<![CDATA[Multimodal Earth-observation (EO) foundation models increasingly rely on the joint use of optical imagery, synthetic aperture radar (SAR), elevation, and auxiliary geospatial signals to support land-cover mapping, biomass estimation, damage assessment, and change detection. Yet the empirical regime in which such models are usually trained and reported remains overly optimistic: modalities are commonly assumed to be synchronised, complete, and clean. This assumption is not defensible for real deployments. Optical observations are routinely obscured by clouds, SAR coherence and interferometric products can be unavailable or decorrelated, spatial alignment across sources is imperfect, and large geospatial archives inevitably contain broken tiles, missing channels, or temporally inconsistent acquisitions. Starting from the released M3LEO dataset and framework, which already expose multimodal EO learning at continental scale, this paper reconstructs the underlying research direction toward a more technically urgent problem: robust multimodal foundation learning under missing or corrupted modalities. We formalise a modality-resilient framework, RAMEO, that combines modality-specific token encoders, reliability-aware gated fusion, masked cross-modal reconstruction, corruption-aware consistency learning, and calibrated uncertainty estimation. We also define a rigorous evaluation protocol over geographically disjoint splits, structured missingness patterns, and modality-specific corruptions. This paper done a dataset-and-benchmark paper and reconstruction distinguishes carefully between source-anchored evidence reported for M3LEO and new robustness analyses that are specified as executable experiments. The result is therefore technically coherent, and reproducible, while remaining honest about what has and has not yet been empirically established.]]>
Copyrights © 2026
