Sea Ice Concentration Estimates Research Articles

Sea ice concentration satellite retrievals influenced by surface changes due to warm air intrusions: A case study from the MOSAiC expedition

Warm air intrusions over Arctic sea ice can change the snow and ice surface conditions rapidly and can alter sea ice concentration (SIC) estimates derived from satellite-based microwave radiometry without altering the true SIC. Here we focus on two warm moist air intrusions during the Multidisciplinary drifting Observatory for the Study of Arctic Climate (MOSAiC) expedition that reached the research vessel Polarstern in mid-April 2020. After the events, SIC deviations between different satellite products, including climate data records, were observed to increase. Especially, an underestimation of SIC for algorithms based on polarization difference was found. To examine the causes of this underestimation, we used the extensive MOSAiC snow and ice measurements to model computationally the brightness temperatures of the surface on a local scale. We further investigated the brightness temperatures observed by ground-based radiometers at frequencies 6.9 GHz, 19 GHz, and 89 GHz. We show that the drop in the retrieved SIC of some satellite products can be attributed to large-scale surface glazing, that is, the formation of a thin ice crust at the top of the snowpack, caused by the warming events. Another mechanism affecting satellite products, which are mainly based on gradient ratios of brightness temperatures, is the interplay of the changed temperature gradient in the snow with snow metamorphism. From the two analyzed climate data record products, we found that one was less affected by the warming events. The low frequency channels at 6.9 GHz were less sensitive to these snow surface changes, which could be exploited in future to obtain more accurate retrievals of sea ice concentration. Strong warm air intrusions are expected to become more frequent in future and thus their influence on SIC algorithms will increase. In order to provide consistent SIC datasets, their sensitivity to warm air intrusions needs to be addressed.

Super-Resolution-Aided Sea Ice Concentration Estimation From AMSR2 Images by Encoder–Decoder Networks With Atrous Convolution

Passive microwave data is an important data source for the continuous monitoring of Arctic-wide sea ice concentration (SIC). However, its coarse spatial resolution leads to blurring effects at the ice-water divides, resulting in the great challenges of fine-scale and accurate SIC estimation, especially for regions with low SIC. Besides, the SIC derived by operational algorithms using high-frequency passive microwave observations has great uncertainties in open water or marginal ice zones due to atmospheric effects. In this article, a novel framework is proposed to achieve accurately SIC estimation with improved spatial details from original low-resolution Advanced Microwave Scanning Radiometer 2 (AMSR2) images, with joint the super-resolution (SR) and SIC estimation network. Based on the SR network, the spatial resolution of original AMSR2 images can be improved by four times, benefiting to construct AMSR2 SR features with more high-frequency information for SIC estimation. The SIC network with an encoder–decoder structure and atrous convolution, is employed to accurately perform the SIC retrieval by considering the characteristics of passive microwave images in the Arctic sea ice region. Experimental results show that the proposed SR-Aided SIC estimation approach can generate accurate SIC with more detailed sea ice textures and much sharper sea ice edges. With respect to MODIS SIC products distributed in Arctic scale, the proposed model achieves a root-mean-square error (RMSE) of 5.94% and mean absolute error (MAE) of 3.04%, whereas the Arctic Radiation and Turbulence Interaction Study (ARTIST) Sea Ice (ASI) SIC results have three and two times greater values of RMSE and MAE.

Predicting Sea Ice Concentration With Uncertainty Quantification Using Passive Microwave and Reanalysis Data: A Case Study in Baffin Bay

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.

Baltic Sea Ice Concentration Estimation From C-Band Dual-Polarized SAR Imagery by Image Segmentation and Convolutional Neural Networks

In this study application of convolutional neural networks (CNNs) preceded by synthetic aperture radar (SAR), image segmentation for sea ice concentration (SIC) estimation over the Baltic Sea from dual-polarized C-band SAR imagery is studied. Three algorithm variants were studied and trained using FMI ice chart SIC or a synthetic SIC dataset with different SIC values generated by combining pure open water and sea ice blocks by applying binary masks. The first two algorithm variants were trained using only open water and fully ice-covered patches, based on the FMI ice charts, and they had a similar CNN structure. These two algorithm variants differ only in deriving the segmentwise SIC from the CNN output. In the third algorithm, variant synthetic SIC data derived as a mixture of open water and fully ice-covered ice patches according to the ice charts were used in training. The estimation results were evaluated with respect to the FMI ice chart SIC for an independent test dataset. The results were very encouraging for operational purposes and significantly better than for our earlier SAR-based SIC estimation algorithms. The algorithm version trained with synthetic SIC data clearly outperformed the two other algorithm versions and our earlier SIC estimation results based on dual-polarized SAR, using independent FMI ice chart SIC as a reference.

Simulated Geophysical Noise in Sea Ice Concentration Estimates of Open Water and Snow-Covered Sea Ice

Sea ice concentration algorithms using brightness temperatures (<i>T_B</i>) from satellite microwave radiometers are used to compute sea ice concentration (<i>c_ice</i>), sea ice extent, and generate sea ice climate data records (CDRs). Therefore, it is important to minimize the sensitivity of <i>c_ice</i> estimates to geophysical noise caused by snow/sea ice thermal microwave emission signature variations, and presence of water vapor and clouds in the atmosphere and/or near-surface winds. In this study, we investigate the effect of geophysical noise leading to systematic <i>c_ice</i> biases and affecting <i>c_ice</i> standard deviations (STD) using simulated top of the atmosphere (TOA) <i>T_B</i>s over open water and 100 % sea ice. We consider three case studies for the Arctic and the Antarctic and eight different <i>c_ice</i> algorithms, representing different families of algorithms based on the selection of channels and methodologies. Our simulations show that, over open water and low <i>c_ice</i>, algorithms using gradients between V-polarized 19 GHz and 37 GHz <i>T_B</i>s shows the lowest sensitivity to the geophysical noise, while the algorithms exclusively using near 90 GHz channels have by far the highest sensitivity. Over sea ice, the atmosphere plays a much smaller role than over open water and the <i>c_ice</i> STD for all algorithms is smaller than over open water. The hybrid and low frequency (6 GHz) algorithms have the lowest sensitivity to noise over sea ice, while the polarization type of algorithms have the highest noise levels.

Joint Total Variation With Nonnegative Constrained Least Square for Sea Ice Concentration Estimation in Low Concentration Areas of Antarctica

Sea ice concentration (SIC) is an indispensable parameter for the study of polar sea ice. The existing methods can obtain accurate SICs for most situations, but they usually perform poorly in low SIC regions because of the spatial differences in the neighboring pixels induced by the discontinuity of the sea ice cover. In this letter, to cope with the difficulty of this problem, an improved SIC estimation method is proposed to retrieve SIC, focusing on low SIC regions. The proposed method introduces the spatial relationships into SIC estimation by employing a total variation (TV) regularizer. Moreover, nonnegative constrained least squares (NCLS) is used to derive the optimal solutions from the SIC estimation equation. Verification was conducted in low SIC regions (0%-50%) of the Antarctic utilizing ship-based in situ data and the Moderate Resolution Imaging Spectroradiometer (MODIS), and the results were compared with those of some of the mature methods. The results indicated that the proposed method can obtain a superior accuracy with a smaller root-mean-square error (RMSE) (6.0%-14.61%) than the other algorithms in low SIC regions. Furthermore, the proposed method can accurately estimate the SIC of both first-year ice and multiyear ice. The findings of this study confirm the need to consider the spatial relationships in the processing of SIC estimation.

Development of adjoint-based ocean state estimation for the Amundsen and Bellingshausen seas and ice shelf cavities using MITgcm–ECCO (66j)

Abstract. The Antarctic coastal ocean impacts sea level rise, deep-ocean circulation, marine ecosystems, and the global carbon cycle. To better describe and understand these processes and their variability, it is necessary to combine the sparse available observations with the best-possible numerical descriptions of ocean circulation. In particular, high ice shelf melting rates in the Amundsen Sea have attracted many observational campaigns, and we now have some limited oceanographic data that capture seasonal and interannual variability during the past decade. One method to combine observations with numerical models that can maximize the information extracted from the sparse observations is the adjoint method, a.k.a. 4D-Var (4-dimensional variational assimilation), as developed and implemented for global ocean state estimation by the Estimating the Circulation and Climate of the Ocean (ECCO) project. Here, for the first time, we apply the adjoint-model estimation method to a regional configuration of the Amundsen and Bellingshausen seas, Antarctica, including explicit representation of sub-ice-shelf cavities. We utilize observations available during 2010–2014, including ship-based and seal-tagged CTD measurements, moorings, and satellite sea-ice concentration estimates. After 20 iterations of the adjoint-method minimization algorithm, the cost function, here defined as a sum of the weighted model–data difference, is reduced by 65 % relative to the baseline simulation by adjusting initial conditions, atmospheric forcing, and vertical diffusivity. The sea-ice and ocean components of the cost function are reduced by 59 % and 70 %, respectively. Major improvements include better representations of (1) Winter Water (WW) characteristics and (2) intrusions of modified Circumpolar Deep Water (mCDW) towards the Pine Island Glacier. Sensitivity experiments show that ∼40 % and ∼10 % of improvements in sea ice and ocean state, respectively, can be attributed to the adjustment of air temperature and wind. This study is a preliminary demonstration of adjoint-method optimization with explicit representation of ice shelf cavity circulation. Despite the 65 % cost reduction, substantial model–data discrepancies remain, in particular with annual and interannual variability observed by moorings in front of the Pine Island Ice Shelf. We list a series of possible causes for these residuals, including limitations of the model, the optimization methodology, and observational sampling. In particular, we hypothesize that residuals could be further reduced if the model could more accurately represent sea-ice concentration and coastal polynyas.

Sea Ice Concentration Estimation: Using Passive Microwave and SAR Data With a U-Net and Curriculum Learning

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.

Retrieval of ice/water observations from synthetic aperture radar imagery for use in lake ice data assimilation

High-resolution lake ice/water observations retrieved from satellite imagery through efficient, automated methods can provide critical information to lake ice forecasting systems. Synthetic aperture radar (SAR) data is well-suited to this purpose due to its high spatial resolution (approximately 50 m). With recent increases in the volume of SAR data available, the development of automated retrieval methods for these data is a priority for operational centres. However, automated retrieval of ice/water data from SAR imagery is difficult, due to ambiguity in ice and open water signatures, both in terms of image tone and in terms of parameterized texture features extracted from these images. Convolutional neural networks (CNNs) can learn features from imagery in an automated manner, and have been found effective in previous studies on sea ice concentration estimation from SAR. In this study the use of CNNs to retrieve ice/water observations from dual-polarized SAR imagery of two of the Laurentian Great Lakes, Lake Erie and Lake Ontario, is investigated. For data assimilation, it is crucial that the retrieved observations are of high quality. To this end, quality control measures based on the uncertainty of the CNN output to eliminate incorrect retrievals are discussed and demonstrated. The quality control measures are found to be effective in both dual-polarized and single-polarized retrievals. The ability of the CNN to downscale the coarse resolution training labels is demonstrated qualitatively.

The Arctic Ocean Observation Operator for 6.9 GHz (ARC3O) – Part 2: Development and evaluation

Abstract. The observational uncertainty in sea ice concentration estimates from remotely sensed passive microwave brightness temperatures is a challenge for reliable climate model evaluation and initialization. To address this challenge, we introduce a new tool: the Arctic Ocean Observation Operator (ARC3O). ARC3O allows us to simulate brightness temperatures at 6.9 GHz at vertical polarization from standard output of an Earth System Model. To evaluate sources of uncertainties when applying ARC3O, we compare brightness temperatures simulated by applying ARC3O on three assimilation runs of the MPI Earth System Model (MPI-ESM), assimilated with three different sea ice concentration products, with brightness temperatures measured by the Advanced Microwave Scanning Radiometer Earth Observing System (AMSR-E) from space. We find that the simulated and observed brightness temperatures differ up to 10 K in the period between October and June, depending on the region and the assimilation run. We show that these discrepancies between simulated and observed brightness temperature can be attributed mainly to the underlying observational uncertainty in sea ice concentration and, to a lesser extent, to the data assimilation process, rather than to biases in ARC3O itself. In summer, the discrepancies between simulated and observed brightness temperatures are larger than in winter and locally reach up to 20 K. This is caused by the very large observational uncertainty in summer sea ice concentration and the melt pond parametrization in MPI-ESM, which is not necessarily realistic. ARC3O is therefore capable of realistically translating the simulated Arctic Ocean climate state into one observable quantity for a more comprehensive climate model evaluation and initialization.

Assessment of the Stability of Passive Microwave Brightness Temperatures for NASA Team Sea Ice Concentration Retrievals

Gridded passive microwave brightness temperatures (TB) from special sensor microwave imager and sounder (SSMIS) instruments on three different satellite platforms are compared in different years to investigate the consistency between the sensors over time. The orbits of the three platforms have drifted over their years of operation, resulting in changing relative observing times that could cause biases in TB estimates and near-real-time sea ice concentrations derived from the NASA Team algorithm that are produced at the National Snow and Ice Data Center. Comparisons of TB histograms and concentrations show that there are small mean differences between sensors, but variability within an individual sensor is much greater. There are some indications of small changes due to orbital drift, but these are not consistent across different frequencies. Further, the overall effect of the drift, while not definitive, is small compared to the intra- and interannual variability in individual sensors. These results suggest that, for near-real-time use, the differences in the sensors are not critical. However, for long-term time series, even the small biases should be corrected for. The strong day-to-day, seasonal, and interannual variability in TB distributions indicate that time-varying algorithm coefficients in the NASA team algorithm would lead to improved, more consistent sea ice concentration estimates.

Improved Estimation of Proxy Sea Surface Temperature in the Arctic

AbstractArctic sea surface temperatures (SSTs) are estimated mostly from satellite sea ice concentration (SIC) estimates. In regions with sea ice the SST is the temperature of open water or of the water under the ice. A number of different proxy SST estimates based on SIC have been developed. In recent years more Arctic quality-control buoy SSTs have become available, allowing better validation of different estimates and the development of improved proxy estimates. Here proxy SSTs from different approaches are evaluated and an improved proxy SST method is shown. The improved proxy SSTs were tested in an SST analysis, and showed reduced bias and random errors compared to the Arctic buoy SSTs. Almost all reduction in errors is in the warm melt season. In the cold season the SIC is typically high and all estimates tend to have low errors. The improved method will be incorporated into an operational SST analysis.

Reconstructing spring sea ice concentration in the Chukchi Sea over recent centuries: insights into the application of the PIP25 index

In this study, we aimed to reconstruct spring (April–June) sea ice changes in the western Arctic Ocean over recent centuries (ca. the last 250 years) by measuring biomarker distributions in a multicore (ARA01B-03MUC) retrieved from the Chukchi Shelf region and to evaluate outcomes against known or modelled estimates of sea ice conditions. Specifically, we analyzed for the Arctic sea ice proxy IP25 and assessed the suitability of a further highly branched isoprenoid (HBI) lipid (HBI III), epi-brassicasterol, and dinosterol as complementary biomarkers for use with the so-called phytoplankton marker-IP25 index (PIP25; PIIIIP25, PBIP25, and PDIP25, respectively). The presence of IP25 throughout core ARA01B-03MUC confirms the occurrence of seasonal sea ice at the study site over recent centuries. From a semi-quantitative perspective, all three PIP25 indices gave different trends, with some dependence on the balance factor c, a term used in the calculation of the PIP25 index. PIIIIP25-derived spring sea ice concentration (SpSIC) estimates using a c value of 0.63, determined previously from analysis of Barents Sea surface sediments, were likely most reliable, since SpSIC values were high throughout the record (SpSIC > 78%), consistent with the modern context for the Chukchi Sea and the mean SpSIC record of the 41 CMIP5 climate models over recent centuries. PBIP25-based SpSIC estimates were also high (SpSIC 108%−127%), albeit somewhat over-estimated, when using a c value of 0.023 obtained from a pan-Arctic distribution of surface sediments. In contrast, PDIP25 values using a pan-Arctic c value of 0.11, and PIP25 data based on the mean biomarker concentrations from ARA01B-03MUC, largely underestimated sea ice conditions (SpSIC as low as 13%), and exhibited poor agreement with instrumental records or model outputs. On the other hand, PBIP25 values using a c factor based on mean IP25 and epi-brassicasterol concentrations exhibited a decline towards the core top, which resembled recent decreasing changes in summer sea ice conditions for the Chukchi Sea; however, further work is needed to test the broader spatial generality of this observation.

Sea Ice Remote Sensing Using GNSS-R: A Review

Knowledge of sea ice is critical for offshore oil and gas exploration, global shipping industries, and climate change studies. During recent decades, Global Navigation Satellite System-Reflectometry (GNSS-R) has evolved as an efficient tool for sea ice remote sensing. In particular, thanks to the availability of the TechDemoSat-1 (TDS-1) data over high-latitude regions, remote sensing of sea ice based on spaceborne GNSS-R has been rapidly growing. The goal of this paper is to provide a review of the state-of-the-art methods for sea ice remote sensing offered by the GNSS-R technique. In this review, the fundamentals of these applications are described, and their performances are evaluated. Specifically, recent progress in sea ice sensing using TDS-1 data is highlighted including sea ice detection, sea ice concentration estimation, sea ice type classification, sea ice thickness retrieval, and sea ice altimetry. In addition, studies of sea ice sensing using airborne and ground-based data are also noted. Lastly, applications based on various platforms along with remaining challenges are summarized and possible future trends are explored. In this review, concepts, research methods, and experimental techniques of GNSS-R-based sea ice sensing are delivered, and this can benefit the scientific community by providing insights into this topic to further advance this field or transfer the relevant knowledge and practice to other studies.

Sea Ice Climate Normals for Seasonal Ice Monitoring of Arctic and Sub-Regions

The climate normal, that is, the latest three full-decade average, of Arctic sea ice parameters is useful for baselining the sea ice state. A baseline ice state on both regional and local scales is important for monitoring how the current regional and local states depart from their normal to understand the vulnerability of marine and sea ice-based ecosystems to the changing climate conditions. Combined with up-to-date observations and reliable projections, normals are essential to business strategic planning, climate adaptation and risk mitigation. In this paper, monthly and annual climate normals of sea ice parameters (concentration, area, and extent) of the whole Arctic Ocean and 15 regional divisions are derived for the period of 1981–2010 using monthly satellite sea ice concentration estimates from a climate data record (CDR) produced by NOAA and the National Snow and Ice Data Center (NSIDC). Basic descriptions and characteristics of the normals are provided. Empirical Orthogonal Function (EOF) analysis has been utilized to describe spatial modes of sea ice concentration variability and how the corresponding principal components change over time. To provide users with basic information on data product accuracy and uncertainty, the climate normal values of Arctic sea ice extents (SIE) are compared with that of other products, including a product from NSIDC and two products from the Copernicus Climate Change Service (C3S). The SIE differences between different products are in the range of 2.3–4.5% of the CDR SIE mean. Additionally, data uncertainty estimates are represented by using the range (the difference between the maximum and minimum), standard deviation, 10th and 90th percentiles, and the first, second, and third quartile distribution of all monthly values, a distinct feature of these sea ice normal products.

Impact of intermediate ice concentration training data on sea ice concentration estimates from a convolutional neural network

ABSTRACTConvolutional neural networks (CNNs) are being used increasingly for classification and regression tasks in remote sensing. However, a challenge in the remote sensing field is having a sufficient quantity of data to train a CNN, in particular, for conditions that are observed less frequently. In this study, a CNN is trained and tested to estimate sea ice concentration from synthetic aperture radar imagery (SAR). We first investigate the importance of including samples of intermediate ice concentration in the training data, which correspond to samples from the marginal ice zone (MIZ). In the present study, ice concentration from image analysis charts is used to provide training data labels. The MIZ for these ice charts are believed to be less accurate than those of the high and low concentrations, but nevertheless, our results support including these samples in the training data set. Additional experiments are then carried out increasing the number of MIZ sample, both from similar regions and a different ice region. It is found the MIZ is represented best in the test data when more samples from a similar region are included. Overall the results improve upon earlier studies, increasing the classification accuracy of the MIZ from 0.66 to 0.74.

Disentangling the coupling between sea ice and tundra productivity in Svalbard

The rapid decline in Arctic sea ice poses urgent questions concerning its ecological effects, such as on tundra terrestrial productivity. However, reported sea ice/terrestrial productivity linkages have seldom been constrained, and the mechanism governing them remains elusive, with a diversity of spatial scales and metrics proposed, at times in contradiction to each other. In this study, we use spatially explicit remotely sensed sea ice concentration and high-resolution terrestrial productivity estimates (Normalised Difference Vegetation Index, NDVI) across the Svalbard Archipelago to describe local/sub-regional and large-scale components of sea ice/terrestrial productivity coupling. Whereas the local/sub-regional component is attributed to sea breeze (cold air advection from ice-covered ocean onto adjacent land during the growing season), the large-scale component might reflect co-variability of sea ice and tundra productivity due to a common forcing, such as large-scale atmospheric circulation (North Atlantic Oscillation, NAO). Our study clarifies the range of mechanisms in sea ice/terrestrial productivity coupling, allowing the generation of testable hypotheses about its past, present, and future dynamics across the Arctic.

Neural Networks Based Sea Ice Detection and Concentration Retrieval From GNSS-R Delay-Doppler Maps

In this paper, a neural networks (NN) based scheme is presented for detecting sea ice and retrieving sea ice concentration (SIC) from global navigation satellite system reflectometry delay-Doppler maps (DDMs). Here, a multilayer perceptron neural network with back-propagation learning is adopted. In practice, two NN were separately developed for sea ice detection and concentration retrieval purposes. In the training phase, DDM pixels were employed as an input. The SIC data obtained by Nimbus-7 SMMR and DMSP SSM/I-SSMIS sensors were used as the target data, which were also regarded as ground-truth data in this paper. After the training process using a dataset collected around February 4, 2015, these networks were used to produce corresponding detection and concentration estimation for other four sets of DDM data, which were collected around February 12, 2015, February 20, 2015, March 16, 2015, and April 17, 2015, respectively. Results show high accuracy in sea ice detection and concentration estimation with DDMs using the proposed scheme. On average, the accuracy for sea ice detection is about 98.4%. In terms of estimated SIC, the mean absolute error is less than 9%, whereas the correlation coefficient is as high as 0.93 compared with the reference data. It was also found that low sea state and wind speed could lead to an overestimation of SIC for partially ice-covered region.

Sea Ice Concentration Estimation during Freeze-Up from SAR Imagery Using a Convolutional Neural Network

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.

Validation of sea ice models using an uncertainty‐based distance metric for multiple model variables

AbstractWe implement a variance‐based distance metric (Dn) to objectively assess skill of sea ice models when multiple output variables or uncertainties in both model predictions and observations need to be considered. The metric compares observations and model data pairs on common spatial and temporal grids improving upon highly aggregated metrics (e.g., total sea ice extent or volume) by capturing the spatial character of model skill. The Dn metric is a gamma‐distributed statistic that is more general than the χ2 statistic commonly used to assess model fit, which requires the assumption that the model is unbiased and can only incorporate observational error in the analysis. The Dn statistic does not assume that the model is unbiased, and allows the incorporation of multiple observational data sets for the same variable and simultaneously for different variables, along with different types of variances that can characterize uncertainties in both observations and the model. This approach represents a step to establish a systematic framework for probabilistic validation of sea ice models. The methodology is also useful for model tuning by using the Dn metric as a cost function and incorporating model parametric uncertainty as part of a scheme to optimize model functionality. We apply this approach to evaluate different configurations of the standalone Los Alamos sea ice model (CICE) encompassing the parametric uncertainty in the model, and to find new sets of model configurations that produce better agreement than previous configurations between model and observational estimates of sea ice concentration and thickness.