AquaMamba: Selective state space modeling for underwater scene segmentation in marine robotics

Underwater scene segmentation is challenging because image degradation and domain shift vary greatly in marine environments, and the behavior of selective state-space models (SSMs) in this environment remains insufficiently characterized. We propose AquaMamba, a ResNet34-UNet segmentation framework that replaces the bottleneck operator with stacked Mamba layers, and compare it with five established CNN-based methods and a constructed Transformer-bottleneck baseline in the SUIM benchmark. AquaMamba achieves an mIoU of 0.669, 1.9 percentage points higher than the controlled Trans-UNet baseline (0.650), while using fewer parameters (29.5M vs. 33.9M). DeepLabV3 + remains the strongest overall model, reaching 0.712 mIoU. On the unlabeled UIEB dataset, AquaMamba shows the highest confidence retention among four representative models, retaining 93.5% of its source-domain prediction confidence. Due to the lack of semantic annotation in UIEB, this result reflects prediction confidence rather than target-domain segmentation accuracy. Scale- and boundary-specific diagnostics within the selected representative subset further show complementary behavior among AquaMamba, Trans-UNet and DeepLabV3 + : AquaMamba is competitive in several large object categories and obtains the lowest boundary error point estimate, while DeepLabV3 + remains stronger for human divers and plant/seagrass, and AquaMamba is limited on the reported small-RO subset. These findings position AquaMamba as a controlled SSM baseline for underwater semantic segmentation and emphasize the need for future SSM designs to better integrate multi-scale and high-resolution features.