Year: 2026
Type: preprint
Abstract: Satellite flood mapping faces a fundamental spatiotemporal trade-off: Synthetic Aperture Radar (SAR) missions provide high-resolution flood masks but revisit any given location only every 6–12 days, while GNSS Reflectometry observations from the CYGNSS constellation are available daily but at kilometer scale. Neither sensor alone suffices to track flood events evolving over 24–72 hours. We propose a physically-constrained deep learning framework that fuses daily CYGNSS inundation signals with Sentinel-1 SAR temporal anchors, conditioned on MERIT Hydro topographic variables, to produce daily flood maps at 90 m resolution. Flood observations are organized as temporal triplets (A→B→C), where B is the target date and A and C are SAR acquisitions bracketing B by 5–7 days. A temporal U-Net with three parallel encoding branches and learnable cross-attention gates fuses these inputs into spatially explicit flood probability maps. Trained and evaluated on a 47,653-tile dataset spanning seven flood-prone regions across three continents and seven years (2019–2025), the framework achieves a micro-averaged IoU of 0.703, outperforming the SARonly state of the art (IoU = 0.67) despite solving a harder downscaling problem. Ablation establishes that SAR temporal anchors dominate performance, while the CYGNSS contribution is driven by the climatological anomaly channel and peaks on moderately fast-changing flood scenes. A conditional diffusion model trained as a post-processing module produces calibrated ensemble flood maps with spatially explicit uncertainty estimates, a capability absent from existing operational systems. The framework is designed for direct transfer to the NASA-ISRO NISAR L-band mission, extending applicability to forested tropical floodplains inaccessible to C-band SAR.
DOI: 10.31223/x5222g
Authors: Gauthier Malandrin, Cynthia Gerlein-Safdi
Institution: École Polytechnique
Language: English
Subfield: Global and Planetary Change
Field: Environmental Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7212270167