The machine learning system the team used was eventually able to come up with a list of 4,096 features it could use to look at a particular area on the map and predict its level of nighttime lights.