Identification of Frost in Martian Hirise Images

Abstract

Understanding Martian landscapes is crucial for unraveling the mysteries of Mars' seasonal frost cycle and its implications on the planet's climate over the last 2 billion years. Discriminating images depicting frost or background in Martian territory are crucial to identifying low latitude frosted microclimates, offering valuable insights into climate patterns and surface evolution. Accurate classification of these images is paramount in understanding Martian climate dynamics. This work proposes a comprehensive system to build classifiers that distinguish images of Martian terrain from frost. The dataset, sourced from the Jet Propulsion Laboratory (JPL Open Repository), California Institute of Technology, NASA, provides a rich repository of Martian images, forming the foundation for our classification endeavors. Neural networks are adept at capturing complex patterns and subtle features within image data, making them well-suited for discerning the nuances associated with Martian terrain. The proposed CNN+MLP architecture employs convolutional layers for feature extraction, capturing spatial hierarchies in the images, followed by MLP layers for further abstraction and classification.