AI vision models for geolocation: capabilities and limits

Block

11 · AI as Practitioner Tool

Topic

AI Geolocation

Tools

GPT-6 Astra · Gemini 3.1 Pro · SunCalc · Google Earth Pro

Difficulty

Intermediate

Prerequisites

Capstone: verifying crisis footage end-to-end.

Vision models accelerate geolocation triage but cannot replace primary verification

Large vision models can identify architectural styles, vegetation zones, road markings and signage across hundreds of frames in the time it takes an analyst to examine one. That speed is operationally useful. The output is not evidence.

A vision model presented with a photograph returns candidate locations based on pattern matching against its training distribution, with no access to ground truth; it will confabulate a plausible-sounding location with the same surface confidence it uses to report a correct one. The discipline this tutorial builds is constraint: a two-stage workflow where the model generates and ranks candidate locations and surfaces visual features worth pursuing, then SunCalc, Google Earth Pro and open-source corroboration confirm or eliminate each candidate through primary-source methods.

In the field

On 4 April 2022 the New York Times published a satellite and video analysis of Bucha, Ukraine, rebutting Russian claims that civilian deaths there occurred after Russian forces withdrew. The Times matched the exact street positions of bodies visible in ground video filmed on 1 and 2 April against Maxar satellite imagery dated as early as 9 March, when Russian forces still controlled the town. The BBC independently reproduced the finding using separate Planet Labs and Apollo Mapping imagery, and AFP photographers confirmed the scene on the ground.

Positional matching. Objects matching the size and position of human bodies, identified in satellite imagery dated 9 to 19 March, were matched to the exact street positions of bodies filmed in ground video from 1 to 2 April.
Independent satellite corroboration. The BBC separately compared the same street against Planet Labs and Apollo Mapping imagery, reproducing the Times' finding using a different imagery provider.
Ground-truth confirmation. AFP photographers entered Bucha directly and confirmed the presence of the bodies described in both the satellite and video analysis.

The New York Times, "Satellite Images Show Bodies Lay in Bucha for Weeks, Despite Russian Claims" · corroborated by BBC Reality Check · April 2022

Learning outcomes

By the end of this tutorial you will be able to:

Submit imagery to a large vision model with a structured prompt that extracts candidate location features without leading the model toward a specific answer
Evaluate vision model geolocation candidates against the LST-001 confidence tiers and identify which outputs qualify as Possible versus which require additional corroboration to reach Probable
Apply primary-source verification methods (SunCalc, Google Earth Pro, satellite imagery) to confirm or eliminate model-generated candidates
Identify the six failure modes in which vision models produce confidently stated but incorrect geolocation outputs
Document a vision-model-assisted geolocation workflow to evidentiary standard, distinguishing model output from verified finding

The rest of this tutorial is for Signal subscribers.

What remains: the decision framework, the tool configuration, the failure modes, and the evidentiary standard required to use the finding defensibly. Signal is €90 a year, or €9 a month. Students, €49 a year.

Join Signal

A Signal subscription gives you:

Full OSINT Reference Card library, 21 domains
Methods, every tradecraft tutorial in full
AI in OSINT, every prompt and field report
Shadow Analysis, every forensic dossier

AI vision models for geolocation: use and limits

Vision models accelerate geolocation triage but cannot replace primary verification

Learning outcomes

The rest of this tutorial is for Signal subscribers.

A Signal subscription gives you:

Keep Reading