In this study, a convolutional neural network (CNN) is used to estimate sea ice concentration using synthetic aperture radar (SAR) scenes acquired during freeze-up in the Gulf of St. Lawrence on the east coast of Canada. The ice concentration estimates from the CNN are compared to those from a neural network (multi-layer perceptron or MLP) that uses hand-crafted features as input and a single layer of hidden nodes. The CNN is found to be less sensitive to pixel level details than the MLP and produces ice concentration that is less noisy and in closer agreement with that from image analysis charts. This is due to the multi-layer (deep) structure of the CNN, which enables abstract image features to be learned. The CNN ice concentration is also compared with ice concentration estimated from passive microwave brightness temperature data using the ARTIST sea ice (ASI) algorithm. The bias and RMS of the difference between the ice concentration from the CNN and that from image analysis charts is reduced as compared to that from either the MLP or ASI algorithm. Additional results demonstrate the impact of varying the input patch size, varying the number of CNN layers, and including the incidence angle as an additional input.

Sea ice observations through satellite imaging have led to advancements in environmental research, ship navigation, and ice hazard forecasting in cold regions. Machine learning and, recently, deep learning techniques are being explored by various researchers to process vast amounts of Synthetic Aperture Radar (SAR) data for detecting potential hazards in navigational routes. Detection of hazards such as sea ice floes in Marginal Ice Zones (MIZs) is quite challenging as the floes are often embedded in a multiscale ice cover composed of ice filaments and eddies in addition to floes. This study proposes a segmentation model tailored for detecting ice floes in SAR images. The model exploits the advantages of both convolutional neural networks and convolutional conditional random field (Conv-CRF) in a combined manner. The residual UNET (RES-UNET) computes expressive features to generate coarse segmentation maps while the Conv-CRF exploits the spatial co-occurrence pairwise potentials along with the RES-UNET unary/segmentation maps to generate final predictions. The whole pipeline is trained end-to-end using a dual loss function. This dual loss function is composed of a weighted average of binary cross entropy and soft dice loss. The comparison of experimental results with the conventional segmentation networks such as UNET, DeepLabV3, and FCN-8 demonstrates the effectiveness of the proposed architecture.

Ice concentration estimates are typically acquired from algorithms using passive microwave satellite data, and from image analysis charts, but these have limitations. Estimates acquired from passive microwave data have coarse spatial resolution, may have errors due to atmospheric contamination, and often perform poorly in marginal ice zones. Image analysis charts are not as precise, subject to analyst interpretation, and only available over specific geographic areas. We have implemented a U-net with synthetic aperture radar images as inputs and use ice concentration estimates retrieved 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 previous work with convolutional neural networks that use fully connected layers at the output. Data augmentation and an L1 loss function were applied along with a novel training scheme that leverages curriculum learning. In this training scheme, the model is first trained with samples from open water and consolidated ice regions before incorporating samples from marginal ice regions. In a tenfold cross validation experiment, we achieve 3-4% mean absolute error comparing to estimates using passive microwave data and observe curriculum learning models having more stable training. Predictions on four with-held SAR scenes with difficult ice conditions were evaluated with image analysis charts. A mean absolute error of 7.18% is achieved, which is lower than errors associated with passive microwave data alone. Qualitative improvements in marginal ice zone estimates are achieved, while still preserving smooth consolidated ice regions, and openings in ice cover.

Bayesian neural networks (BNNs) have been demonstrated to be effective in accurate retrieval of sea ice concentration (SIC) from multi-source data, while providing estimates of uncertainty, which are essential for downstream services. However, uncertainty obtained by BNNs are intrinsically uncalibrated, which indicates that it may not correlate well with model error. To address this issue, we investigate a new approach that combines an auxiliary prediction interval (PI) estimator with the BNN-based SIC mean estimator to develop a well-calibrated SIC retrieval model that is both accurate and reliable. We adopt a training strategy called “uncertainty matching" to train the model, which ensures that the estimated uncertainties match the estimated PIs. We use a subset of AMSR2 brightness temperature data and ERA5 atmospheric data collected from 2014 to 2015 in the Baffin Bay area as input features of the model. Comparison between model inference and SIC labels obtained from the enhanced NASA Team (NT2) algorithm shows that the proposed approach is able to produce well-calibrated uncertainty with more accurate predictions in marginal ice zones.

Synthetic aperture radar (SAR) sea ice imagery is a promising source of data for sea ice data assimilation. Classification of SAR sea ice imagery into ice and water is of particular relevance due to its relationship with ice concentration, a key variable in sea ice data assimilation systems. With increasing volumes of SAR data, automated methods to carry out these classifications are of particular importance. Although several automated approaches have been proposed, none look at the impact of including an estimate of uncertainty of the model parameters and input features on the classification output. This article uses an established database of SAR image features to train a multilayer perceptron (MLP) neural network to classify pixel locations as either ice, water, or unknown. The classification accuracies are benchmarked using a recently developed logistic regression approach for the same database. The two methods are found to be comparable. The MLP approach is then enhanced to allow uncertainty to be estimated at each pixel location. Following methods proposed in the deep learning community, two kinds of uncertainty are considered. The first, epistemic uncertainty, is that due to uncertainty in the MLP weights. The second kind of uncertainty, aleatoric uncertainty, is that which cannot be explained by the model, and is therefore associated with the input data. It is found that including these uncertainties in the MLP models reduces their accuracies slightly, but also reduces misclassification rates. This is of particular importance for data assimilation applications, where misclassifications could severely degrade the analysis.

In recent years, the adoption of deep learning (DL) techniques for predicting sea ice concentration (SIC) given both passive microwave (PM) data and reanalysis data has seen a growing interest. For use in downstream services, these sea ice concentration estimates should be accompanied by uncertainty estimates. To provide these estimates, we utilize a heteroscedastic Bayesian neural network (HBNN), which can estimate both model (epistemic) and data (aleatoric) uncertainty. We use both PM and atmospheric data as our input features, and demonstrate that both are needed for accurate SIC estimates. Results show that over an annual cycle, the months of melt onset, such as April, May, and June produce the highest uncertainties relative to other months, with total (epistemic + aleatoric) uncertainties of approximately 20%, while areas in the marginal ice zone, contributed highest total uncertainty of 25% spatially. When considering an average over the test year, the level of uncertainty due to the data (aleatoric) is consistent with other studies, at 10%-15%. The advantage of our approach is that the uncertainties are specific to the data instance, and both model and data uncertainty are estimated.