Natural Language-Driven Global Mapping of Martian Landforms

Technical Summary: Natural Language-Driven Global Mapping of Martian Landforms

Problem Statement
Planetary science currently faces a fundamental mismatch between human conceptual reasoning and data organization. While scientists analyze planetary surfaces using high-level semantic concepts expressed in natural language (e.g., aeolian, volcanic, fluvial processes), vast orbital image archives remain organized at the pixel level. This disconnect limits scalable, open-ended exploration. Traditional approaches rely on manual interpretation, which is labor-intensive and difficult to scale, or supervised learning, which requires large, carefully labeled datasets and often generalizes poorly across heterogeneous landforms. Consequently, the field remains "data-rich but interpretation-limited," lacking a system capable of fast, label-free retrieval of diverse landforms or generating comprehensive, quantitative, cross-comparable maps of planetary geomorphology.

Methodology
The authors introduce MarScope, a planetary-scale vision-language framework designed to align planetary images and text within a shared semantic space. The system is built on the following technical components:

• Training Data: The model was trained on over 200,000 curated image-text pairs. While focused on Mars, the dataset includes geomorphological features from the Moon, Mercury, and icy satellites to ensure cross-planet extensibility. The text data was refined using a Large Language Model (LLM) for relevance filtering, key-phrase extraction, and semantic rewriting, alongside data augmentation (synonym expansion and paraphrasing) to enhance linguistic diversity.
• Architecture: MarScope utilizes a contrastive vision-language encoder. It learns high-dimensional embeddings that capture diagnostic aspects of surface morphology by maximizing similarity between matched image-text pairs and minimizing it between mismatched pairs.
• Search Infrastructure: The system operates on a global mosaic of Context Camera (CTX) data, subdivided into overlapping tiles at two resolutions: 0.2° (default, for kilometer-scale features) and 0.02° (high-resolution, for sub-kilometer features). Approximately 130 million tiles are encoded into a shared embedding space and indexed using a Facebook AI Similarity Search (FAISS) approximate nearest-neighbor database. This process compresses the raw imagery by a factor of ~160 while preserving key semantic information.
• Query Modes: The platform supports three complementary retrieval strategies:
  1. Text-based: Queries using landform names or formation processes.
  2. Image-based: Queries using representative visual examples to find morphologically similar features without predefined labels.
  3. Multimodal: Queries combining visual exemplars with textual constraints to refine results for complex morphologies.
• Evaluation: Retrieval performance is quantified using dynamic F1 scores (Precision@K and Recall@K) against six published global catalogues (alluvial fans, glacier-like forms, landslides, pitted cones, yardangs, and dark slope streaks). A proximity-based matching criterion with a 0.5° tolerance radius is used to account for the continuous nature of geospatial coordinates.

Key Results

• Performance: MarScope achieves retrieval of Martian landforms in approximately 5 seconds with F1 scores reaching up to 0.978. Performance varies by landform type; for instance, pitted cones and yardangs are best retrieved via text queries, while alluvial fans and landslides benefit from multimodal queries. Image-only queries, despite relying on only six exemplars, demonstrate robustness in retrieving relevant matches.
• Global Mapping: The system successfully generated global distribution maps for diverse landform systems, including aeolian (dunes, yardangs, slope streaks), glacial/periglacial (pedestal craters, glacier-like forms, scalloped depressions), and volcanic/tectonic features (fossae, lava flows, wrinkle ridges).
• Process-Oriented Discovery: Beyond morphological classification, the framework enables mapping based on inferred formation processes. Queries for "ice-related," "stress-related," or "flow-related" mechanisms successfully retrieved genetically related but morphologically diverse features across the planet.
• Rare Feature Detection: The system demonstrated the ability to identify previously understudied or unnamed landforms through visual similarity search. Specific examples include the detection of doublet craters (paired impact structures) and inverted craters (sedimentary infill resistant to erosion), facilitating the construction of statistically robust inventories for features rarely documented in existing catalogues.

Significance and Claims
The paper posits that MarScope establishes a new paradigm where natural language serves as a direct interface for scientific discovery over massive geospatial datasets. Its primary significance lies in reframing geomorphological mapping from a predefined classification task into an open-ended semantic retrieval problem. This shift allows for:

1. Label-free exploration: Researchers can query planetary surfaces using flexible semantic prompts rather than rigid taxonomies.
2. Scalability: The framework enables near-instantaneous global mapping, overcoming the bottlenecks of manual interpretation and task-specific supervised learning.
3. Foundation for Specialized Models: MarScope is designed as a foundation model that can serve as a backbone for specialized models fine-tuned for specific mission goals, reducing the need for extensive custom model development for new planetary datasets.

The authors acknowledge that the system is optimized for flexible, large-scale exploration rather than pixel-level accuracy, noting limitations in representing features significantly larger or smaller than the tile sizes. However, they assert that MarScope complements detailed physical modeling and expert interpretation by providing a versatile infrastructure for hypothesis-driven exploration across planetary bodies and scientific domains.