Marine debris segmentation in underwater environments remains a challenging task owing to visual degradation caused by light absorption, scattering, and turbidity. Although deep-learning-based vision systems have shown promising results, their prediction confidence under such conditions is often unreliable. In this study, we propose an entropy-guided uncertainty quantification framework using the HyenaPixel model, an attention-free architecture capable of capturing global context with large convolutional kernels. We apply Monte Carlo Dropout during inference to estimate epistemic uncertainty and analyze its correlation with per-class segmentation performance. Experiments conducted on a curated split of a public Ocean Debris Segmentation dataset containing 22 classes demonstrate that uncertainty is a meaningful indicator of model reliability, and may reveal overconfident mispredictions and unstable classes. The proposed lightweight framework can be readily integrated into underwater vision pipelines, offering interpretable uncertainty measures that are essential for reliable and risk-aware decision-making.