Probabilistic Surface Classification for Rover Instrument Targeting

Abstract

Communication blackouts and latency are significant bottlenecks for planetary surface exploration; rovers cannot typically communicate during long traverses, so human operators cannot respond to unanticipated science targets discovered along the route. Targeted data collection by point spectrometers or high-resolution imagery requires precise aim, so it typically happens under human supervision during the start of each command cycle, directed at known targets in the local field of view. Spacecraft can overcome this limitation using onboard science data analysis to perform autonomous instrument targeting. Two critical target selection capabilities are the ability to target priority features of a known geologic class, and the ability to target anomalous surfaces that are unlike anything seen before. This work addresses both challenges using probabilistic surface classification in traverse images. We first describe a method for targeting known classes in the presence of high measurement cost that is typical for power- and time-constrained rover operations. We demonstrate a Bayesian approach that abstains from uncertain classifications to significantly improve the precision of geologic surface classifications. Our results show a significant increase in classification performance, including a seven-fold decrease in misclassification rate for our random forest classifier. We then take advantage of these classifications and learned scene context in order to train a semi-supervised novelty detector. Operators can train the novelty detection to ignore known content from previous scenes, a critical requirement for multi-day rover operations. By making use of prior scene knowledge we find nearly a 100 percent increase in the number of abnormal features detected over comparable algorithms. We evaluate both of these techniques on a set of images acquired during field expeditions in the Mojave Desert.