The nice folks at the New York Times approached us with an intriguing question: could we use satellite imagery to map how the United States has changed over the past decade?
We focused our efforts on one aspect of this broad question: could we find new construction? From a remote sensing standpoint, construction can be thought of as a pixel that contains an impervious surface.
To detect impervious surfaces, we trained a UNet segmentation model on Landsat imagery. For training data, we used the urban imperviousness dataset from 2016 from the National Land Cover Database (NLCD).
To train the model, we generated image/target pairs for segmentation. The imagery was constructed from a cloud-free median composite of Landsat 8 imagery over the continental United States. The corresopnding targets for each image were constructed from the NLCD impervious surface dataset, converted to a binary image. We utilized six bands (red, green, blue, nir, swir1, swir2) from Landsat imagery. Once trained, the model given an input image generates a probability map of every pixel belonging to an impervious surface class at 30 meter resolution.
The model was deployed over two composites: a cloud-free median Landsat 5 composite from 2008/2009, and a cloud-free median composite Landsat 8 composite from 2018/2019. The outputs from both composites were then differenced to ultimately produce the map above: a map of all newly constructed surfaces in the United States over the past decade.
Check out the article for more details and fun anecdotes that we found along the way.