A research team analyzing Mars surface data needed to build a frost map identifying water and CO₂ frost formations. With data coming from orbiters in multiple formats high-res images and thermal data they had to create a structured, annotated training dataset that combined expert input across data types.
Using Databrewery, the team broke large images into smaller, model-ready segments. They built a dynamic labeling guide to help non-experts contribute accurately and aligned thermal readings with imagery by time and location. This iterative process improved annotation quality while scaling across the dataset.
The frost dataset was created over several labeling cycles and is now being prepared for release to the broader scientific community where it will support future discoveries and collaborative analysis of Martian climate patterns.
For over two centuries, scientists have observed Martian frost, first through telescopes and later through spacecraft imagery. As early as the 1960s, missions to Mars revealed surface features that hinted at now-vanished glaciers. In the last decade, tools like ground-penetrating radar and orbiting thermal detectors have confirmed that both water and carbon dioxide ice lie hidden beneath the planet’s surface, shielded by dust and rock.
To deepen our understanding of how this frost persists under extreme conditions, a planetary science team is working to generate a monthly frost map of Mars. This project is designed to track freeze-thaw cycles and provide insights into how seasonal frost activity might be shaping the planet’s landscape. With the help of AI and machine learning, the team aims to pinpoint not just where frost appears but how and why it forms.
“On Mars, just like Earth, freeze-thaw cycles can contribute to surface erosion,” said Dr. Evelyn Tan, a planetary data scientist on the team. “Understanding how frost develops and fades across seasons gives us valuable clues about long-term surface changes.”
To build the frost map, researchers gathered and organized image data from multiple instruments, including the Mars Reconnaissance Orbiter. These high-resolution images, along with lower-resolution captures from additional onboard cameras, were paired with thermal data columns provided by the Mars Climate Sounder, which measures temperature from the surface to the upper atmosphere.
Given the sheer size of the images, the team split each one into smaller square tiles to make them easier to label and process. These tiles were then randomized and prepared for annotation.
Labeling proved complex. Frost doesn’t always appear in obvious ways. Subtle changes in surface brightness, cracks, sublimation marks, and terrain albedo made identification difficult. To ensure quality, each image was labeled by three separate annotators in Databrewery. Annotators drew polygons outlining suspected frost areas, explained their rationale, and rated their confidence. Images with low consensus were escalated for group review to analyze discrepancies and improve consistency.
Some formations were so subtle that even expert annotators disagreed, so the team introduced a strategy early in the project: maintaining a continuously evolving labeling guide. This guide was updated with each edge case, including visual examples and annotation standards, giving new labelers clear context and helping the full team stay aligned.
After the annotation phase, the team further segmented images into 300×300 pixel tiles to train their first model. Once trained, the model was tested on separate validation data, and its weak points were identified, often tied to specific terrains or frost signatures. The team then curated additional training data to target those model failure areas, creating an iterative loop to refine performance over time.
Combining this annotated image data with thermal measurements posed another challenge. To align the two datasets, they matched image tiles and thermal profiles by shared metadata such as the collection time and geographic location on Mars.
We're applying a kind of confidence fusion,” said Dr. Tan. “If the model’s prediction aligns with the thermal data’s signal, we can trust that frost is present. If they conflict, it tells us that we need to take a closer look.”
This dual-confidence approach is allowing the team to improve frost detection across different surface types and eventually share the dataset with the broader scientific community. The goal is to open new doors for collaboration, allowing researchers worldwide to refine, question, and build upon the foundation they've created.