Automatic Experimental Design Using Deep Generative Models of Orbital Data

Abstract

The mapping and characterization of planetary surfaces relies on the analysis of data collected by spacecraft and orbiters. Their instruments provide extensive contextual information, but factors such as sparsity, resolution, and noise leave uncertainty in the orbital analysis. Hence the need to send robotic explorers to refine these models through the collection of definitive, in situ measurements. Since planetary rovers face many operational challenges and constraints, it is important to identify sampling locations that maximize information value. This paper describes a deep generative method that learns a probabilistic model relating remote and in situ data, which then allows formal experimental design and measurement planning using tools from information theory. We apply this method to spectroscopic observations of the Cuprite Hills in Nevada. The results indicate that our model is capable of inferring high resolution features from orbital data, and that it also identifies effective in situ sampling locations.