Map data that supports the nation, made faster and more reliable with AI.

Geotrans, Inc. builds AI that analyzes aerial, satellite, and drone imagery to produce the spatial data that government agencies and infrastructure companies need. Its mission is to support the world's infrastructure with advanced technology. Tasks that once relied on human interpretation, such as detecting building changes for property-tax management, classifying vegetation, and analyzing forests, are now handled with AI.

The company's models analyze aerial photographs and orthoimages to extract features such as buildings and vegetation. Since 2026, ANOSUPO has supported Geotrans with reviewing inference results and creating training data from aerial imagery.

Producing a client-ready deliverable from AI inference results means checking for false and missed detections and making the necessary corrections. At Geotrans, the AI engineers handled this step themselves.

In a project covering hundreds of thousands of buildings, for example, the team had to compare imagery from two time periods against the AI's output and add or remove polygons as needed. Because the required output format differs from project to project, forecasting the review workload and securing enough people for it was a constant challenge.

Consistency was another issue. In interpretation work, judgments vary from person to person, so aligning review criteria matters. Geotrans decided that engineers should focus on development and inference, and that the review and correction needed to finish deliverables could draw on outside help.

The first request was to review the inference results of a building-detection model. Geotrans provided a complete QGIS project and the aerial imagery, and our annotators visually searched for buildings the AI had missed and marked each with a point.

The next request was to create training data. From aerial photographs of several designated areas, we traced the specified features as polygons and returned them as training data.

Every annotator's output receives a final visual check against the agreed quality criteria before being delivered in batches. Transferring several gigabytes of imagery and confirming it displayed correctly in QGIS took time on the first project, but once the environment was set up, Geotrans was able to hand over the full sequence of tasks.

According to Mr. Sakuma, the time he spends checking deliverables has not changed dramatically. What has changed is how he checks. With shared quality criteria and a sense of what the deliverables tend to look like, he can direct his attention to the areas that need it most. When corrections are needed, he can point to specific locations and discuss them.

The outlook on workload has also shifted. Once imagery and the working environment can be handed over smoothly, several tasks can be commissioned together. Geotrans now has the option of drawing on outside capacity depending on a project's scale and output format.

Transferring multi-gigabyte datasets remains an open issue, and both companies are exploring more reliable ways to share large files. Video annotation is a possible next step in the collaboration.