Google researchers pointed 48 AI agents at NASA satellite imagery and found a…

The convergence of satellite imagery, machine learning, and archaeology represents a fundamental shift in how humans can discover and document the physical world. Understanding this intersection requires exploring how each component works independently, what happens when they combine, and why this matters beyond any single discovery. Remote sensing from space Satellites have orbited Earth for decades collecting visual data across multiple wavelengths of light, including wavelengths invisible to the human eye. NASA and other space agencies maintain extensive libraries of this imagery covering most of the planet's surface. The raw material exists in massive, freely available archives. A single satellite image might cover hundreds of square kilometers at varying resolutions. Human analysts examining these datasets manually would require months or years to process what modern computing systems can analyze in hours or days. The sheer volume of data creates a fundamental problem: humans cannot scale their attention across such vast areas efficiently. What lidar actually measures Lidar is a remote sensing technology that uses laser pulses to measure distances to Earth's surface. The system fires pulses and times how long they take to return after bouncing off terrain and vegetation. By doing this millions of times per second, lidar creates precise three dimensional maps of landscape features. Critically, lidar can penetrate dense vegetation. Where traditional optical imagery shows only the green canopy of a jungle, lidar reveals the ground beneath. Archaeological features like mounds, platforms, and altered terrain that might be invisible under thick forest cover become visible in lidar data. This capability fundamentally changed archaeological prospecting in challenging environments like Central America, where dense jungle has historically hidden human structures. Machine learning pattern recognition Machine learning systems are trained on labeled examples to recognize patterns. If you show a system thousands of images of known archaeological features and tell it what a pyramid or ceremonial platform looks like in lidar data, the system learns to recognize similar patterns in new imagery. This is fundamentally different from hand coded rules. The system doesn't have a pre programmed definition of a pyramid. Instead it develops mathematical representations of features that humans have marked as important. Multiple agents can run in parallel, each analyzing different sections of a dataset simultaneously. The process is probabilistic rather than certain. The system generates confidence scores about whether a particular pattern matches the features it learned to recognize. The scalability transformation Traditional archaeology relies on field surveys, historical documents, and accumulated local knowledge to locate sites worth excavating. These methods are necessarily limited in scope. A team might spend years surveying a specific region. Remote sen