Classification of sea ice types in Sentinel-1 synthetic aperture radar images

Abstract

A new Sentinel-1 image-based sea ice classification algorithm using a machine-learning-based model trained in a semi-automated manner is proposed to support daily ice charting. Previous studies mostly rely on manual work in selecting training and validation data. We show that the readily available ice charts from the operational ice services can reduce the amount of manual work in preparation of large amounts of training/testing data. Furthermore, they can feed highly reliable data to the trainer by indirectly exploiting the best ability of the sea ice experts working at the operational ice services. The proposed scheme has two phases: training and operational. Both phases start from the removal of thermal, scalloping, and textural noise from Sentinel-1 data and calculation of grey 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 operationally. Test results from the two datasets spanning winter (January？March) and summer (June？August) seasons acquired over the Fram Strait and the Barents Sea showed that the classifier is capable of retrieving three generalized cover types (open water, mixed first-year ice, old ice) with overall accuracies of 87% and 67%in winter and summer seasons, respectively. For the summer season, the classifier failed in distinguishing mixed first-year ice from old ice with accuracy of only 12 %; however, it performed rather like an ice？water discriminator with high accuracy of 98% as the misclassification between the mixed first-year ice and old ice was between them. The accuracy for five cover types (open water, new ice, young ice, first-year ice, old ice) in the winter season was 60 %. The errors are attributed both to incorrect manual classification on the ice charts and to the semi-automated algorithm. Finally, we demonstrate the potential for near-real-time service of the ice map using daily mosaicked Sentinel-1 images.