Friday, November 1, 2024

Enhance!

 

AI, Satellite Images, Sci-Fi, and Magic

I remember watching Blade Runner as a kid—the original from 1982, when we were using DOS on 386 computers with CRT monitors. I was particularly fascinated by the scene where Harrison Ford mumbles instructions to a computer that seems to understand everything perfectly. What really baffled me was the "enhance" command. From a low-quality scan of a paper photo, the computer extracts absurdly detailed parts of the image through a series of panning, zooming, and "enhance" commands. I vividly thought, "It's impossible for the computer to retrieve information that isn’t there." This "enhance" trick defies the laws of information theory, thermodynamics, or whatever other principle there might be —it's just cinematic magic.

Fast forward to 2024. I recently stumbled upon a post about how AI models have been used to "enhance" low-resolution satellite images. This led me to the Satlas project at the Allen Institute for Artificial Intelligence. I was particularly excited by the "Super Resolution" model, which takes Sentinel-2 satellite images (those used in GeoFS) with a maximum resolution of 10 meters per pixel and refines them to 1 meter per pixel using some clever AI techniques.

Sentinel-2 (left) and Super Resolution (right)

Generative AI can be a tricky process; it may sometimes hallucinate or create strange artifacts. But, as Arthur C. Clarke put it, the results can be "indistinguishable from magic." The information doesn’t come from nowhere, though. The model was trained using temporal series data (Sentinel-2 images are updated regularly and cover many years) and freely available high-resolution images (such as NAIP).

Hallucinating AI

The resulting dataset was made available to GeoFS by the Allen Institute's "Super Resolution" team, led by Piper Wolters, whom I want to thank for helping to get this data ready for GeoFS. Her team, which includes Favyen Bastani and Ani Kembhavi, has published a paper about their research, which you can read here.

This new data will replace the aging, blurry Sentinel-2 images in GeoFS. While this is still an experimental phase and availability may change, you can currently enjoy GeoFS in "SR" Super Resolution, worldwide and for free. Just say: "Enhance!"


This post was enhanced by AI