Improving Predictive Modeling for Robotic Resource Mapping with Non-Stationary Process Effects and Sensor Noise Estimation

Abstract

Accurate modeling of the distribution of water ice on the moon is crucial for in situ resource utilization (ISRU) and is the focus of remote exploration missions such as NASA’s VIPER rover. Current knowledge about the distribution of ice is limited, so probabilistic models, such as Gaussian process regression, should be used to provide uncertainty estimates along with predictions at unobserved locations. Analytic knowledge of relevant physical processes may also be capable of providing information about the presence and abundance of ice. Probabilistic models that incorporate physical processes have been shown to produce improved estimates of spatially distributed phenomena with a small number of observations; however, such models often fail to account for location-specific process effects.
Our approach advances prior work to estimate sensor observations from a neutron spectrometer as a function of modeled physical processes using Gaussian process regression. Unlike previous works, we allow the model’s parameters to vary locally such that the importance of the process variables differs by spatial location. In addition, we incorporate improved sensor noise estimation based on knowledge of how observations are accumulated over time. These two modifications taken together result in a 10-12% decrease in error and 96% decrease in average variance compared to the base model. Such improved predictions can enable faster and more targeted approaches for ISRU, while knowledge of where process variables are important can contribute to understanding the link between these processes and the abundance of water ice.