Presentation Abstracts

Skillful sea ice forecasts from days to years ahead are becoming increasingly important for the operation and planning of human activities in the Arctic. Here we analyze the potential predictability of the Arctic sea ice edge in six climate models. We introduce the integrated ice-edge error (IIEE), a user-relevant verification metric defined as the area where the forecast and the “truth” disagree on the ice concentration being above or below 15%. The IIEE lends itself to decomposition into an absolute extent error, corresponding to the common sea ice extent error, and a misplacement error. We find that the often-neglected misplacement error makes up more than half of the climatological IIEE. In idealized forecast ensembles initialized on 1 July, the IIEE grows faster than the absolute extent error. This means that the Arctic sea ice edge is less predictable than sea ice extent, particularly in September, with implications for the potential skill of end-user relevant forecasts.

We assess prediction skill in the Arctic and Antarctic in the prototype MiKlip prediction system that is based on the Max Planck Institute Earth System Model (MPI-ESM). The three-dimensional temperature and salinity fields in the ocean in the decadal hindcasts are initialized from GECCO2 and ORA-S4 and the atmosphere is initialized with ERA40 and ERA-Interim, while sea ice is not initialized. We show that the initialization can improve the predictability of sea surface temperatures and air temperature in certain polar regions. In the Arctic these areas include the North Pacific Ocean, the North Atlantic Ocean and the East-Siberian Sea. The skill improvement mainly derives from a better representation of observed variability patterns, since the initialization of oceanic and atmospheric parameters enables the simulations to start from the correct phase of natural variability. We additionally analyse prediction skill in sea-ice variables using different metrics.

Stand-alone sea ice simulations with a physical based melt pond model reveal a strong correlation between the simulated spring pond fraction and the observed as well as simulated September sea ice extent for the period 1979 to 2015. This is explained by a positive feedback mechanism: more ponds reduce the albedo; a lower albedo causes more melting; more melting increases pond fraction. We implemented the Los Alamos sea ice model CICE 5 including our physical based melt pond model into the latest version of the Hadley Centre coupled climate model, HadGEM3. The model surface shortwave radiation scheme has been adjusted to account for pond fraction and depth. We performed three 55-year HadGEM3 simulations with constant external forcing for the years 1985, 2010 and 2035. In all three simulations we find a strong correlation between the April/May pond fraction and the September sea ice conditions. We demonstrate that spring melt ponds are an important driver for summer ice melt and enable us to make skilful predictions for the consequent September sea ice in most Arctic regions for current and future climate conditions.

The indigenous populations of the Arctic region possess an immense knowledge of their environments, based on centuries of living close to nature. Living in and from the richness and variety of complex ecosystems, they have an in-depth understanding of the properties of plants and animals, the functioning of ecosystems and the techniques for using and managing them in a sustainable way. However, when the subject is sea ice prediction and ice retreat there are potential hazards triggering the exposure of communities to extreme climatic events dependent on specific geographical patterns that, consequently, restrict their economic, social and cultural activities. Traditional knowledge can benefit immensely from western science research and communication to minimise the exposure of Arctic local communities to climatic adversities. For this reason, it is fundamental that the sea ice predictions research teams design and provide effective communication strategies to translate relevant data of research findings to address local social needs and priorities. In this sense, it is essential to make Arctic science accessible to indigenous communities by improving as much as possible cross-community communication, multi-cultural dialogue and the communities’ understanding of useful data about ice, ocean and atmospheric processes. Cross-cultural understanding is fundamental to investigate and transplant concepts from the indigenous world to the western culture and vice-versa in order to contribute effectively to more consistent patterns of adaptation.

A number of previous studies have provided insights on the mechanisms driving predictability of sea ice on seasonal to decadal timescales. Traditionally, many of these studies were focused on large-scale sea ice properties, such as the total northern hemisphere ice extent. These studies were important contributions to our understanding of factors that influence sea ice forecasting skill. However, users of sea ice forecasts have the need for more specific regional and temporal information, for example on the ice edge location or the timing of ice retreat in various locations. Here I will discuss some of these stakeholder needs and some recent work that explores the predictability of related properties. I will also present some new analysis that uses simulations from the Community Earth System Model (CESM) to assess mechanisms affecting predictability of some societally-relevant sea ice properties.

Variations in Arctic sea ice thickness and volume affect regional heat and freshwater budgets and patterns of atmospheric circulation at lower latitudes. Estimates of Arctic sea ice thickness and volume can be obtained from observation and model outputs, but so far, an in-depth comparison of these outputs has not occurred. Such a comparison is crucial for the validation of models that form the basis of future Arctic and global climate projections. In order for an in-depth comparison to be successful, expert knowledge is required of numerous model configurations, as well as the observation methods used to estimate sea ice thickness. This presentation will aim to encourage collaboration between modellers and observers by summarising the method used to calculate sea ice thickness from CryoSat-2 radar altimeter satellite observations, and the difficulties and limitations associated with obtaining these.

It is also our hope that the availability of Arctic-wide sea ice thickness data, especially in near real time (NRT), will enable evaluation and improved skill in the prediction of sea ice thickness distributions by climate models. Although it is possible to calculate Arctic sea ice thickness using measurements acquired by CryoSat-2, the latency of the nominal data set typically exceeds 1-2 months due to the time required to determine precise satellite orbits. The second part of this presentation will introduce our new NRT Arctic sea ice thickness dataset, based on preliminary orbits, which provides Arctic-wide sea ice thickness estimates just three days after acquisition from the satellite and is publicly available. A NRT sea ice thickness estimate is delivered, on average, within 64 km of each location in the Arctic every 14 days.

In this talk I will provide a quick overview on Arctic-midlatitdues climate linkage especially by way of stratosphere-troposphere coupling. Evidence is emerging from both observations and model results that under present climate conditions Arctic sea-ice loss results in modulated propagation of planetary-scale Rossby waves, leading to changes in the strength of polar vortex with consequent surface signals. Having gained some knowledge on this process we are beginning to put various components together for sub-seasonal to sub-decadal prediction, for which sea ice serves as both forcing and an output. In the second part, I will discuss about how this implies in sea ice prediction, in particular in the areas of turbulent heat flux and sea-ice regional variability.

We identified an atmospheric circulation dipole anomaly in the Pacific-Arctic sector and we showed that it is related to the following September sea ice in the Beaufort-Chukchi Sea, using sea ice observations and model-generated data from PIOMAS (Pan-Arctic Ice-Ocean Modeling and Assimilation System), and the European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis. The dipole anomaly is the second leading EOF mode of spring (April-June) sea level pressure (SLP) in the Pacific-Arctic (600N-900N, 1200E-1200W) and it accounts for 21% of the variance. This dipole anomaly, which we denote as the Pacific-Arctic Dipole, has a positive anomaly in the Beaufort Sea and a negative anomaly extending from East Siberia to Northwest America, and it exhibits co-variance with the Beaufort High and the Aleutian Low. The dipole mode also reflects the re-distribution of cyclone activities in the Pacific-Arctic sector, with fewer cyclones in the Beaufort Sea and central Arctic and more cyclones in the subpolar Pacific. We also define a cyclone activity dipole index using the difference between the cyclone system density in the Arctic (700N-900N, 900E-600W) and that in the subpolar Pacific (500N-600N, 1500E-2100W), which is highly correlated with the time series of the Pacific-Arctic Dipole. We found that the spring Pacific-Arctic Dipole accounts for about 20% of the interannual variance of the following summer SIC in the Beaufort-Chukchi Sea. A positive Pacific-Arctic Dipole has an enhanced Beaufort high and the resulting intensified eastern winds in the Beaufort Sea lead to enhanced ice advection and weakened sea ice thickness. Ice is advected into warm Alaskan coastal waters which results in extensive melting. Moreover, less cyclone activity leads to less middle level cloud cover and less warm air advection in the Beaufort Sea and central Arctic, which causes a net surface heat flux gain, resulting in further reductions in sea ice.

We apply a Lagrangian sea ice tracking software to the problem of transport between the exclusive economic zones (EEZs) of the coastal Arctic nations over about 3 decades: 1979-2011. Lagrangian tracking is combined with sea ice concentration maps to distinguish sea ice formation and melt events from ice drift. We describe trends in formation and melt locations and sea-ice transport pathways. Most ice, ca. 60%, melts within 100 km of where it is formed; only ca. 15% escapes from its "native" EEZ. Of the ice that does leave its region of formation, the majority is ultimately exported from the Arctic through Fram Strait, melting in the East Greenland Current. While only a small fraction of sea ice is exported from one nation’s exclusive economic zone to another’s, this transport nonetheless amounts to tens of thousands of square kilometers of sea ice. Over the last three decades, as the ice has thinned, ice travelling between nations has accelerated by about 18% per decade. As a result, the transit times between different shelf seas of the Arctic have declined significantly. The total area of sea ice transported between EEZs has increased, with the rate of increase varying regionally. The rapid thinning (approximately 50% over the satellite era) of ice, however, means that the volume of freshwater transport between EEZs in solid form has probably nonetheless declined. As the summer melt expands, melt competes with transport and even fast-moving ice can be “caught” by the melt front before it can “escape” from its EEZ. We apply our tracking method to climate model output to consider this and other trends will evolve as the Arctic transitions to a seasonally ice-free state.

This study considers the influence of both stratospheric and tropospheric aerosol changes on Arctic sea ice extent on interannual to decadal timescales. While it is well-known that accounting for volcanic aerosol increases the skill of interannual to decadal predictions of surface temperature, its impact on sea ice prediction has not been widely considered. Here we analyze the impact of historical volcanoes on Arctic sea ice extent using a large ensemble of 20th-century climate simulations, and demonstrate that an increase in Arctic sea ice extent following these eruptions persists for up to a decade. Even though we do not find a detectable response in observations, these results suggest that inclusion of the effect in seasonal to interannual predictions of sea ice extent should improve skill on these timescales. Secondly, while the influence of future greenhouse gas changes on sea ice extent is known to be important for predictions on decadal timescales, the influence of projected tropospheric aerosol changes has received less attention. We examine the simulated response of Arctic sea ice to projected aerosol and aerosol precursor emissions changes under the Representative Concentration Pathway (RCP) scenarios, and show that projected aerosol emissions decreases drive approximately 20-30% of the projected decrease in annual mean Arctic sea ice extent on decadal timescales.

Phytoplankton are a fundamental component of the Arctic ecosystem and carbon cycle. Through their growth and decay, they form the foundation of the oceanic food web and are a sink for atmospheric CO2. Phytoplankton populations undergo periods of exponential growth, known as ``blooms", which occur seasonally in many of the world's oceans. In 2011, a phytoplankton bloom was observed underneath a region of the Arctic fully covered by sea ice, unexpected as sea ice is typically understood to transmit little solar radiation to the ocean below. To investigate the likelihood and location of sub-ice blooms, we develop a critical-depth model for regions of the ice-covered Arctic ocean that incorporates the transmission of solar radiation through regions of sea ice that are covered by melt ponds. We find that favorable conditions for sub-ice blooms exist over a large portion of the modern Arctic. The development of bloom-permitting regions of the ice-covered Arctic, which comprise greater than 30\% of the latitudes above 70N in July has occurred only in the recent two, coinciding with the thinning of Arctic sea ice and an increase in melt pond coverage. Our results demonstrate that these biological events may be both possible and likely in regions previously considered off-limits to photosynthetic activity. Projections of a thinner Arctic sea ice cover in a warming world suggests that the likelihood and extent of sub-ice phytoplankton blooms will increase in the future.

The Antarctic Circumpolar Current Separates the cold polar oceans (density dominated by salinity) from the warm subtropical waters (density controlled by temperature). In the polar oceans cooling of the surface waters in autumn has little effect on the density of the mixed layer, allowing a rather thin surface layer to freeze to the freezing point before sea ice formation. Moving across the ACC, the cooling has a stronger affect on the density and causes convective deepening, eventually deepening the surface to depths that are so deep there is not enough winter sea-air heat loss to cool such a thick layer to the freezing point. There is a location where sea ice simply cannot grow because of this surface layer thickening—that location forming the virtual sea ice wall (SIW). This presentation discusses the underlying physics and shows the location of the SIW. Its climatological position aligns remarkably well with the climatological ice edge suggesting that the wall may actually be restricting the northern extent of the sea ice edge. Discussing the wall is preceded by a presentation of the Antarctic ocean-sea ice interaction, as well as a discussion comparing and contrasting such processes in the Arctic (mainly contrasting).

Improved sea ice forecasting must be based on improved model representation of coupled system processes that impact the sea ice thermodynamic and dynamic state.  Pertinent coupled system processes remain uncertain and include surface energy fluxes, clouds, precipitation, boundary layer structure, momentum transfer and sea-ice dynamics, interactions between large-scale circulation and local processes, and others.

We will present results, comparisons, process-oriented diagnostics, and parameterization assessment from sea ice forecasts using a version of the Regional Arctic System Model (RASM) adapted for short-term Arctic sea ice forecasting. Specifically, cloud, atmosphere, and ocean observations for studies of atmospheric predictability, air-ocean turbulent fluxes, and sea ice conditions collected in the marginal ice zone from ship-based campaign and coastal Arctic land stations will be used. We will also outline future model improvements based on a comparison of observations that include replacing the mixed-layer ocean model with a multi-layer upper ocean model to allow for observed mixed layer variability, such as storing heat below the surface layer that is transferred up in large storms and horizontal ocean advection.

RASM is a limited-area, fully coupled ice-ocean-atmosphere-land model. It includes the Weather Research and Forecasting (WRF) atmospheric model, the LANL Parallel Ocean Program (POP) and Community Ice Model Version 5 (CICE5) and the NCAR Community Land Model (CLMv4.5) configured for the pan-Arctic region. These components are coupled using a regionalized version of the CESM flux coupler (CPL7), which includes modifications important for resolving the sea ice pack’s inertial response to transient (i.e. weather) events.

In order to optimize the model for short-term forecasts the dynamic level ocean model has been replaced with a mixed–layer ocean model, the horizontal domain is limited to the Arctic, and all components are run with 10km horizontal resolution. This model is run with a bulk double-moment cloud microphysics scheme for droplets and frozen hydrometeors that allows both size and number of hydrometeors to vary in response to environmental conditions (Morrison et al. 2009).

Daily 5-day forecasts with RASM-ESRL were run for the 2015 freeze-up season, initialized with GFS atmosphere and AMSR2 sea ice analyses and forced by 3-hourly GFS forecasts at the lateral boundaries. The forecasts were delivered daily and used for guidance on the UNOLS research vessel Sikuliaq during the SeaState campaign. These daily forecasts have been validated with observations of surface fluxes and vertical profiles of cloud ice and liquid at land sites, and with observations of surface fluxes and sea ice characteristics from recent ship campaign and ice mass balance buoys. These relatively short forecasts are currently being used to validate and improve simulations of synoptic evolution, atmospheric boundary-layer structure and surface energy fluxes over sea ice and the adjacent ocean.

This study quantifies the role of Antarctic sea ice in the transformations of water masses within the Southern Ocean State Estimate (SOSE). Winds drive a strong export of sea ice away from the continent towards the open ocean. The resulting freshwater fluxes at the ocean surface dominate the Southern Ocean freshwater budget (compared with direct precipitation and glacial melt), and these strong fluxes have a major impact on density, stratification, and circulation. Using Walin’s water mass transformation framework, we isolate the contributions of brine rejection, ice melt, and snow interception on the modification of seawater density. Together with direct atmospheric precipitation - evaporation, glacial melt, surface heat flux, and interior mixing, these processes provide the thermodynamic transformations necessary to sustain the meridional overturning circulation. The transformation analysis reveals that sea-ice freshwater fluxes are the main contributor to the transformation of upwelling Upper Circumpolar Deep Water, pushing it primarily towards lighter Antarctic Intermediate and Subantarctic Mode Water but also partly toward denser classes. We also examine the seasonal cycle in transformation, revealing a subtle interplay between brine rejection and upper ocean mixing. Overall these results indicate a tight coupling between Antarctic sea ice and the upper branch of the Southern Ocean overturning circulation.

Seasonal predictions at the Met Office are made using the GloSea5 coupled forecasting system which has contributed September sea ice predictions to SIPN since its implementation in 2013. The ocean and sea ice components of GloSea5 are initialised using analysis fields from the FOAM ocean-sea ice analysis and forecast system. Both GloSea and FOAM are run daily at the Met Office and use the Nucleus for European Modelling of the Ocean (NEMO) model coupled to the Los Alamos sea ice model (CICE).

Data assimilation in the FOAM analysis system is performed using the NEMOVAR 3D-Var First Guess at Appropriate Time (FGAT) scheme. Satellite and in-situ observations of temperature, salinity, sea level anomaly and sea ice concentration are assimilated by FOAM each day. Sea ice assimilation is performed using SSMI/S sea ice concentration data obtained from the EUMETSAT OSI-SAF. At present assimilation of sea ice is unbalanced from the rest of the ocean-sea ice system and is performed in a separate NEMOVAR minimisation. In this talk a brief overview of the challenges associated with assimilating sea ice concentration in a multi-category sea ice model will be given and the various assumptions made will be discussed. Details will be provided of planned sensitivity studies intended to increase our awareness of how these assumptions affect the predictability of sea ice on short-seasonal time scales.

Although the FOAM analyses used to initialise GloSea benefit from assimilation of sea ice concentration data, other aspects of the sea ice - particularly thickness - is unconstrained by observations. Work is under way at the Met Office to implement assimilation of sea ice thickness data derived from satellite altimetry (i.e., CryoSat-2) within the FOAM-GloSea system. A key goal of this involves assessing the potential impact of sea ice thickness initialisation for the GloSea forecasts. Details of planned work designed to achieve these goals will be presented which includes exploring the link between winter thickness biases/errors and the evolution of concentration forecast errors in GloSea. This assessment can include participation/involvement from other modelling groups within SIPN and we would be interested in exploring potential collaboration in this regard.

From March through October 2015, the National Oceanic and Atmospheric Administration (NOAA) Climate Prediction Center provided monthly sea ice outlooks to the National Weather Service Alaska Region. The outlooks were based on output from an experimental set-up of the Climate Forecast System Version 2 (CFSv2) model, which ingests initial sea ice thickness data from the Pan-Arctic Ice Ocean Modeling and Assimilation System (PIOMAS). All other initialization data come from the Climate Forecast System Reanalysis (CFSR). Twenty ensemble model runs were integrated from the 8th through the 12th of each month and several products were delivered including sea ice concentration, sea ice concentration standard deviation, sea ice probability, first projected sea ice melt date, and first projected sea ice freeze date (starting in June). Here we present an overall verification of different products against available observations.

Results show stronger skill in the prediction of sea ice extent and sea ice concentration than its operational counterpart and illustrates the importance of using an improved forecast system as sea ice prediction continues to be a crucial factor for a variety of Arctic interests. Predictions of ice melt date and ice freeze date, while not as widely used, also proved valuable. Exact prediction of these variables is near impossible due to atmospheric processes that cannot be predicted months in advance. However, it is shown that through modeled trends and standard deviations, that it is indeed plausible to issue a meaningful forecast of first ice freeze date and ice melt date given a spread is quantified. Further refinements to the experimental model will continue and plans are to resume monthly sea ice outlooks beginning in March 2016.

In order to meet the urgent need of sub-seasonal to seasonal forecast NCEP started their 9-month seasonal forecasts from 2012 using Climate Forecast System version 2 (CFSv2). In this study, we examined the predictability of Arctic sea ice and atmospheric forcing fields produced by CFSv2. What we found is that the extra large sea ice cover at the end of summer are actually caused by the relative large cold temperature biases in the same region. The predicted atmospheric forcing fields are evaluated against several reanalysis products including the European Center for Medium-Range Weather Forecasts (ECMWF) Interim Re-Analysis (ERA-Interim), NASA the Modern Era Retrospective-analysis for Research and Applications, version 2 (MERRA2), NCE-CFSR, and NCEP/NCAR-Reanaysis (R1). While there are discrepancies among the reanalysis products, we do find that some are better performed than the others on selected fields. The variables we investigated are those regional or stand-alone ocean-ice models will take as their atmospheric forcings. The quality of these variables is therefore has direct impact on the regional or stand-alone ocean-ice model simulation results. Further analyses reveal that biases in turbulent heat fluxes are believed to be the main factor leading to the sea ice forecast biases. The forecast bias in the surface short wave and long wave radiative forcing can mainly be attributed to the cloud cover uncertainty in the Arctic.

Studies have shown that by using improved ice thickness distribution to initialize the Climate Forecast System, version 2 (CFSv2) can result in improved sea ice simulation. Here we examine whether such sea ice thickness initialization can have positive impact as well on surface atmospheric forcing in the Arctic by a number of ensemble seasonal prediction experiments. Sea ice thickness from the Climate Forecast System Reanalysis (CFSR) and the Pan-Arctic Ice Ocean Modelling and Assimilation System (PIOMAS) were used to initialize the hindcasts. The predicted atmospheric forcing is evaluated against several reanalysis products including the European Center for Medium-Range Weather Forecasts (ECMWF) Interim Re-Analysis (ERA-Interim) and the National Centers for Environmental Prediction-National Center for Atmospheric Research Reanalysis (NCEP/NCAR-R1). While there are discrepancies among the reanalysis products, we do find that some are better performed than the others. The variables we investigated are those regional ocean-ice model will be taken as atmospheric forcings. The quality of these variables therefore has direct impact on the regional or stand-alone ocean-ice model simulation results. Preliminary results show that seasonal predictions of surface atmospheric forcing are sensitive to the initial sea ice thickness. This suggests that local meteorological forecasts can be improved by assimilating sea ice thickness into the coupled climate model.

Sea ice thickness (SIT) has a potential to contain substantial climate memory and predictability in the northern hemisphere (NH) sea ice system. We use 5-member NH SIT, reconstructed with an ocean-sea-ice general circulation model (NEMOv3.3 with LIM2) with a simple data assimilation routine, to determine NH SIT modes of variability disentangled from the long-term climate change. Specifically, we apply the K-means cluster analysis - one of nonhierarchical clustering methods that partition data into modes or clusters based on their distances in the physical – to determine optimal number of NH SIT clusters (K=3) and their historical variability. To examine prediction skill of NH SIT clusters in EC-Earth2.3, a state-of-the-art coupled climate forecast system, we use 5-member ocean and sea ice initial conditions (IC) from the same ocean-sea-ice historical reconstruction and atmospheric IC from ERA-Interim reanalysis. We focus on May 1st and Nov 1st start dates from 1979 to 2010. Common skill metrics of probability forecast, such as rank probability skill score (RPSS) and ROC (relative operating characteristics - hit rate versus false alarm rate) show that our dynamical model predominately perform better than the 1st order Marko chain forecast (that beats climatological forecast) over the first forecast year. On average May 1st start dates initially have lower skill than Nov 1st start dates, but their skill is degraded at slower rate than skill of forecast started on Nov 1st. Besides this study the full spectrum of Arctic research activities at the Earth Sciences Department at the Barcelona Supercomputing Center will be also presented.

We examine how the choice of a particular satellite-retrieved sea-ice concentration dataset used for initializing seasonal climate forecasts impacts the prediction skill of Arctic sea-ice area and Northern hemispheric 2-meter air temperatures.

To do so, we performed two assimilation runs with the Max Planck Institute Earth System Model (MPI-ESM) from 1979 to 2012, where atmospheric and oceanic parameters as well as sea-ice concentration were assimilated using Newtonian relaxation. The two assimilation runs differ only in the sea-ice concentration dataset used for assimilating sea ice. In the first run, sea-ice concentrations as derived by the NASA-Team algorithm are used, while in the second run sea-ice concentrations computed from the Bootstrap algorithm are assimilated. A major difference between these two sea-ice concentration data products involves the treatment of melt ponds. While for both products melt ponds appear as open water in the raw satellite data, the Bootstrap algorithm more strongly attempts to offset this systematic bias by synthetically increasing the retrieved ice concentration during summer months.

For each year of the two assimilation runs we performed a 10-member ensemble of hindcast experiments starting on 1 May and 1 November with a hindcast length of 6 months. For hindcasts started in November, initial differences in Arctic sea-ice area and surface temperature decrease rapidly throughout the freezing period. For hindcasts started in May, initial sea-ice area differences increase over time. By the end of the melting period, this causes significant differences in 2-meter air temperature of regionally more than 3°C. Hindcast skill for surface temperatures over Europe and North America is higher with Bootstrap initialization during summer and with NASA Team initialization during winter. This implies that the choice of the sea-ice data product and, thus, the observational uncertainty also affects forecasts of teleconnections that depend on Northern hemispheric climate indices.

Sea ice forecasts for the Arctic are investigated. Sea ice forecasts are generated for 6 hours to 9 days using the Marginal Ice Zone Modelling and Assimilation System (MIZMAS) driven with 6 hourly forecasts of atmospheric forcing variables from the NOAA Climate Forecast System (CFSv2). Forecast sea ice drift speed is compared to observations from drifting buoys and other observation platforms. Forecast buoy positions are compared with observed positions at 24 hours to 9 days from the initial forecast. The initial thickness fields are compared to aircraft remote sensing data from Operation IceBridge (OIB). Forecast Ice concentrations and ice edge positions are compared to passive microwave products. Forecast skill is assessed relative to reference forecasts based on climatology or persistence. RMS errors for ice speed are found in the order of 5 km/day for 24 h to 48 h using the sea ice model vs. 12 km/day using climatology. Following adjustments in the sea ice model to remove systematic biases in direction and speed, predicted buoy position RMS errors are 6.5 km for 24 hour forecasts and 15 km after 72 hours. Using the forecast model increases the probability of tracking a target drifting in sea ice with a 10x10 km sized image to 95% vs. 50% using climatology. Initial ice thickness computed by MIZMAS compares well with OIB data. Skill assessment of ice edge position forecasts faces challenges due to observational uncertainty and shows room for improvement. The results are generated in the context of planning and scheduling the acquisition of high resolution images which need to follow buoys or research platforms for scientific research but additional applications such as navigation in the Arctic waters may benefit from this accuracy assessment. Ideas for future improvement of short term sea ice forecasts are explored. Implications for intra-seasonal forecasts are considered.

In recent years, the demand for improved sea ice forecasts in the Arctic has intensified as maritime transport and offshore exploration increase. To provide these forecasts, a few centers around the world have implemented operational sea ice forecasting systems. Here, I will provide an overview of the state of operational short-term (1-10 days) sea ice forecasting. These systems are hindered by a number of challenges associated with issues related to observation uncertainties (e.g., surface melt in passive microwave retrievals), data assimilation methods and model initialization. Additionally, the absence of a few key physical processes not represented in models (e.g. wave-ice interactions) is thought to contribute to large forecast error growth in certain regions. Some recent model developments (e.g., landfast ice parameterization, form drag) should help to improve sea ice forecasts. An additional challenge to sea ice forecasting is in how to bridge the gap with operational needs and translate forecasts into operational products (ice edge, ice internal pressure, etc.). Finally, the importance of verification methods to assess the forecast quality and to provide guidance in the development of sea ice forecasting systems will also be discussed.

The Met Office seasonal forecast system (GloSea) submitted both a June outlook, and an August update to the 2015 SIPN outlook. Here we will assess the GloSea performance through last summer's sea ice melt season - our first using the latest coupled configuration of the Met Office Unified Model - Global Coupled configuration 2 (GC2; Williams et al., 2015).

One particular aspect that has been of concern to us has been the performance of the forecast system with respect to the hindcast system, which is used to both estimate the skill of the system, and to provide a means to calibrate the system for bias. The hindcast sea ice analysis, forced by the ERA-I atmospheric analysis has somewhat thicker ice than the forecast sea ice analysis which is forced by the Met Office NWP atmospheric analysis. We will show relationships between these thickness differences in the analysis, and sea ice concentration differences in forecast.

Finally, we should be able to offer you a first glimpse of our 2016 outlook for September sea ice.

Seasonal forecast quality assessment relies on two sources of information: a forecast (produced by a model) and a reference for verification. It is frequent to evaluate various forecasts against a common reference, to detect the effects of e.g. model improvements on prediction skill. The emergence of multiple observational datasets allows to ask the dual question: how does the skill of an individual forecast depend on the quality of the underlying reference dataset? We address this question for the seasonal prediction of summer Arctic sea ice extent over 1993-2008, using six different observational products. We find that skill scores, measured using sample correlation, can vary substantially (up to 0.1) from product to product, and for the same forecast. The reasons behind these variations, sometimes as large as variations noticed when upgrading the model itself, are discussed.

Here we assess seasonal analyses and forecasts of sea ice and Arctic hydrographic conditions produced by the successful NASA Global Modeling and Assimilation Office (GMAO) sea ice forecasting system, and in part driven by the Modern Era Retrospective analysis for Research and Applications (MERRA2). The GMAO GEOS-5 ocean and sea ice models are built on a 0.5ox0.5o resolution ocean model with GFDL MOM V4 numerics and CICE V5 sea ice. The ocean hydrography is updated from a historical data base using an ensemble OI scheme with a 10-day data window. Assimilated data includes SST, hydrography. An updating scheme for sea ice relies on gridded passive microwave estimates of concentration. Independent comparison data sets include ice thickness, and drift, as well as the WHOI ice-tethered profiler data. In the first part of this talk we examine the historical Arctic Ocean hydrographic and sea ice properties, focusing on the past decade. Our goal is to understand the connection between sea ice changes, changing weather conditions, and the contribution of changing near-surface heat flux. In the second part of the talk we present results from an examination of the seasonal forecasts.

For several years the Naval Research Laboratory has participated in the Sea Ice Outlook (SIO) to predict the Arctic September minimum sea ice extent. In earlier years these forecasts used the regional Arctic Cap Nowcast/Forecast System, but have more recently used the Global Ocean Forecast System (GOFS) 3.1, which is comprised of the two-way coupled Community Ice CodE (CICE) and the HYbrid Coordinate Ocean Model (HYCOM). The horizontal resolution is ~3.5 km near the North Pole and the system runs daily in pre-operational mode at the Naval Oceanographic Office assimilating satellite ice concentration data. Using ice and ocean initial conditions from this data-assimilative system at the appropriate start date, the ice/ocean models are integrated in forecast mode through October 1st. Atmospheric forcing is from the National Centers for Environmental Prediction Climate Forecast System Reanalysis (CFSR) for the years 2005-2014. This ensemble of 10 members gives an indication of how sea ice can respond to variable atmospheric conditions during summer and the projected sea ice extent for September is the average across all ensemble members. Each year’s forecast ice concentration bias is pre-determined by comparing a non-assimilative GOFS forecast with a data-assimilative GOFS hindcast. Monthly ice concentration error is translated into a monthly heat flux offset that is added to the net surface heat flux during the final SIO forecasts. This methodology has proven effective in reducing GOFS forecast ice concentration biases.

In this contribution we extend the pan-Arctic analysis presented in Williams et al. (2015) to produce a regional forecast of sea-ice conditions for each of the Arctic peripheral seas including the Beaufort, Chukchi, East Siberian, Laptev and Kara seas. The model builds on the finding that the mean large scale sea ice circulation the previous winter determines where the following summer minimum sea ice extent. We show that, in particular, the mid to late winter coastal sea-ice divergence plays a major role. When ice divergence occurs late in the winter, new ice forms but it does not have the time to grow to sufficient thickness to survive the following summer melt. To produce a regional forecast, we use a Lagrangian sea-ice model to backtrack an imaginary line defining the boundary of a given sea, starting from the beginning of the melt season in June and ending in November of the previous year. Results show that the position of the ice edge in each of the peripheral seas of the Arctic is well correlated with the previous winter’s divergence. The maximum correlation is obtained when the synthetic ice edge is backtracked from May to February. Note that, unlike the previous pan-Arctic study of Williams and colleagues in which the net ice divergence could be correlated with the winter mean Arctic Oscillation index, this regional analysis requires sea ice drifts in order to calculate the winter mean ice divergence along the coast. This finding supports the need for continuous production of real-time satellite-derived sea-ice velocity vectors, which can now be used for observation-based regional forecasting.

In recent years, dynamical seasonal forecast systems have started to incorporate interactive sea ice components, allowing for sea ice forecasts on seasonal (1-12 month) timescales. Previous investigations on the forecast skill in such dynamical forecast systems have mainly focused on area-integrated sea ice quantities, which are of little use to end-users. Here we assess the skill in predictions of ice-free and freeze-up dates - variables that are highly relevant to shipping - in a set of hindcasts performed with the Canadian Seasonal to Interannual Prediction System (CanSIPS).

We find statistically significant skill for both ice-free and freeze-up dates in large parts of the Arctic, including Hudson's Bay, Baffin Bay/Labrador Sea and the Barents, Kara and Chukchi Seas. Skill is particularly large for freeze-up dates with significant skill at lead times up to 6-12 months. While most of the skill of ice-free dates stems from persistence of sea ice concentration anomalies, the longer-lead time skill for freeze-up dates is due to a reemergence mechanism involving ocean temperatures. These results highlight the potential for dynamical forecast systems to provide valuable forecasts of socio-economically relevant sea ice variables on seasonal timescales.

Due to its persistence on seasonal timescales, Arctic sea ice thickness (SIT) is a potential source of predictability for summer sea ice extent (SIE). New satellite observations of SIT represent an opportunity to harness this potential predictability via improved thickness initialization in seasonal forecast systems. The winter and spring SIT field may also encode information about the future spatial patterns of summer SIE anomalies. In this work, the evolution of Arctic SIT anomalies is studied using a 1400-year control integration and initialized ensemble forecasts from a fully-coupled global climate model. Our analysis is focused on the September sea ice zone, as this is the region where thickness anomalies have the potential to impact summer SIE. It is found that, in addition to a general decay with time, sea ice volume anomalies display a summer enhancement, in which anomalies tend to grow between the months of May and August. This summer enhancement is relatively symmetric for positive and negative volume anomalies and occurs between May and August regardless of the initial month. Analysis of the surface energy budget reveals that the summer volume enhancement is driven by a positive feedback between the SIT state and the surface albedo. The SIT state affects surface albedo through changes in the melt onset date, ice thickness distribution, and sea ice concentration field, yielding an anomaly in the total absorbed shortwave radiation between May and August, which enhances the existing SIT anomaly. This phenomenon highlights the crucial importance of accurate SIT initialization in seasonal forecast systems.

By now, drawing inference from predictions is considered as the most powerful statistical approach to generalization. While that is already known for over 30 years (Leo Breiman 2001), most practitioners and funders still ignore this basic scheme, and get hold up instead with lower performing mechanistic, parsimonious, linear and/or similar (‘homebrew’) models that do not live up to relevant scrutiny and performance metrics. The consistently under-predicting summer sea ice models for the Arctic make that very clear for years, as do the Population Viability Analysis (PVAs) by IUCN, medical disease models, permafrost models and many of the IPCC forecasts. This leaves us with disastrous impacts.

This presentation presents over 50 projects that have already used Machine Learning, namely ensemble algorithms like bagging and boosting, across a diverse set of applications worldwide to obtain best possible ‘classifiers’ (predictions) explicit in time and space. I will give a synthesis and an overview from these reviewed and published (polar) studies. Further, I will explain some of the roadblocks of the current scientific quagmire. Finally, I will present a good way forward on how to make predictions better and how to create a best-possible science-based institutional model culture for pre-cautionary decision-making and for reaching ‘true’ global sustainability while many problems (ecological, sociological, economical) currently appear to go ‘through the roof’.

This study examines the predictability of Arctic sea ice concentration at the intra-seasonal time scale through developing a stochastic model using weekly time series of sea ice concentration (SIC), oceanic and atmospheric variables from 1979 to 2012. The first order selection of predictors was achieved through anomaly correlations (AC) between SIC and oceanic and atmospheric variables. Sea surface temperature (SST), surface air temperature (SAT), 300mb geopotential height (GH300) and downward long-wave radiation (DLR) stands out than other variables. A linear Markov model was then developed in a multi-variate space of SIC, SST, SAT, GH300 and DLR. Model variables were further selected through prediction skill measured by AC and root-mean-squared error (RMSE) between predictions and observations. The SIC and SST combination is sufficient to capture predictable variability in such a model. Due to different ocean-sea ice coupled relationships in the Pacific and Atlantic sectors of the Arctic, the final model was built in two regions' multi-variate spaces for capturing regional co-variability. Weekly predictions of SIC were made at grid points up to 12-week lead-time, initialized by each week of the time series. The model skill was evaluated by the cross-validated model experiments. Because SIC has relatively high persistence, the model usually cannot beat persistence in first two-week predictions in most study areas but posses predictable skill for longer lead time predictions. The model can out-performs persistence for the predictions with lead-time longer than 2-week in the Chukchi Sea and East Siberian Sea but only beats persistence with lead time longer than 6-week in the Baffin Bay and Bering Sea. The model has predictive skill (AC>0.6) with RMSE less than 20% for predictions up to 6-week lead-time in the most areas in the Arctic Basin during the warm season (May to September).

We present results of data-driven predictive analyses of sea ice over the main Arctic regions. Our approach relies on the Multilayer Stochastic Modeling (MSM) framework of Kondrashov, Chekroun and Ghil [Physica D, 2015] and it leads to probabilistic prognostic models of sea ice concentration (SIC) anomalies on seasonal time scales. This approach is applied to monthly time series of state-of-the-art data-adaptive decompositions of SIC and selected climate variables over the Arctic. We evaluate the predictive skill of MSM models by performing retrospective forecasts with “no-look ahead” for up to 6-months ahead. It will be shown in particular that the memory effects included intrinsically in the formulation of our non-Markovian MSM models allow for improvements of the prediction skill of large-amplitude SIC anomalies in certain Arctic regions on the one hand, and of September Sea Ice Extent, on the other. Further improvements allowed by the MSM framework will adopt a nonlinear formulation and explore next-generation data-adaptive decompositions.

Motivated by Arctic intra-annual variability phenomena such as SST \& sea ice re-emergence, we use a prediction approach for sea ice anomalies based on analog forecasting. Traditional analog forecasting relies on identifying a single analog in a historical record, usually by minimizing Euclidean distance, and forming a forecast from the analog's historical trajectory. We use an ensemble of analogs for our forecasts, where the ensemble weights are determined by a dynamics-adapted kernel, which take into account nonlinear geometry on the underlying data manifold. We apply this method for forecasting pan Arctic and regional sea ice concentration anomalies from CCSM4 model data, and in many cases find improvement over the persistence forecast, notably a 2 month increase in predicting September sea ice extent.

A Vector Auto-Regressive (VAR) model is evaluated for predicting the 1979-2012 summer time (May through September) daily Arctic sea ice concentration on the intra-seasonal time scale, first using only the daily sea ice data and without direct information of the atmosphere and ocean. The cross-validated forecast skill of the VAR model is found to be superior to both the anomaly persistence and damped anomaly persistence at lead times of 20~60 days, especially over Northern Eurasian marginal seas and the Beaufort Sea. The daily forecast of ice concentration also leads to predictions of ice-free dates and September mean sea ice extent. This study reveals for the first time that Arctic sea ice can be predicted statistically with reasonable skills at the intra-seasonal time scales given the small signal-to-noise ratio of daily data.

Each Arctic summer since 2008, the Sea Ice Outlook (SIO) has invited researchers and members of the public to contribute their predictions regarding the September mean extent of Arctic sea ice. The SIO collects and publishes these contributions online in three cycles having deadlines at the start of June, July and August each year. Post-season reports summarize how predictions compared with the observed September extent, aiming to provide feedback and insights for improvement. The unique public character of the SIO, with its focus on predicting a single number whose true value soon becomes known, brings an element of constructive gamification to the science process as well. Here we analyze the performance of more than 400 individual predictions from the SIO’s first eight years, testing for differences in ensemble skill across different years, different months, and five general types of method: heuristic, statistical, mixed, and ice-ocean or ice-ocean-atmosphere modeling. In general, prediction accuracy reveals a strong pattern of easy and difficult years. Difficult years, in which most predictions are far from the observed September extent, tend to be those with large positive or negative excursions from the overall downward trend. In contrast to these large interannual effects, ensemble improvement from June to August is comparatively small. Among method types, predictions based on statistics and ocean-ice-atmosphere modeling perform better than heuristic methods.

In the Sea Ice Outlook prediction system, we find that dynamical models not only show poor skill in forecasting observed September sea ice extent but are also equally unsuccessful at predicting each other, indicating a large divergence in model physics and/or initial conditions. Motivated by this, we have performed a modeling experiment with SIO models that include both fully coupled global models, and global/regional ice-ocean models. We run two sets of simulations — a control set initialized with a climatological May 1 sea ice thickness and an experiment set initialized with May 1 2015 sea ice thickness. This allows us to investigate both model uncertainty and error growth, and the importance of post-processing to forecast uncertainty. We find that without post-processing, model uncertainty is the dominant source of forecast uncertainty, while with post-processing, overall forecast uncertainty is greatly reduced and error growth dominates forecast uncertainty. This result suggests that uniform bias correction across SIO models’ forecasts will greatly reduce the forecast spread. We also investigate the spatial patterns of forecast uncertainty and find that forecast uncertainty grows at a faster pace along coastal areas compared to the central Arctic basin. Potential ways of offering forecast information based on probabilistic metrics are also explored.