Automatic detection of Arctic polynyas using hybrid supervised-unsupervised deep learning

Abstract. Polynyas are areas with no- or thin-ice within the ice pack. They play a crucial role for the Earth system, from deep water to cloud formation, causing large gas exchanges, and acting as hotspots for marine life. Yet their monitoring in the Arctic is challenging because polynya detection is non-trivial, owing to the Arctic's complex geometry. Recently, a labelled dataset was released in which daily winter Arctic sea ice concentration since 1978 was turned into a polynya mask. After oversampling to reduce the class imbalance from 0.1 to 2.5 %, we use this labelled dataset to train a Unet autoencoder to detect polynya pixels in daily sea ice concentration images. We further filter out the false positives in the marginal ice zone using an unsupervised Gaussian Mixture Model classifier. False negatives are virtually absent from 2012 onwards, when noise in the labelled dataset is reduced by combining ice concentration and thickness masks. False positives exhibit a significant trend with time and anticorrelation with the Arctic sea ice area. Coupled with our expert assessment of individual images, we argue that most ``false positives'' are in fact correct, detecting patterns of reduced ice within the changing, more unpredictable ice cover that the rigid traditional methods with fixed thresholds cannot identify. We also successfully apply our trained model to detect polynyas in daily and monthly climate model output at low computing costs. As Arctic sea ice continues to decrease, pushing traditional methods to their limits, we expect such machine-learning methods to become the norm.