Artificial intelligence-Enabled Lightweight Flood Segmentation Model with Polarization Fusion and Attention

Abstract

Rapid and reliable flood maps are essential for emergency response and disaster management. However, many deep learning models for flood segmentation are computationally demanding, limiting real-time use and deployment on modest hardware. This paper presents PFALNet, a reduced-parameter segmentation network that integrates dual-polarization synthetic aperture radar image backscatter and a polarization-ratio channel. The model employs concurrent spatial and channel squeeze-and-excitation attention to enhance multi-scale feature fusion for flood delineation. Experiments on a public Sentinel-1 flood-mapping benchmark show that PFALNet achieves strong performance on the Florence test region (F1 = 90.53%, IoU = 82.70%, κ = 88.80%) while remaining lightweight (0.31M parameters, 5.88 GFLOPs) and enabling real-time inference (6.23 ms per 256 × 256 tile, ∼ 160 tiles/s) on a single GPU. These results indicate that carefully designed lightweight synthetic aperture radar (SAR) segmentation models can deliver competitive performance with substantially reduced computational cost.