The Last Remote Sensing Model Built by Humans? Casting Any Earth Observation Task as Recursive Self-Improvement
The paper says EO model-building can be reduced to a harness, a verifier, and an agent improving its own outputs.
The authors argue that any verifiable Earth observation task can be framed this way, using labels, independent sensors, physical models, or repeated observation as the check. In their tests across optical, SAR, and multispectral time-series benchmarks, agent-built models beat reported state of the art. They also report retaining most of that accuracy with about 1% of the training labels. The project is marked as ongoing. ArXiv · AI/CL/LG's note
The authors argue that any verifiable Earth observation task can be framed this way, using labels, independent sensors, physical models, or repeated observation as the check. In their tests across optical, SAR, and multispectral time-series benchmarks, agent-built models beat reported state of the art. They also report retaining most of that accuracy with about 1% of the training labels. The project is marked as ongoing. ArXiv · AI/CL/LG's note
score 4