From Passive Observation to Active Multi-Agent Sensing in Earth Observation

Xiaojian Wang, Zhengxiong Li

Keywords: multimodal Earth observation, active sensing, multi-agent observation

TL;DR: Multimodal Earth observation can be designed as an active loop between satellites and robots.

Abstract: Multimodal Earth observation has framed the "imperfect lens" of cloud cover, resolution shortfall, asynchronous radar/optical acquisitions, and irregular revisit as a modeling problem, addressed with fusion, imputation, cross-modal translation, and foundation-model pretraining. We argue these imperfections are also a sampling problem. Implicit in this framing is an assumption that observation is taken as given: revisit schedules are set by orbit, pre/post-event pairs are whatever was collected, and ground truth is annotated long after the fact. What is missing is a sensing layer that is now widely deployed in the world: autonomous robotic platforms such as aerial robots, ground robots, and autonomous vessels, which can be tasked on the basis of what satellites cannot see. We position robots as active sampling agents in a multimodal Earth observation loop: they consume satellite-derived uncertainty maps as a tasking signal, contribute high-resolution observations as a modality alongside satellite data, and provide in-situ ground truth that closes the loop on online model adaptation. We diagnose four mismatches between current passive-observation practice and what active multi-agent observation requires, and trace three concrete design moves toward an Earth observation stack in which satellites and robots act as co-observers rather than separate data pipelines.

Submission Number: 14