Semantic Segmentation of Flooded Vegetation in SAR Imagery: A Benchmark Dataset and Comparative Study of Deep Learning Models

Abstract

Accurate flood extent mapping plays a critical role in disaster management and response. With the increasing availability of high-resolution and frequently acquired satellite imagery, synthetic aperture radar (SAR) has become a valuable tool for flood delineation. However, one of the main challenges in SAR-based flood mapping remains the presence of vegetation, which obscures water surfaces and complicates detection. In this study, we introduce VegFlood, a reference dataset designed for training, validating, and testing deep learning models for semantic segmentation of flooded areas, including regions covered by vegetation. The dataset was constructed using publicly available Sentinel-1 SAR imagery, combined with high-resolution digital terrain models and hydrological observations from gauging stations. VegFlood comprises 1707 image-mask pairs from 39 gauging stations, geographically limited to the territory of Poland. To establish baseline performance on the VegFlood test set, we conducted a series of experiments using state-of-the-art deep learning models for flood segmentation. The experimental results consistently demonstrate that FV is the most difficult class to detect, achieving substantially lower Intersection over Union (IoU), User Accuracy (UA), and Producer Accuracy (PA) values compared to OW and NF classes. Among the evaluated models, RS-Mamba achieved the highest performance, reaching a mean IoU of 0.566, with UA of 0.689 and PA of 0.689.Further analysis shows that segmentation accuracy is strongly influenced by vegetation density, flood water level, and local geomorphological conditions. These results confirm the limitations of C-band SAR imagery for flooded vegetation detection and highlight the importance of dedicated datasets that explicitly account for FV as a separate semantic class.