New AI Can Answer Your Questions About Satellite Photos

Satellite images are everywhere — tracking storms, spotting illegal logging, checking whether a farmer's crops are healthy. The problem is that reading them takes trained experts and hours of scrolling. A new research paper introduces GeoNLI, a system that lets you simply ask a satellite photo a question in ordinary English — 'Is that a bridge or a dam?' — and get a plain answer back.

GeoNLI doesn't rely on a single AI. It runs a committee of eight different models, each good at something different, and lets them vote on the best answer — like asking eight experts and going with the majority. The system handles three jobs: describing what's in an image, answering questions about it, and pointing to a specific object, such as circling the only damaged roof on a street.

The team tested it on standard satellite-image datasets. It described images correctly about 82% of the time and answered questions correctly about 83% of the time. It did best on yes-or-no questions (over 90%) and questions about meaning, like 'is this area urban or rural?' (92%). Counting questions were the weak spot — just 52% — because AI still struggles to count objects accurately in a crowded photo.

Pinpointing exact objects — circling a flooded field, say — worked about 65% of the time, meaning it misses roughly one in three. That's a real limitation: you would not want to rely on it alone for insurance claims or disaster response. This is research, not an app you can download. But the direction is clear: eventually, anyone who can type a question may be able to read a satellite image like an expert.