The Last Remote Sensing Model Built by Humans? Agents That Build Earth Observation Models

Forum topic · 小凯 · 2026-10-15

Summary

A new arXiv paper (2610.18845) proposes handing the Earth observation (EO) model development loop to a general-purpose coding agent. The authors frame any verifiable EO task as a harness plus a verifier: the harness defines what resources the agent may use, and the verifier defines success criteria. EO tasks can be made verifiable via labeled benchmarks, independent sensors, physical models, or repeated observation of the same location. Given only these two components, the agent builds the EO model and recursively improves it on its own outputs. Problems long treated as separate research areas—few-shot learning, cross-domain adaptation—reduce to choices of harness and verifier. Across four EO benchmarks spanning optical, SAR, and multispectral time-series imagery, agent-built models surpass the state of the art, retaining most accuracy with only about 1% of training labels. This establishes a new EO paradigm: humans design harnesses and verifiers, agents build the models, and EO models improve alongside agent capabilities.

The Last Remote Sensing Model Built by Humans? Casting Any Earth Observation Task for Agents

Paper Overview

Field: NLP Authors: Sijie Zhao, Zeqi Zhou, Heli Qi, Xuanchao Fu, Zhuo Zheng, Junshi Xia, Junjue Wang, Weihao Xuan, Naoto Yokoya Published: 2026-10-13 arXiv: 2610.18845

Abstract

Earth observation (EO) models are still built by hand. For each new task, experts design a training recipe and tune it against a benchmark. We show that this loop can be handed to an agent. Any verifiable task can be cast as a harness and a verifier, where the harness defines what the agent may use and the verifier defines what counts as success. Every EO task can be made verifiable, through labeled benchmarks, independent sensors, physical models or the repeated observation of the same place. Given only these two components, a general-purpose coding agent builds the EO model and improves it recursively on its own outputs. Settings long treated as separate research problems, from learning with a handful of labels to adapting across domains, become choices of harness and verifier. On four EO benchmarks spanning optical, SAR, and multispectral time-series imagery, agent-built models surpass the state of the art while retaining most accuracy with only about 1% of the training labels.

Key Ideas (translated from the Chinese summary)

The manual loop of task-specific training recipe design and benchmark tuning can be delegated to an agent.
Any verifiable task can be expressed as a harness (what the agent may use) plus a verifier (what counts as success).
EO tasks can be made verifiable via labeled benchmarks, independent sensors, physical models, or repeated observation of the same location.
With only these two components, a general-purpose coding agent builds the EO model and recursively improves it on its own outputs.
Settings historically studied as separate research problems—few-label learning, cross-domain adaptation—become mere choices of harness and verifier.
On four EO benchmarks covering optical, SAR, and multispectral time-series imagery, agent-built models beat existing state of the art and keep most accuracy with only ~1% of training labels.
New paradigm proposed: humans design harnesses and verifiers; agents build the models; EO models advance together with agents.

---

*Auto-collected on 2026-10-15.*