Machine Learning for Planetary Science: Challenges and Opportunities
Abstract
The capabilities of machine learning (ML) and deep learning (DL), subsets of artificial intelligence (AI), have enabled new opportunities in planetary science in the past decade. DL-based computer vision has automated content-based image classification, identifying and categorizing landmarks such as craters and dunes in satellite imagery. Computer vision applied to video imagery also enables motion and anomaly detection. Its use in both mission development and mission operations is well-documented, from charting mission trajectories to prioritizing novel observations. A wide range of techniques, from clustering and support vector machines (SVMs) to random forests (RFs) and convolutional neural networks (CNNs) have been utilized for planetary science. There is also emerging research harnessing generative adversarial networks (GANs) to detect and reconstruct planets. Further, domain adaptation and transfer learning have reduced computational burdens and allowed AI applications to be utilized even when available domain-specific data is scarce. Integrating ML and physics-based modeling remains a popular research question to tackle. Outstanding challenges in the application of ML to planetary science and astronomy at large include increasing the explainability of algorithms and reducing bias in training data. An additional challenge is determining whether particular research problems are more aptly approached through traditional statistical modeling or more complex ML techniques. As the quantity of astronomical data continues to increase in volume, it is crucial that computer scientists and planetary scientists collaborate to bridge siloed efforts and advance the field in a way that is coordinated and sustainable. We also encourage increased investment in programs focused around data analysis (such as in NASA's Planetary Science Division) and the development of ML-ready software and hardware.
- Publication:
-
American Astronomical Society Meeting Abstracts
- Pub Date:
- January 2023
- Bibcode:
- 2023AAS...24110553C