Mapping the Unknown: Generative Geospatial Layers

If this article interests you, consider reaching out. We’re actively looking for ambitious researchers to join our mission. This article is for a general curious audience.

Some of the most valuable questions about the Earth have little to no data to answer them. There is no survey of buildings in Havana scored by structural safety. Nobody has surveyed the architectural style of every street in the world. Nobody has labeled which stretches of coastline are "scenic," which neighborhoods feel "well-kept," or where a flood is most likely next season. The grounded data simply doesn't exist yet, and surveying is currently expensive, slow, and limited.

This is where our proprietary Generative Geospatial Layers “Smart layers” come in. First, we want to define what we mean by generative geospatial layers because the phrase gets used loosely. There's a thinner version of "generative maps" that some products ship today, a wrapper around a language model that reorganizes existing points of interest into a nicely arranged map. It synthesizes a new arrangement of things that were already labeled.

Columbus Generative Geospatial Layers

Columbus' Smart layers are different. Our generative layers synthesize new data and information, not just rearrange POIs, but produce something that wasn't in the dataset at all. A new dataset such as a predicted heat map, a future-pattern layer, or an extrapolated estimate for a region nobody has surveyed. For example, a Smart Layer in combination with the recipes the model has learned has the ability to reason about how a species' habitat might shift after a mining project. The endpoint isn't a prettier map, it's a new map of something that was not visible before. This direction has research precedent supporting it (Jakubik et al. (2025), TerraMind: Large-Scale Generative Multimodality for Earth Observation).

The Benefits of generative Smart Layers

Used well, Smart Layers let us answer questions that have no survey behind them and enable new creativity in cartography. They can also surface patterns that raw data implied, but no human had inferred. They’re also useful because they're produced in minutes rather than the months a field survey takes (see Columbus Pro Smart Layers in use).

The Technology Behind It

Mechanically, this rests on a hard distinction: ground-truth attributes versus generative attributes.

• A ground-truth attribute is something we know is true at a coordinate. There's a metro stop here, operated by this transit authority; the building is 30 meters tall; the survey records the formation as porphyry.
• A generative attribute is a predicted value the model produces: the likelihood that this coordinate is "scenic," that this block has "ancient architecture," that this area is rich in some mineral. We assign each a score from 0 to 1, and critically, label the prediction as synthetic.

Generative Layer Safety

A known failure mode in generative AI is model collapse. A potential failure mode when a model is recursively trained on synthetically generated datapoints. At Columbus, ground truth is among our most important commitments. That’s why we have vetting systems in place, as well as internal validation auditsInternal validation auditsMethodology for internal validation audits are always developing. Currently, we perform cross-validation against proxy variables as well as probabilistic consistency checks, to assess whether the model's predictions are statistically meaningful. In addition, we routinely conduct proprietary ground surveying to validate data with certainty. for each dataset that is used in reasoning or sold to customers.

Additionally, Smart Layers are clearly labeled as ‘generated data’, and synthetic predictions we generate are stored in a separate repository. They are presented to the user, but are never fed back into the model for training unless they have been validated. With transparent labeling, confidence scoring, cross-validation, and proper UX disclosures within our product, we believe Generative Layers can be safely introduced into the world of Geospatial Intelligence through our business and consumer products.

Generative Layers are a new capability we’re building in Magellan (our model) for our products that our supporters and we are excited about.

If you’re interested in joining the voyage to develop this technology, come on board.