Attention placement and the precision–recall balance in few-shot urban Typhoon flood mapping: a case study of Shanghai during Typhoon In-Fa

Heterogeneous urban surfaces and dark backgrounds confuse spectral water indices, hampering flood mapping in coastal megacities. Using Lujiazui (a high-density business district) and Lingang (a coastal storm-surge area) in Shanghai during Typhoon In-Fa (2021) as a case study, we examined how attention type and placement affect few-shot U-Net flood segmentation from Sentinel-1 SAR and pre-event Sentinel-2 data. Models were pre-trained on Sen1Floods11 and fine-tuned with 105 target-domain patches. Three U-Net variants showed comparable agreement with VH-derived reference labels (IoU 91.60–91.75%). In a six-configuration ablation, decoder-side CBAM had higher mean recall than encoder-side CBAM (97.22% vs 96.47%) and lower mean precision (94.18% vs 94.80%); with matched attention initialisation, these differences kept their direction but were not significant. In proxy dark-background areas, CBAM-U-Net did not show a lower false-positive rate than Vanilla U-Net or Attention U-Net. Agreement depended strongly on VH backscatter, and a SAR-only model showed no detectable IoU difference from the six-channel model (+0.25 pp; p = 0.311). Under a 500-pixel connected-component rule, the area classed as flood was 5.23 km2, but it ranged from 2.35 to 12.86 km2 across thresholds of 100–10,000 pixels. Whether placement can favour precision or recall requires further testing.