Multi‐Reference External Validation of Synthetic Aperture Radar Flood Mapping in Hanoi With Volunteered Geographic Information–Based Infrastructure Exposure Assessment

ABSTRACT Sentinel‐1 SAR flood mapping in data‐scarce urban deltas typically relies on algorithmically generated pseudo‐labels, creating circular accuracy assessments. We present an operational SAR‐to‐VGI workflow with external validation against two methodologically independent operational reference products, applied to two typhoon events in Hanoi, Vietnam—Yagi (September 2024) and Matmo (October 2025). Model recall against the manually delineated Copernicus Emergency Management Service EMSR754 reference is 0.882 ± 0.016; raw IoU is 0.096 ± 0.010, scoped as a structural lower bound under a two‐day SAR/CEMS temporal offset and pseudo‐label‐calibration effects rather than as a calibrated accuracy claim. Against the same‐date Sentinel‐1 JRC Global Flood Monitoring product, recall is 0.996; a GFM 12–15 September recession check independently confirms approximately 57% post‐peak flood‐area reduction over 3 days. Cross‐event transfer is strongly asymmetric (Yagi → Matmo: ΔIoU = −0.347; Matmo → Yagi: +0.057), cautioning against single‐event‐trained operational deployment within a single metropolitan area. Three robustness probes, label‐noise sensitivity at N = 12, a hydrological‐consistency loss regularization and an attention‐gate architectural comparison, are reported as null or sub‐threshold rather than omitted. The VGI exposure pipeline estimates approximately 1516 buildings and 107.8 km of roads affected by Typhoon Yagi across three central‐southern districts of Hanoi, with explicit per‐district OpenStreetMap completeness disclosure.