AI4SeaIce: Task Separation and Multistage Inference CNNs for Automatic Sea Ice Concentration Charting

• Main Authors: Stokholm, Andreas, Kucik, Andrzej, Longépé, Nicolas, Hvidegaard, Sine Munk
• Format: Article in Journal/Newspaper
• Language: English
• Published: Copernicus Publications 2023
• Subjects: article; Verlagsveröffentlichung; Sea ice
• Online Access: https://doi.org/10.5194/egusphere-2023-976; https://noa.gwlb.de/receive/cop_mods_00067435; https://noa.gwlb.de/servlets/MCRFileNodeServlet/cop_derivate_00065895/egusphere-2023-976.pdf; https://egusphere.copernicus.org/preprints/2023/egusphere-2023-976/egusphere-2023-976.pdf

We investigate how different Convolutional Neural Network (CNN) U-Net models specialised in addressing partial labelling tasks related to mapping Sea Ice Concentration (SIC) can improve performance. We use Sentinel-1 SAR images and human-labelled ice charts as the reference to train models that benefit from advantages gained from different model optimisation objectives by utilising a multistage inference scheme. We find our multistage model inference approach that apply a classification (CrossEntropy or Earth Mover's Distance squared) optimised model to separate open water, intermediate SIC and fully covered ice in conjunction with a regression (Mean Square Error or Binary CrossEntropy) optimised model, that assigns specific intermediate classes, to perform the best. To evaluate the models we introduce several specific metrics illustrating the performance in key areas, such as the separation of macro classes, intermediate class, and an accuracy metric better encapsulating uncertainties in the reference data. We achieve R2-score of ~93 %, similar to state-of-the-art in the literature (Kucik and Stokholm, 2023). However, our models exhibit significantly better open water and 100 % SIC detections. The multistage synergises high open water and fully covered sea ice accuracies achieved with classification optimised objectives with good intermediate class performance obtained by regressional loss functions. In addition, our findings indicate that the number of classes that the intermediate concentrations are compressed into does not influence the result significantly it is the loss function used to optimise the model that assigns the specific intermediate class to have the largest impact.