Satellite systems such as Global Positioning System (GPS) used for radionavigation have revolutionized our modern society with a plethora of military and non-military applications. For decades, one specific type of satellites, denominated Earth observation satellites has provided high quality imagery of our planet, transforming our understanding of meteorology, geology, and cartography. Satellite imagery of a city provides important information such as the number of vehicles and the particular distribution of these throughout a specific area. This data is commercially and logistically relevant. For instance, determining the amount and density of vehicles could predict less-congested routes, as well as, demographics statistics. Nevertheless, due to the size of satellite imagery databases, processing these images by hand is an extremely time-consuming task. Here, we apply a machine-learning classification algorithm called K-Nearest Neighbor to identify red vehicles on satellite imagery. A training dataset was used to extract specific features of the vehicles to train a model (via cross-validation) and make predictions on nonlabelled test imagery. After several iterations defining optimal parameters, the algorithm reliably predicted the location of red vehicles on test satellite image.

See PDF File