Solution · Computer Vision & Inspection

Aerial Segmentation & Measurement

DataWise built and demonstrated a SAM-based system that segments airfields in satellite and aerial imagery and measures them, with human correction and reports.

Updated:

How do you segment and measure features in satellite or aerial imagery without weeks of manual digitizing? By letting a foundation model do the tracing and keeping a professional in charge of the result. DataWise built and demonstrated a SAM-based (Segment Anything) system that automatically segments airfields in satellite and aerial imagery and measures them with precision tooling, wrapped in a purpose-built UI with manual correction and editing, object addition, and report generation. It turns the costliest phase of geospatial vision work into a supervised, minutes-long task.

How it works

The pipeline runs in four stages, and the human sits inside it, not after it.

Automatic segmentation. The system builds on SAM, a vision foundation model with strong general segmentation ability, adapted to the target feature class, in our demonstrated case, airfields, so boundaries that would take an analyst hours to digitize are proposed in minutes.

Precision measurement. Segments are not pictures, they are geometry: the tooling measures the segmented features precisely, which is what downstream engineering, planning, or assessment work actually consumes.

Correction as a first-class feature. A segmentation that is 95% right and uneditable is useless to a professional who signs off on the result, so the purpose-built UI supports manual correction and editing, adding missed objects, and adjusting boundaries. The model does the heavy lifting; the person corrects and approves.

Report generation. The output is a report your organization can act on, not a folder of masks.

This shape, automation plus supervision, is deliberate. In geospatial work the expensive part is not detection, it is trust, and the tooling assumes from the start that the model will sometimes be wrong.

What we’ve built

  • The system described above: built and demonstrated, on the airfield segmentation and measurement task, end to end from imagery to corrected, measured report.
  • Vision inside closed networks: architected and specified. For a security-sensitive organization, DataWise architected vision object-detection as part of a fully isolated AI enclave, everything running inside the closed network. Imagery that cannot leave the building is a deployment constraint we design for, not an exception.

From the wider industry, the exact pattern DataWise built is now both government-scale and academically validated. The US National Geospatial-Intelligence Agency runs a five-year, $290 million program (Maxar’s Luno A) to detect and measure vehicles, aircraft, vessels, and infrastructure across satellite imagery, per reporting on the contract. And foundation-model segmentation now works at scale without task-specific training: researchers ran Segment Anything on free Sentinel-2 imagery to delineate farm-field boundaries across 32 million hectares at 0.86 IoU against manual mapping, per a peer-reviewed study. Closer to the ground, ASTERRA’s satellite-based analysis found 50 water leaks for the town of North Kingstown, Rhode Island, worth about $204,807 per year with a 10.4-month payback, per the town’s own published report. Imagery analytics is delivering audited, municipal-grade ROI, and that is the standard aerial AI should be held to.

FAQ

Our features aren't airfields. Does this transfer?

The demonstrated system targets airfields; the architecture, a segmentation foundation model plus measurement tooling plus a correction UI, is built to be adapted to other feature classes. We validate on your imagery and your feature class during feasibility before any build commitment, and if your imagery cannot support the accuracy your decision needs, we say so at that stage.

How much labeled data do we need to start?

Often far less than classical vision projects, because SAM-class models start from strong general segmentation ability. Feasibility on your imagery tells you how much labeling your specific task actually needs before you spend on it.

Our imagery is classified. Can this run disconnected?

Yes, that constraint is familiar to us: we architected vision workloads as part of a zero-egress enclave for a security-sensitive organization. The deployment follows your data's rules.