Date of Publication : May 2026

Abstract: Floods are one of the most destructive natural hazards in South Asia, and it is difficult to map during the monsoon season because cloud cover blocks optical satellites days at a time. Sentinel-1 SAR avoids this: it can work behind clouds, during the day or night, at 10-meter resolution and with repeat cycles of 6-12 days. The paper is a review of the various ways deep learning architectures to flood segmentation of Sentinel-1 have evolved, starting with fully convolutional networks, and moving on to modern vision transformers. We compare three families of models: UNet++ (CNN), Swin-UNet (transformer) and SegFormer-B0 (efficient transformer) on a regionally stratified dataset with 55% South Asian coverage using Sen1Floods11, Kuro Siwo and UrbanSARFloods as benchmarks. With only 3.7M parameters, SegFormer-B0 has an IoU of 0.6404, which is very small but still provides a benefit over other models that are both smaller and larger at the same time. There are four gaps: urban double-bounce scattering, sparse regional labeled data, topographic false positives, and near-real-time inference. We map the way ahead to each.

Reference :