Sea Ice Concentration Estimation: Using Passive Microwave and SAR Data with Fully Convolutional Networks

Abstract

Sea ice concentration is of great interest to ship navigators and scientists who require regional ice cover understanding. Passive microwave data and image analysis charts are typically used to estimate ice concentration, but these have limitations. Estimates obtained from passive microwave data have coarse spatial resolution, may be biased due to weather filters that reduce atmospheric contamination, and often perform poorly in marginal ice zones. Image analysis charts are not as precise and subjective to analyst interpretation. Synthetic aperture radar (SAR) images are finer resolution satellite images that can be used to observe oceans. However, the complex interactions between the SAR signal and water and ice make it a difficult process to estimate sea ice concentration. Previous studies have found that deep learning is a viable avenue to estimate ice concentration from SAR images. In these studies, convolutional neural networks (CNNs) have been successful due to their ability to learn spatial features in their convolutional layers. To overcome the shortcomings of ice concentration estimation, we have uniquely implemented a U-net with SAR images as inputs and use estimates obtained from passive microwave data as training labels. The U-net, due to not being sensitive to patch size, is shown to be an improvement over the CNN models used in previous studies. Data augmentation and a mean absolute error (L1) loss function were applied as well as a curriculum learning method that introduces more open water and consolidated ice regions before incorporating marginal ice regions. The key objectives of this study are (a) to overcome shortcomings of using passive microwave data for model training and (b) to improve ice concentration predictions in marginal ice zones. Evaluating with image analysis charts, a mean absolute error of 7.18\% is achieved, which is lower than errors associated with estimation algorithms using passive microwave data alone. Through qualitative analysis, we also show instances where our proposed model has more precise estimates in the marginal ice zones than traditionally used passive microwave data based ice concentration retrieval algorithms. In this thesis, we also evaluate our proposed model to scenes from a higher latitude, which contain different ice types, to evaluate the ability of the model to extend to different scenes. It was found that model performance is worse in when extending to the new region. This suggests that the model was unable to learn features required to make accurate predictions on this dataset. Model performance may be improved if similar regions are included for training. Lastly, we evaluate the effects of using training labels from different passive microwave data based retrieval algorithms. We show that model performance is affected by the training labels used and using different training labels have unique benefits and shortcomings. To capitalize on benefits exclusive to specific training labels, we also attempt a method of staged training where the model is trained with one set of training labels before being trained with another set of training labels. We found that predictions made from models using staged training has qualities of the individual training labels used to train it. We also found that the final set of training labels used in staged training has the strongest impact on model predictions. From this experiment, we show that through staged training, we can teach deep learning models important information exclusive to different sets.