Automated Geological Landmarks Detection on Mars Using Deep Domain Adaptation From Lunar High-Resolution Satellite Images

Abstract

The diversity in geological characteristics on the planetary surface, such as distribution (density), size, shapes, floor structures, ages, and availability of various input data types such as optical, thermal images, and digital elevation maps pose numerous challenges for detecting geological landmarks (e.g., rockfalls, craters, etc.). Several automatic detection methods are proposed to identify geological landmarks. However, the insufficiency of the labeled dataset is a challenging problem. It requires exceedingly time-consuming and expensive manual annotation. In this article, we use the domain adaptation technique to transfer deep learning from the planetary surface to another (lunar surface into Martian surface). We test the feasibility of transfer learning of the convolutional neural networks in optical images and elevation maps to distinguish landmarks such as rockfalls and craters from the background. The experimental results demonstrate the effectiveness of the proposed method. It achieves high F1-scores compared to the state-of-the-art methods with 58.32± 2.3 and 57.51± 2.4 in detecting rockfall regions in optical lunar and Martian images. It also achieves 65.32± 1.8, 67.39± 2.4, 77.37± 2.2, and 72.56± 2.3 in detecting crater regions in optical images and digital elevation maps of Moon and Mars. This method can be a potential approach to identify landmarks for coming Mars missions.