Expert-graded signal for NASA JPL's Martian frost model
Problem
NASA has launched multiple probes and orbiters to study Mars, each generating large quantities of data — images of different types and resolutions, plus thermal data. JPL data scientists and engineers are using it to build a frost map of the surface, identifying water and CO2 frost and various formations. To power it, they first had to construct a training dataset their model could consume — combining multiple data types with annotations from domain experts.
Solution
With Labelbox Annotate, the team produced its training signal through a strategic, iterative approach: breaking high-resolution images into smaller parts easier to label and process, maintaining a living labeling guide with edge cases, and organizing thermal data by month and location to match it with corresponding images.
Result
The frost-map dataset was built over many iterations in Labelbox and will soon be published to the broader scientific community, where other scientists and experts can ask questions, add context, and refine the signal even more.
NASA JPL is building a model to map water and CO2 frost on Mars. Labelbox produced the multi-expert, confidence-scored signal that makes subtle frost trainable.
Note: This story is a recap based on the session, “How to build ML-ready data sets: Recent research from NASA/JPL,” at Labelbox Accelerate 2022.
The challenge
NASA's Jet Propulsion Lab is building a holistic map of frost on the Martian surface, released monthly to track the planet's freeze-thaw cycle, using data science and machine learning to understand where, how, and why water and carbon dioxide frost appear.
Just like on earth, freeze-thaw cycles can cause erosion which is why we want to look at how the Martian frost cycle is shaping the surface of Mars,” says Mark Wronkiewicz, data scientist at JPL.
The data comes from many sources — high-resolution imagery from the Mars Reconnaissance Orbiter, lower-resolution cameras, and thermal data from the Mars Climate Sounder. To train a model, JPL first had to turn all of it into a training dataset. The hard part is the signal: frost shows up in subtle, variable ways, confounded by cracks, albedo, and sublimation.
The approach
With Labelbox Annotate, JPL produced the signal with an iterative, consensus-driven approach. High-resolution images were broken into smaller squares, randomized, and labeled. Each image was labeled by three contributors, who drew a polygon identifying frost, described why they believed frost was present, and rated their own confidence. The team reviewed low-consensus images as a group to improve consistency and find the causes of low confidence.
(Pictured above) Cracks, sublimation spots, and albedo captured in image data of Mars's surface can make it difficult to accurately identify frost patterns. These images will need to be labeled and reviewed by planetary scientists with years of expertise in recognizing various terrain formations.
A living labeling guide captured edge cases as they emerged, so each new set of contributors had reliable reference material. To train the model, images were broken into 300px x 300px tiles. The team tested on a separate dataset, measured low-confidence areas, and curated new training data targeting the model's problem terrains. Then it combined the labeled image signal with Mars Climate Sounder thermal data, paired by metadata like time and location.
We're basically doing a form of model adjustment. So given the ML prediction on one axis, you can take that model confidence and then cross it with the confidence that you have that frost exists based on a second dataset,” said Wronkiewicz.
Viewing confidence on both datasets together showed where frost actually exists: matching scores were likely accurate, while diverging scores were pulled aside for domain-expert study.
The outcome
The frost-map dataset was built over many iterations in Labelbox and will be published to the broader scientific community, where other scientists can ask questions, add context, and refine it. Over the next two to three years, JPL will add more data types from other orbiters and tools to build a more reliable frost map and launch new studies.
Where this goes
Planetary science is a grounding problem: subtle, expert judgment turned into structured signal, cross-checked across instruments. The same confidence-scored signal that maps Martian frost is how you make any model trustworthy on hard cases.