Accurate, real-time flood mapping from synthetic aperture radar image is vital for disaster response but is hindered by a persistent trade-off between model accuracy and computational efficiency. This paper introduces FM-Mamba, a novel network that leverages a state space model to break this bottleneck. Its core innovation is an encoder built with non-causal Mamba blocks, which captures essential long-range spatial context with linear complexity, paired with a parameter-efficient decoder designed for precise boundary recovery. Evaluated on the Sen1Floods11 and S1GFloods benchmarks, FM-Mamba achieves leading segmentation accuracy, matching or exceeding state-of-the-art methods in F1-score and IoU. Crucially, it accomplishes this with only 3.93 million parameters and a drastically reduced computational footprint, demonstrating a superior balance of performance and efficiency that is ideal for operational, real-time flood mapping.
Wang et al. (Mon,) studied this question.