Optimal U-Net Architecture for Object Recognition Problems in Multispectral Satellite Imagery

Geospatial data follows Moore's law. On the back of improvements of optical Earth observation satellite hardware [1] (weight, propulsion systems, signal transmission and resolution) as well as reduced costs of rocket launch carrying these satellites, number of nanosatellites deployed to Lower Earth Orbit (LEO) in 2018 was larger than in the previous 10 years combined [2]. This allowed an exponential growth in satellite imagery data production that is available for military and commercial use. Machine learning tools enable us to process high resolution, multi-spectral satellite imagery data to recognize objects at scale [3] and generate insights with practical industry applications. It allowa us to calculate global oil reserves, track tanker ships or estimate retail revenue based on the car count, all exceptionally valuable financial information. In this paper, we investigate various computer vision techniques to develop an optimal machine learning technique for object recognition problems at this unique type of the dataset: multi-spectral satellite imagery.