, I have been frustrated with the lack of widespread, easy access to construction permit data, which can tell us a lot about how neighborhoods are changing now. Without these data, we are left to rely on Census data for neighborhoods, which is often very dated.

So, with more computing power and tools available, I wanted to see if it was possible to use satellite imagery to get such information and sidestep the government. Though I failed to find a model that worked, I post my methods and findings here in case someone comes along with a method that can make some progress. If you'd like data or code, feel free to contact me on twitter at @GrahamIMac.

Abstract

Policymakers looking to implement place-based policymaking often struggle to find up-to-date data with which to use to make sure their money is targeted to the appropriate, “needy” areas. One measure that might indicate a neighborhood is changing rapidly is new construction. In most areas, the only information available on new construction is residential permit data, which is provided only at the local level, and very rarely provided in electronic form to the public for easy analysis.

Satellite imagery from the USGS Landsat 8 program is made freely available to the public, and the satellite surveys the entire earth once every 16 days. This study examines Landsat 8 data over Washington, DC and Montgomery County in 2013 and 2014 to see if satellite imagery might be used to detect new residential construction.

The study uses 16 indicators in 4 models – logistic regression on single pixels and pixel averages, and random forest classification on single pixels and pixel averages – to try and predict new construction – as determined by permit data – from satellite imagery reflectance values. The random forest classification model also includes all possible second-order interaction terms. Unfortunately, the indicators and models used in this study were unable to predict new construction any better than the baseline assumption that all areas do not have new construction occurring.

It is possible that better indicators, better models, or integrating data from outside sources might improve our ability to predict new residential construction from satellite imagery, but these assumptions were not tested in this paper. Until a better model is found, politicians and advocates should focus their efforts on making local residential permit data more easily accessible.

Download the full pdf here.