In 2019 I flew to Albania and made a road trip through the country. One of the first things I noticed was the staggering number of tiny concrete bunkers that are literally everywhere. I googled it and found out that they were built during the rule of Enver Hoxha from the 1960s to 1980s. He feared that Albania would be invaded by the NATO or Warsaw Pact and decided it would be a good idea to build 750,000 bunkers for the 3 million people country.

Today, they are decaying and serve as spots for dark tourism. It is too expensive to dismantle them. Some are used as stables, cafés or even tattoo studios. It’s unclear how many of those bunkers were eventually completed. Estimates range from 173,00 to 745,000.

I figured it would be a cool project to use deep learning on satellite images to get an estimate. There are a lot of research projects dealing with the extraction of information from satellite images. E.g., EuroSAT is a dataset consisting of 27,000 Sentinel-2 images that are grouped into 10 classes. It is used for land use and land cover classification. While Sentinel-2 data is publicly available and free it just offers a maximum spatial resolution of 10 meters, i.e. an area of 10 by 10 meters is represented by one pixel. This is insufficient as the smallest bunker type has a diameter of just 3 meters. There are commercial satellites that offer sub-meter resolutions but are expensive for the scope of the project. For scientific research projects ESA offers access to DigitalGlobe WorldView-3 data that offers resolutions up to 31 cm. The downside is that one needs to submit a full-fledged project proposal that doesn’t seem to be suited for hobby projects like this.

Update 2021-02-08: A couple of days ago, I read on HN about Albedo - a startup that promisses inexpensive satellite image data with 10 cm resolution. I’m highly anticipating this one.

The next step would be to manually create a dataset and start counting, utilizing object detection algorithms such as {Fast, Faster, Mask} R-CNN or YOLOv3.