Answers are grounded in this paper and cite their sources.
Sentinel-1 acquires VV and VH backscatter simultaneously, but most deep learning methods for flood mapping simply stack them as a two-band input. The two channels do not measure the same thing: VV is sensitive to the contrast between open water and dry land, while VH keeps a flood signal under partly inundated vegetation, where VV often fails. We propose PolarFuseNet, a compact dual-encoder model that processes each channel in its own branch before merging them through a Cross-Polarization Fusion Block (CPF-Block). The CPF-Block applies Squeeze-and-Excitation recalibration on each branch first, then runs bidirectional cross-gating, so one polarization can guide the other at full spatial resolution. We test the model on Sen1Floods11, STURM-Flood and S1S2-Water. PolarFuseNet has 2.03 million parameters, 15 times fewer than U-Net. On Sen1Floods11, U-Net is slightly ahead (69.35% vs 68.26% flood IoU), a gap of about one percentage point despite this large difference in model size. Over five random seeds, the block improves on a single-encoder baseline by 0.78 points of flood-class IoU on Sen1Floods11 and 1.83 points of Water F1 on STURM-Flood, for 2.96% more parameters; the ablation results point to per-branch recalibration as the primary source of the observed mean gain, whereas the additional contribution of cross-gating is not measurably supported on the two flood benchmarks. Analysis of the learned gates shows that, across all ten events tested, VV features receive lower gate weights over flooded pixels than over dry pixels, whereas VH features receive higher gate weights and are therefore less strongly suppressed in flooded areas. Leave-one-event-out gives 56.93±18.34% flood-class IoU, with the lowest scores in arid regions where dry soil and shallow water look almost identical in a single SAR acquisition.
No takes yet. Share an insight, caveat, or question.
Kzadri et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: