Using FME + AI for ArcGIS Feature Service metadata creation
In our most recent series of the Avineon Tensing FME AI Skill Booster workshop, we decided to have a little fun and test the location awareness skills of AI with a game of GeoGuessr. GeoGuessr is an online game that takes you on a virtual journey across the globe. The game drops you in a random location on Google Street View, and you have to guess where you are by navigating the streets, observing landmarks, and using your geography knowledge. It's a fun and interactive way to test your geography skills and learn more about different places around the world.
The premise is simple: we give an AI a picture of a location somewhere in the world, and it has to guess where it is. As the discussion around the role of AI in the geospatial industry continues, I thought I would test out the spatial reasoning and simple geography skills of an LLM.
Starting out, I was sceptical about how AI would be able to identify the features in an image to pinpoint a location around the world. I thought it would rely on the EXIF metadata in the image containing the location information. Using FME, I made sure that none of this metadata was uploaded with the image. I also used an extensive prompt that provided AI with explicit rules and instructions, forcing it to rely solely on visual evidence. The AI was instructed to think like a detective, examining architectural styles, licence plates, vegetation, topography, and any other clues in the image. The goal was to ensure a pure test of its visual recognition and reasoning capabilities, not its ability to find hidden data.
Want to see how it works? Take a look at the image below. Do you know where this is?
Here's our FME Flow application in action. I am uploading the image and providing the location of the image so we can find the AI's GeoGuesser score and see how far away the AI guess is from the correct location.
The result from the app was scarily good. The AI was able to select the correct location and identify the Statue of Liberty replica at the roundabout in Colmar, France. The point is just over a kilometre away. The detailed reasoning displayed in the results shows that the AI analysed a variety of clues, from local infrastructure and architectural styles to geographical features, before settling on its conclusion. A shortlist of alternative locations was also presented, as part of its analysis was to cross-check its prediction with similar locations.
The result from the app was scarily good. The AI was able to select the correct location and identify the Statue of Liberty replica at the roundabout in Colmar, France. The point is just over a kilometre away. The detailed reasoning displayed in the results shows that the AI analysed a variety of clues, from local infrastructure and architectural styles to geographical features, before settling on its conclusion. A shortlist of alternative locations was also presented, as part of its analysis was to cross-check its prediction with similar locations.
Throughout the development of this FME Flow app, I tested a wide variety of images. I found that even with what appeared to be a completely random road, offering very few clues, the AI would still return a correct region or country. If you want to play GeoGuessr with AI or think you might be able to trick the system, have a go at our Avineon Tensing Labs.
