MOMO: An RGB–Pseudo-Event Spiking Network for Robust Underwater Object Detection in Aquatic Ecosystem Monitoring

Computer-vision-based monitoring of aquatic organisms and benthic targets requires detectors that remain reliable across variable water bodies and acquisition conditions. Underwater imagery is affected by wavelength-dependent attenuation, scattering, turbidity, illumination interference, and color cast, which can degrade models trained in source environments when they are deployed in unseen monitoring conditions. Long-duration observation with autonomous or edge platforms also requires favorable accuracy–efficiency characteristics. This study presents MOMO (Multi-Representation Optical Monitoring of Objects), an RGB–pseudo-event spiking detector for robust ecological visual monitoring. Here, “multi-representation” refers to two complementary representations of the same optical input: the RGB appearance view and the frame-derived pseudo-event structural view; it does not denote independently acquired sensing modalities. The method combines RGB appearance with a frame-derived pseudo-event structural representation in a quantized spiking neural network. A training-free generator converts each RGB frame into a three-channel pseudo-event image through grayscale projection, Prewitt gradients, polarity separation, and sample-wise normalization, Under a spatially constant positive affine photometric transformation, the representation remains unchanged apart from discretization and numerical stabilization. A dual-branch spiking backbone extracts RGB semantic features and pseudo-event structural features, which are fused by multi-scale spiking cross-attention. Under the standard unseen-domain evaluation protocol of S-UODAC, MOMO achieves 57.2% [email protected], outperforming the strongest evaluated SNN baseline by 3.5 percentage points. Across the haze-like, color-cast, and illumination-interference subsets of RUOD, the mean mAP@[0.5:0.95] is 43.7%, with stronger performance under color and illumination variation. Ablation studies and feature analyses indicate that both the pseudo-event representation and spike-compatible fusion improve target–background discrimination under cross-domain water-condition shifts. The theoretical neural-network inference energy is estimated at 24.21 mJ per frame from arithmetic operations under a representative 45 nm model.