Remote sensing-based sea ice concentration estimation via weighted deep learning networks

Sea ice concentration is an essential monitoring parameter in climate change processes and in navigation and environmental planning in polar regions. Although conventional encoder–decoder of convolutional neural networks perform well in sea ice segmentation, sharp spatial edges and marginal ice zones are difficult to segment without overburdening the computational overhead. A new architecture named feedforward weighted group convolution (FF-WGC), which combines group convolution, asymmetric channel weighting, and lightweight feedforward augmentation, is developed in this research to improve the performance and efficiency of segmentation. The FF-WGC model performs selective discriminative feature group emphasis and intragroup feature learning by progressively involving convolutional transformations. Tests on Sentinel-2 imagery and Canadian Ice Service charts in Hudson Bay reveal that FF-WGC outperforms the current segmentation baselines DeepLab, SegNet, and FCDenseNet in all major indicator measures, including the intersection over union (IoU), accuracy, precision, recall, and F1 score. The proposed architecture produced an IoU of 47.5 % with non-statistically significant additional complexity and offers a good compromise between the required performance and prolonged operation in large-scale sea ice applications. The current work can be extended by incorporating attention mechanisms and testing over multitemporal sequences to forecast sea ice conditions, with validation in other Arctic and Antarctic regions. • Proposed a high-performance model for automated sea ice concentration estimation. • Group convolution module enhances feature learning for multiclass segmentation tasks. • Asymmetric weight design enables targeted learning tailored to unique sea ice features.