When Clouds Silence Optical — Self-Gated Evidential Fusion for Cloud-Robust Flood Mapping

Abstract

Sentinel-1 / Sentinel-2 fusion is critical for flood inundation mapping under cloud cover, yet shared-attention Remote Sensing Foundation Models (RSFMs) have no built-in mechanism to down-weight an occluded modality. On the Sen1Floods11 Bolivia held-out split, a TerraMind baseline collapses on heavy-cloud tiles (IoU_Flood = 0.000), confirming that modality laziness persists at foundation-model scale.

We propose a multi-branch evidential head with a law-of-total-variance gate that routes per-pixel weight across S1, S2, and fused branches at inference — entirely label-free, with no auxiliary cloud detector. The method recovers heavy-cloud performance (IoU_Flood 0.328 → 0.426; Recall 0.431 → 0.555) while improving precision, and yields well-calibrated per-branch uncertainties as a free by-product (AUROC up to 0.952 in-distribution).