Answers are grounded in this paper and cite their sources.
Preserving ocean ecosystems and achieving UN Sustainable Development Goal 14 require automated marine pollution surveillance. Nevertheless, marine satellite-based semantic segmentation of marine pollution faces extreme scale imbalance, with pollution classes comprising less than 0.5% of pixels relative to extensive marine water, leading to traditional algorithms showing majority-class gradient bias in training and precision-recall biases in inference. To overcome these shortcomings, this article proposes MEP, a unified model that addresses class imbalance through combined architectural, optimization, and inference-level innovations. The Multiscale Attention Swin Hamburger architecture uses a Swin Transformer encoder with shifted-window self-attention for hierarchical feature decomposition, alongside a Hamburger decoder that uses nonnegative matrix factorization to model global context. The Equal Adaptive Gradient Learning Engine automatically estimates per-batch inverse-frequency weights for classes, weighted focal loss, and class-balanced Dice loss to ensure a democratic contribution of the gradient, regardless of pixel count. During inference, Precision-Aware Selective Test-Time Augmentation combines eight geometric augmentations and applies confidence-sensitive rejection to avoid systematic overprediction without discarding actual detections. State-of-the-art results on the MADOS benchmark have been achieved, with a mean F1-score of 79.0%, a mean IoU of 69.2%, and an overall accuracy of 94.1%, representing improvements of 3.0%, 4.9%, and 5.0% over the original published MariNeXt results. The framework achieves strong performance across 14 of 15 classes, though the Foam class remains challenging due to fundamental spectral similarity with Turbid Water at the available sensor resolution. Furthermore, cross-validation on the MARIDA benchmark yielded promising results without dataset-specific tuning, confirming generalizability across marine environments.
No takes yet. Share an insight, caveat, or question.
Sangaiah et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: