Geospatial AI: Blind-Testing LandTrace

When you build a system designed to detect physical changes in land parcels using satellite embeddings, the hardest part isn't the AI—it's proving that the AI isn't hallucinating. Shadows change. Seasons change. Cameras get replaced. We needed to prove that LandTrace only flags real physical changes.

The Setup

We fed the engine random coordinates across the US, asking it for year-over-year deltas. Hidden among those random coordinates were two specific, known events: the devastating Camp Fire in Paradise, California, and the rapid construction of the Tesla Gigafactory in Texas.

The engine had no prior knowledge of these events. It only saw raw satellite embeddings and federal disaster records.

The Results

It caught both.

For the Camp Fire, the embedding distance spiked violently in the correct year. The system automatically cross-referenced the coordinates with federal fire databases and generated a report confirming massive vegetation loss and structural destruction.

For the Gigafactory, it tracked the transition from empty dirt to a massive industrial footprint, mapping the exact quarters where the heaviest construction occurred.

This is why LandTrace works. It doesn't just look for visual differences; it understands semantic, structural change and verifies it against ground-truth data.