Confronting the few-event data ceiling in a prototype flood early-warning system for Hanoi: weak bi-temporal SAR labels and continuous self-verification

Deep-learning flood mapping presumes abundant labelled events, but a single-city early-warning system rarely has them. We report an operational-prototype flood early-warning system for Hanoi, Vietnam, built entirely on open data and open-source tools that forecasts flood extent 6–48 hours ahead and ask how far model sophistication helps given the use of only three hand-labelled events. The system comprises Sentinel-1 GRD, GPM IMERG, ERA5 and river-gauge data and runs a four-model stack (physics, XGBoost, U-Net, LoRA-adapted Clay) combined with IoU-weighted voting. A verification loop scores each forecast against auto-generated bi-temporal SAR change masks (same-orbit, >3 dB drop, HAND/land-cover masking); weak labels carrying speckle and threshold uncertainty are not definitive observations. Historical floods mined from the 2015–2026 archive (62 scenes) add weak labels. The hand-mapped Copernicus EMS product EMSR754 (Typhoon Yagi 2024) is the independent evaluation reference; this and the Yagi scene were withheld from model training and weak-label generation, although per-model thresholds and ensemble weights were tuned against it, so reported IoUs were tuned and not blind. On the held-out Yagi event, AOI-clipped IoU rises from 0.00 (physics) to 0.09 (U-Net), 0.16 (Clay) and 0.23 (XGBoost) and to 0.28 (ensemble; 18.7 vs 21.6 km² mapped). Across tested configurations, added complexity did not improve Yagi skill: lower LoRA rank (0.16→0.10), augmentation (→0.13) and two pretrained-decoder inits underperformed for cold initialization. Bi-temporal weak labels removed steady-state land cover (rice paddy ~31%, tree canopy ~45%), which single-scene Otsu conflates with flood and archive mining, growing the training set from 3 to 10 events, including a dam-release flood with 0.3 mm concurrent rainfall. Here, data availability appears to be the binding constraint. Auto-generated bi-temporal labels plus self-verification offer an open, low-cost route to more data; the prototype runs in shadow mode through the 2026 monsoon; and a hand-versus-weak ablation follows at season end.