Researchers build an AI that reads satellite data directly instead of converting it to text first

A new system lets large language models process dense geospatial data (satellite imagery, population movement, climate patterns encoded as mathematical vectors) as direct inputs instead of converting them to text descriptions first. This cuts out the middle step, saves computational tokens, and preserves numerical precision — meaning AI models can reason about space and place faster and more accurately.

Until now, the bottleneck was translation: satellite embeddings had to be converted to English descriptions before an AI could reason about them, which is slow and lossy. This framework removes that translation step entirely. It matters because satellite and climate data are central to urban planning, disaster response, environmental monitoring, and development — tasks where speed and accuracy directly affect decisions that affect people. If this pattern holds in production (converting dense spatial data into direct embeddings instead of text), it could reshape how AI assists in real-time geospatial intelligence across infrastructure, agriculture, and climate adaptation.