The Last Remote Sensing Model Built by Humans? Casting Any Earth Observation Task as Recursive Self-Improvement

Imagine you're tasked with monitoring environmental changes using satellite imagery. Traditionally, experts painstakingly design and tune models for each specific task, which can be time-consuming and often leads to suboptimal results. This process can break down when faced with new challenges, as it relies heavily on human expertise and can struggle with limited data or varying conditions — a situation known as the 'data scarcity problem.' As a result, many Earth observation tasks remain underexplored or poorly executed due to the manual nature of model development.

The authors propose a solution where a coding agent takes over the model-building process. By defining a 'harness' that outlines the tools available and a 'verifier' that sets success criteria, any Earth observation task can be automated. This means that the agent can learn and adapt by itself, improving its models based on its own outputs. The approach allows for a more flexible and efficient way to tackle various tasks, from those with limited labeled data to cross-domain adaptations. In tests across four different Earth observation benchmarks, models built by the agent not only surpassed existing state-of-the-art results but also maintained high accuracy with just a fraction of the training labels. This shift means that instead of experts crafting every detail, they can focus on designing the frameworks that guide the agents, leading to continuous improvement in model performance as the agents learn and evolve.

The approach introduces a new paradigm for automating Earth observation model building.

The results are validated across multiple benchmarks, showing consistent performance improvements.