Deep Active Learning with Concept Drifts for Detection of Mercury’s Bow Shock and Magnetopause Crossings

Abstract

Active learning has shown great potential for improving the efficiency of data annotation in scientific applications. In the field of planetary science, where large volumes of data are generated by remote sensing instruments, active learning can significantly reduce the cost and time required for data analysis. However, existing active learning (AL) methods for planetary science applications do not consider the potential for drifts in the data distribution, which can lead to model degradation over time. To address this issue, we propose a drift detection-based active learning approach for planetary science applications. The proposed approach uses a semi-supervised generative adversarial network (GAN) to detect concept drifts and an entropy-based AL sampling procedure to select the most informative orbits from each for training. We test this approach on the use case of detecting bow shock and magnetopause boundaries around Mercury, utilising data obtained from NASA’s MESSENGER mission. Our key results indicate that by employing our approach, a near-maximal information gain can be obtained by training with less than 10% of the available data, surpassing the simpler entropy-based active learning. This approach has the potential to accelerate scientific discoveries in planetary science and other scientific domains that deal with large volumes of remote sensing data.