A Domain-Specific Vision Foundation Model for Mars: Self-Supervised Learning for Planetary-Scale Science Discovery

Planetary missions return far more images than scientists can examine by hand. Modern AI can learn general image features, but models trained on everyday Internet photos often do not transfer well to the textures and lighting conditions of Mars. We trained a new foundation model directly on millions of Mars Reconnaissance Orbiter images, teaching it to recognize patterns in Martian terrain without using human labels. When adapted to tasks such as classifying surface types, detecting small features, and mapping crater outlines, the model performs better than widely used Earth trained alternatives. The same model can also retrieve visually similar scenes from a planet wide image archive in seconds, helping researchers quickly find geomorphic analogs and study where specific landforms occur on Mars.