MambaMPD: a mamba-driven segmentation framework for marine pollution detection from remote sensing imagery

Accurate detection of marine pollution is essential for protecting coastal ecosystems and marine biodiversity. Recently, vision Mamba-based approaches have shown promise in remote sensing semantic segmentation due to their ability to efficiently capture long-range dependencies and global context. However, their potential remains largely unexplored in the context of marine pollution detection (MPD), with distinct challenges, including low signal-to-noise ratios, spatial fragmentation of pollution patterns, and indistinct boundaries caused by strong visual similarity between pollutants and the surrounding marine environment. To address these challenges, we propose MambaMPD, an enhanced Mamba-based framework tailored for marine pollution detection. MambaMPD incorporates two targeted modules that enhance the Mamba encoder with complementary structural priors: the frequency-aware augmentation (FAA) module and the multi-scale edge-guided attention (EGA) module. The FAA module strengthens the encoding process by integrating Wavelet Transforms, which decompose features into multi-scale frequency subbands. This enables the model to efficiently capture both low-frequency contextual semantics and high-frequency structural details, essential for identifying small, low-contrast, and irregular pollution patterns. Meanwhile, the EGA module adaptively integrates hierarchical Laplacian-derived multi-scale boundary cues into deep semantic representations, guiding the refinement of encoder features before decoding, thereby sharpening boundary delineation and alleviating ambiguity in visually confusing and spatially fragmented marine scenes. Additionally, a U-Net-style decoder equipped with squeeze-and-excitation attention and deep supervision is employed to progressively recover and refine semantic and spatial features across multiple scales. Extensive experiments on two benchmark marine pollution detection datasets demonstrate that the proposed method outperforms the compared methods in mIoU while maintaining significantly lower computational cost than foundation-model-based approaches. On MADOS, it surpasses OSDMamba by 3.6% in F1, and on M4D it improves Oil Spill IoU by 6.82% over TransOilSeg. The source code of our approach will be available at https://github.com/Multimodal-Intelligence-Lab-MIL/MambaMPD.