Multimodal GNSS-R self-supervised learning as a generalist Earth surface monitor

Abstract

The increasing frequency of climate extremes and natural disasters demands rapid and scalable Earth surface scans for effective action. Emerging as a novel remote sensing technique, spaceborne global navigation satellite system reflectometry (GNSS-R) plays an increasingly vital role in monitoring Earth’s surface parameters. Recent studies leverage the growing volume of GNSS-R measurements with data-driven approaches to enhance retrieval products over both ocean and land. Yet, these models are typically trained using supervised learning, which requires extensive feature engineering and application-specific annotations. To address these limitations, we propose the first GNSS-R self-supervised learning framework as a generalist Earth surface monitor (GEM). Our model is pretrained on multimodal observables, i.e., delay-Doppler maps (DDMs) and auxiliary parametric data, to learn cross-modal representations from GNSS-R data. To validate the effectiveness of the proposed approach, we fine-tune the pretrained model on various downstream retrieval tasks, including ocean wind speed retrieval, surface soil moisture estimation, and vegetation water content prediction. The results demonstrate that our framework generalizes well across these tasks, providing a versatile solution for GNSS-R-based Earth surface monitoring and facilitating further exploration of novel use cases.