This is a very cool visualization, and I will continue to play around and explore various parts of the country. One thing I noticed immediately, due to where I live, was an interesting consequence of the algorithm. According to the model, there are several very large buildings nearby (see the outskirts of Forest Grove in the picture).Screen Shot 2018-10-12 at 5.30.04 PM

There is one cluster in the lower left (near Ritchey Rd) and another in the upper right near Schefflin. These are not really buildings, they are hoop-houses. Structures made of only plastic sheets and a few metal pipes. You could argue this is nit-picking, and I’m fine with that characterization. But I’d love to press the issue and ask what would it take to train a neural network better… so it knows the difference between what you and I would call a building, and what merely looks like a building from space.

To be more clear, here is a picture of the Ritchey Rd area:

And a full-zoom of the satellite image, corresponding to the top left (northwest) portion of the nursery (the NYT article doesn’t zoom in from here):

Screen Shot 2018-10-12 at 5.37.37 PM

Clearly these are structures, but what would be the next step in developing the algorithm to know they aren’t actually buildings (or do you define them as such?). Certainly including some nurseries in the training set would be a first step.


2 thoughts on “Buildings?

  1. There may be many unused structures in the dataset. One idea that comes to mind is to use night infrared or visible light imagery to identify inhabited structures. I’m guessing a greenhouse could have both of those properties, although the lighting of a greenhouse is often different than the natural lighting we use (for example the ones at OSU are lit purple).

    Lately I’ve been curious if it would be possible to find good places for skateboarding using aerial imagery. I have had a little luck visually using google maps but I would like to try using GIS to see if it finds better places. This would involve finding the large flat concrete or asphalt areas nearby. One step further would be the quality of the pavement (smoother is better) and scenery (more trees is nice, and less traffic).

    • Multi-spectral imaging would certainly be a good tool for categorizing structures. I wonder if it would just help to add some nursery images to the training set. It could be a case where the initial algorithm never has to distinguish such structures.


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s