Sea Ice Concentration Mapping Based on Sentinel-2 Multispectral Imagery Using Multiscale Deep Learning Approach

One of the important climate indicators is the condition of sea ice concentration in the Arctic and Antarctic regions. Measuring this concentration using remote sensing and automated artificial intelligence allows continuous monitoring at relatively low cost. The purpose of this study is to develop a robust and scalable deep learning model to accurately estimate sea ice concentration from satellite imagery, supporting climate monitoring efforts. Hence, Ice-SPP-net (ISPP-net) is introduced as an innovative approach to measuring sea ice concentration using Sentinel-2 imagery, leveraging a novel multiscale convolutional neural network (CNN) architecture. This model is specifically designed to address challenges posed by the complex and heterogeneous nature of sea ice, which conventional fixed-scale CNNs struggle to assess accurately. By incorporating Spatial Pyramid Pooling (SPP) modules, the ISPP-net effectively captures and integrates features across multiple scales, enhancing the model’s ability to discern both fine details and broader contextual information essential for precise measurements. The model was trained and evaluated using annotated Sentinel-2 datasets and validated through a comprehensive ablation study to determine the optimal architecture configuration. Results show that placing the SPP module at the third encoder layer, using three parallel paths, yields the best performance. This configuration outperforms models like U-Net, SegNet, DABNet, TernausNet, and FC-DenseNet, achieving a mean Intersection over Union of 49.20% and an accuracy of 89.30%. ISPP-net provides critical data to refine climate models and improve predictions of future climate scenarios. Future enhancements will integrate more data sources and refine the architecture, supporting SDG 13: Climate Action.