A Cross-Domain Mamba Network with joint spatial-frequency learning for robust SAR oil spill detection

Marine oil spills pose severe threats to ecosystems, prompting the development of synthetic aperture radar (SAR)-based detection technologies. Nevertheless, critical challenges persist. First, prevalent "look-alikes" in SAR imagery manifest as dark spots that closely resemble oil spills, complicating accurate identification. Second, existing approaches predominantly rely on spatial features, often neglecting the potential of frequency-domain analysis. To address these limitations, this paper proposes the Cross-Domain Mamba Network (CDMNet) for robust SAR oil spill detection. CDMNet employs a Cross-Domain Mamba Block (CDMamba) to jointly model spatial and frequency representations, thereby enhancing the discrimination between genuine oil spills and complex look-alikes. Additionally, a Scale-Aware Pyramid Pooling (SAPP) module captures local multi-scale contexts and global structural information via parallel pooling branches, ensuring high boundary fidelity for morphologically complex targets. Moreover, a Multi-Level Feature Fusion Module (MFFM) integrates deep semantic cues with fine-grained spatial representations for the precise localization of thin films and accurate delineation of weak boundaries. Quantitative and qualitative experimental results on the SOS and M4D datasets demonstrate that the proposed CDMNet outperforms currently popular methods in terms of performance. The source code is available at: https://github.com/FF7CA/CDMNet.