This master’s thesis is presenting a novel application of the deep learning model TransUNet for segmenting sea ice into three charts. Sea Ice Concentration, Stage of Development, and floe sizes being the output charts of the model. Sea ice plays a critical role in global climate systems, and it is therefore beneficial to have precise measurements of the sea ice extent. Additionally, the shipping industry will benefit from having near real-time updates on the sea ice extent, as it will make navigation in the arctic regions more predictable. To address these challenges, TransUNet, a CNN-Transformer hybrid offering state-of-the-art segmentation due to its local and global awareness, is deployed in the context of sea ice classification. The developed model is selected by training different TransUNets that differs from each other by having changes in the configuration of the transformer. It is found that the number of layers and the patch size has a large impact on performance, and thus the best performing model is selected for further training. The training was ended after 120 epochs, and the combined validation accuracy ($R^2$ is used for SIC and $F1$ is used for SOD and FLOE) topped at 92.15%. The results from testing show that the model performs in line with state-of-the-art with a combined score of 86.22%. Additionally, an accuracy of 86.91% was achieved on SIC, seeing an improvement of 0.57% compared to previous work. This thesis proves the viability of deploying TransUNets as a semantic segmentation method in remote sensing for predicting sea ice charts.