Improving Sea Ice Deep Learning by Refining Training Labels from Operational Ice Charts: A Study Using Vector-to-Raster Registration and Multispectral Landsat-8 Data in the Baltic Sea

Sea ice classification models trained on manually labeled ice charts from public ice services are often limited by operational constraints and human annotation practices. These charts typically generalize spatial boundaries and provide a conservative representation of ice features. To address these challenges, we employ two label refinement strategies: spectral water masking and vector-to-raster registration. Specifically, spectral water masking resolves fine-scale open water features, while vector-to-raster registration aligns chart boundaries with satellite imagery to correct for sensor misalignment and temporal ice drift. To isolate the impact of these refinements, we apply them to Baltic Sea ice charts provided by the German Ice Service, utilizing the same dataset and neural network architecture as prior work. In controlled and balanced experiments, we find that these spectral and semantic refinements increase the Intersection over Union (IoU) from 0.68 to 0.77, with the incorporation of thermal emission channels providing further improvement to 0.81. The substantial increase in IoU and qualitative inference results demonstrate that the model more accurately captures physical boundaries and fine-scale ice structures. While this study focuses on Landsat-8 multispectral imagery, the vector-to-raster registration is sensor-agnostic and readily applicable to other optical or SAR platforms, such as Sentinel-2 and Sentinel-1.