Accurate navigation is a crucial part of both robotic and crewed exploration of other worlds. For missions to the surfaces of Mars and the Moon, mission planners will require increasingly autonomous guidance systems that support fast, efficient, and agile surface route planning. These navigation systems will rely on access to accurate and regularly updated maps of surface landmarks. In a recent pilot study (Wronkiewicz et al. (2018)), we presented a technique using machine learning to autonomously map surface features (specifically, craters) using satellite images from the Mars Reconnaissance Orbiter (MRO) ConteXt Camera (CTX). Our method could generate accurate crater maps five orders of magnitude faster and at 10x better resolution than the best manual identification efforts. This result suggests we should further explore methods of autonomously generating surface feature maps as a way to reduce the amount of manual effort needed to chart other bodies in the solar system.
We propose building the tools to efficiently map surface features across both the Moon and Mars. Specifically, we will achieve two main technical objectives: (1) create an open-source, cloud-based, reusable software product that can rapidly apply ML algorithms across large image datasets (e.g., satellite images covering the Moon and Mars) and (2) generate publicly available impact crater maps encompassing the entire surface of Mars and the Moon at ten times better resolution than is provided by existing catalogs. We will open source both the software and data products to encourage future research applying ML for planetary-scale mapping. These maps, and the software used to rapidly generate them, will undergird safe and efficient planetary navigation in a manner that exhibits “higher performance and autonomy than currently possible” (NASA STTR T4.01).
The two technical objectives proposed are immediately usable by NASA. The proposed software product will help NASA leverage machine learning on the cloud to analyze planetary satellite image datasets (e.g., within the PDS). The proposed map products will provide mission planners with initial global martian and lunar maps. Using these as a foundation, NASA can better explore the mapping of additional surface features and investigate automated navigation systems for future robotic and crewed missions.
Machine learning models to detect surface features are applicable to Earth for uses such as monitoring surface changes over time (both natural and anthropogenic). Our model's ability to ingest global datasets and create maps from them could potentially also be applied to map notable features in satellite images of Earth to aid in resource exploration, urban planning, farming. and more.