Classification of Sea Ice Types in Sentinel-1 SAR images

Abstract. A new Sentinel-1 image-based sea ice classification algorithm is proposed to support automated ice charting. Previous studies mostly rely on manual work in selecting training and validation data. We show that the use of readily available ice charts from an operational ice services allow to automate selection of large amount of training/testing data void of biased, subjective decisions. The proposed scheme has two phases: training and operational. Both phases start from removal of thermal, scalloping and textural noise from Sentinel-1 data and calculation of gray level co-occurrence matrix and Haralick texture features in a sliding window. In the training phase, the weekly ice charts are reprojected into the SAR image geometry. A random forest classifier is trained with the texture features on input and labels from the rasterized ice charts on output. Then the trained classifier is directly applied to the texture features from Sentinel-1 images in operational manner. Test results from two winter season dataset acquired over the Fram Strait and Barents Sea area showed that the classifier is capable of retrieving 3 generalized cover types (ice free, integrated first-year ice, old ice) with overall accuracies of 85 % and 5 cover types (ice free, new ice, young ice, first-year ice, old ice) with accuracy of 58 %. The errors are attributed both to incorrect manual classification on the ice charts and to the automated algorithm. We demonstrate the potential for near-real time service of ice type classification through an example of ice maps made from daily mosaiced images.