This paper introduces a Bayesian Meta Pseudo Labeling (BMPL) method for optimizing semi-supervised sea ice classification in marine remote sensing imagery. Addressing the challenge of limited labeled data in marine remote sensing, BMPL optimizes the pseudo-labeling process through a Bayesian inference framework and enhances model generalization via meta-learning strategies. Compared to traditional semi-supervised methods, BMPL demonstrates significant advantages in both theoretical foundations and practical performance. Experiments conducted on a comprehensive dataset including open water, new ice, young ice, and first-year ice showcase the method’s superior performance, particularly in complex mixed-ice and marginal ice zones. BMPL significantly improves the accuracy and efficiency of sea ice classification, providing reliable technical support for marine environment monitoring, marine resource assessment, and polar navigation safety. This research not only advances marine remote sensing technology but also offers a powerful tool for marine ecosystem protection, climate change studies, and sustainable fisheries management, contributing to multiple ocean-related issues.