Zero-Shot Martian Landform Classification via Vision–Language Alignment in Foundation Models

Abstract

Martian geomorphological interpretation is fundamental to understanding the geological evolution, environmental changes, and potential habitability of Mars. However, existing automated interpretation methods heavily rely on large-scale annotated data, making them difficult to apply in Martian remote sensing scenarios characterized by limited labeled samples and challenges in recognizing unknown landform types. To address these issues, this letter explores the potential of vision foundation models for planetary remote sensing. Specifically, a pretrained Mars-ViT is employed as the visual encoder, while a CLIP text encoder is used to extract semantic representations of categories. A lightweight visual adapter and a text feature alignment module (TFAM) jointly align visual representations with the textual semantic space. In addition, an InfoNCE loss with class-prototype alignment is incorporated to enhance cross-modal discriminability. Experiments on the DoMars16k dataset demonstrate that the proposed framework achieves strong zero-shot classification performance under various feature alignment strategies, with the vision–text adaptation strategy yielding the best results. These findings demonstrate the effectiveness of the proposed approach without relying on large-scale, high-quality image–text paired data.