Abstract Planetary missions have produced a rapidly growing archive of high‐resolution imagery, yet only a small fraction has been examined in detail. Foundation models (FMs) offer a scalable path by learning transferable representations from unlabeled data, but most vision backbones are trained on Internet imagery or Earth observation data sets that differ substantially from planetary surfaces. We present a CTX‐specific Mars vision FM based on Vision Transformer pretrained with self‐supervised learning on millions of Mars Reconnaissance Orbiter (MRO) Context Camera (CTX) images. Across linear evaluation and fine‐tuning for classification, detection, and segmentation, the CTX‐pretrained encoder consistently outperforms strong baseline FMs. The same embeddings enable rapid similarity search across millions of CTX images, reducing the manual effort required to locate geomorphic features and accelerating exploratory mapping at planetary scale.
Fang et al. (Wed,) studied this question.