Glaciers around the world are losing mass at an accelerating pace with profound implications for global sea-level rise, global climate, local river systems, and their physical, biogeochemical, and ecological properties, as well as natural hazards. Observations indicate a consistent and largely homogenous trend of global retreat and shrinkage of the Earth's mountain glaciers, ice-shelf collapses in Antarctica, and an exceptional acceleration in ice flow and associated mass losses of outlet glaciers in Greenland and West Antarctica in response to atmospheric and ocean warming. These changes highlight the need for further capabilities to monitor, understand and predict the behavior of land ice and its interactions with the atmosphere and the oceans.

In the last decade, new remote sensing technologies and missions, such as NASA's Ice, Cloud, and land Elevation Satellite (IceSAT), the Gravity Recovery and Climate Experiment (GRACE), and the Operation IceBridge have revolutionized our ability to monitor the state and changes of land ice masses, in particular of the two huge ice sheets in Greenland and Antarctica, and glaciers in remote polar regions. These observations indicate that mass loss from glaciers and ice sheets has contributed approximately 60% of total global sea-level rise during the period 2003–2009 with roughly equal shares from the two ice sheets in Greenland and Antarctica combined (259 ± 28 Gt year⁻¹ or 0.71 ± 0.08 mm SLE (sea-level equivalent) year⁻¹) and all ~200,000 glaciers outside the ice sheets (289 ± 49 Gt year⁻¹ or 0.79 ± 0.13 mm SLE year⁻¹; Shepherd et al., 2012; Gardner et al., 2013), highlighting their socio-environmental relevance. Extending these estimates to cover the period April 2002–November 2014, Reager et al. (2016) find that the ice sheet contribution has increased considerably (1.26 mm SLE year⁻¹), while the glacier contribution remained similar (0.65 mm SLE year⁻¹). Hence, the share of the ice sheets relative to the mountain glacier contribution has risen. Their results also indicate that the rapidly increasing cryospheric contribution has partially been masked by an increase in water stored on land.

The discovery that the ice sheets can change much faster than previously considered possible, took the glaciological community by surprise, rendering the ice sheets a “wild card” in any attempts to project future contributions to sea-level rise (Truffer and Fahnestock, 2007; Alley and Joughin, 2012). Some marine-terminating outlet glaciers in Greenland have doubled or even tripled their speed (Joughin et al., 2004) in just a few years, accompanied by thinning and rapid retreat after years of stable front positions (Howat et al., 2007; Stearns and Hamilton, 2007). Flow acceleration has gradually expanded to the north (Rignot and Kanagaratnam, 2006). Some glaciers have then returned to “normal” flow speeds (Murray et al., 2010) indicating that rapid changes in either direction are possible within just a few years. The largely synchronous acceleration in ice flow over large areas is indicative of a common climatic (Joughin et al., 2008) or oceanic driver (Holland et al., 2008; Straneo et al., 2010), although the response to regional and local forcing is complex (Moon et al., 2012, 2014, 2015; Straneo et al., 2013).

Many Greenland outlet glaciers occupy deep troughs extending far inland below sea level setting the stage for inherently unstable glacier behavior. The situation is even more worrisome in West Antarctica, where most of the ice sheet is grounded below sea-level on an inland-sloping bed, providing a delicate setting prone for potential run-away effects once retreat is initiated (Joughin and Alley, 2011), as had been envisioned early on (Mercer, 1978). Rapid changes possibly indicative of such an ice sheet instability have recently been observed at Thwaites and Pine Island Glacier (Mouginot et al., 2014), which combined drain ~5% of the Antarctic Ice Sheet. Ice-sheet simulations by De Conto and Pollard (2016) suggest that rapid mass loss from Antarctica due to marine ice-sheet instabilities may cause a meter of sea-level rise by 2100 and more than 15 m by 2500. Studies investigating the observed dynamic thinning, grounding line retreat and net mass loss suggest that these changes may signal the onset of a collapse of the West Antarctic ice sheet (Favier et al., 2014; Joughin et al., 2014; Rignot et al., 2014; Ritz et al., 2015). Once initiated, a yet not fully understood chain of feedback mechanisms may cascade: reduced buttressing of the upstream ice due to retreat causes ice speed-up, which leads to dynamic thinning and increased calving at the marine-terminating fronts, which in turn leads to further mass loss and retreat. Surface mass-balance processes do not trigger rapid changes directly, but enhanced surface melt has been identified as a precursor to catastrophic ice-shelf collapses (Banwell et al., 2013).

Overall, accelerated ice discharge into the oceans has tripled the mass loss of both ice sheets to 298 ± 58 Gt year⁻¹ (0.83 ± 0.16 mm year⁻¹) during 2000–2011, compared to the period 1992–2000 (Shepherd et al., 2012). Some estimates of multi-year averages of mass loss after 2009 exceed 350 Gt year⁻¹ for Greenland (McMillan et al., 2014) and 150 Gt year⁻¹ for Antarctica (Helm et al., 2014; Khan et al., 2015). Velicogna et al. (2014) found that the significant acceleration of ice sheet mass loss is driven mainly by a few subregions. Sutterley et al. (2014) found that the acceleration of mass loss in the Amundsen Sea area of West Antarctica is nearly 3 times larger during the period 2003–2009 than it was over 1992–2013.

The Fourth Assessment Report (AR4) of the Intergovernmental Panel on Climate Change (IPCC, 2007) pointed out that ice flow may be a key driver of rapid ice sheet change, but recognized that “quantitative projections cannot be made with confidence” mainly due to a lack of process understanding regarding rapid and possibly catastrophic and irreversible changes of ice sheets or parts thereof. The lack of adequate representation of these processes in the previous generation ice sheet models was considered the largest source of uncertainty in projecting twenty first century sea-level rise. In the wake of AR4 a wealth of new observations has been collected, and significant progress has been made in process understanding and the ability of models to account for couplings between ice sheets, ice streams and ice shelves, and their two-way interactions with the atmosphere, their subglacial environment and the ocean. However, models did not mature sufficiently for the Fifth Assessment Report (AR5; IPCC, 2013) to account explicitly for mass losses from rapid dynamic changes. Large uncertainties persist with regard to modeling and projecting ice sheet evolution and possible catastrophic contributions to sea-level change (Vaughan and Arthern, 2007).

Insufficient understanding of the mechanisms of the ice sheets' rapid dynamic changes remains the main culprit in the quest for better projections. Such changes cannot be triggered by the slow flow of ice by internal deformation, which is relatively well understood and represented in models. Hence rapid change can only be caused by interactions of the ice with its boundaries. Better understanding and proper implementation of the relevant processes, in particular at the ice-ocean and ice-bed interface, into predictive models are among the most critical challenges in reducing uncertainties in projections of the ice sheet contribution to sea-level rise.

Processes at the surface-atmosphere boundary (snow accumulation, surface melt) determine the surface mass balance, which constitutes another important boundary condition for glacier and ice sheet behavior. The physical processes are relatively well-understood (Hock, 2005); however, quantifying recent and future surface mass changes still presents a major challenge. For example, ice-sheet wide annual surface mass balances in Greenland derived from regional climate models deviate by up to a factor of two (e.g., Fettweis et al., 2008; Ettema et al., 2009), in particular due to difficulties in modeling snow precipitation (see Section Snow Accumulation). This problem is only exacerbated in the topographically complex mountain ranges that host non-polar glaciers.

Refreezing is a major component of the mass budget, in particular on ice sheets and high-latitude glaciers. Modeling indicates that ~40% of the total surface melt in Greenland may refreeze (Van Angelen et al., 2013). With atmospheric warming, refreezing may initially increase due to increased melt, but eventually it may decrease due to subsurface warming and decreasing pore space. Storage of water in firn may temporarily buffer the contribution of meltwater to sea-level rise (Harper et al., 2012). Modeling needs to account for the subsurface spatio- and temporal variability in temperature, water content and density; however, despite recent observational advances (Forster et al., 2014; Machguth et al., 2016), these processes are often oversimplified, and therefore poorly modeled in ice-sheet or climate models.

To account for feedbacks with the climate system through changes in surface elevation, albedo and freshwater input into the ocean, it is important to fully couple ice-sheet and (possibly also glacier) models to Atmospheric-Ocean General Circulation Models. However, currently most ice-sheet models are stand-alone models without two-way coupling (Vizcaino, 2014; Vizcaino et al., 2015). The necessary large computational resources are an additional barrier to progress in fully coupled modeling. Similarly, downscaling typically coarse-scale climate data to the glacier scale remains a major challenge in modeling and projecting mountain glacier behavior (Ziemen et al., 2016).

Lack of observational data is still a major issue in assessing and predicting change. Glaciological datasets are often short in duration or cover limited areas, although tremendous progress has been made recently. A wide suite of large-scale data sets of the Earth's glaciers at unprecedented spatial and temporal resolution have been collected in the last decade, including surface velocities (Rignot et al., 2011; Moon et al., 2012; Fahnestock et al., 2016), surface elevations (Howat et al., 2014), surface elevation changes (Pritchard et al., 2009), ice sheet bedrock (Bamber et al., 2013), and ice thicknesses (Fretwell et al., 2013). The first complete inventory of all glaciers outside the ice sheets, the so-called Randolph Glacier Inventory (RGI; Pfeffer et al., 2014) combining data from GLIMS (Global Land Ice Measurement from Space) (Raup et al., 2007; Kargel et al., 2014) with new data to achieve complete coverage globally, has removed a major obstacle in modeling these glacier systems. However, lack of accurate bed topography data, and thus initial ice volume, introduces considerable uncertainties in their modeling.

High-resolution data are needed in particular at the ice-ocean interface. In Greenland the fjord arms are largely unmapped, and in Antarctica, the topography under the ice shelves, which strongly affects sub-ice shelf melt rates (Hattermann et al., 2014), is poorly known. Also, too little is known about ocean properties and their short-term variability. Data collection is not trivial given that this domain is often choked with icebergs calling for the development of innovative data collection techniques, such as attaching instruments to deep-diving seals (Sutherland et al., 2013). Data revealing the past, for example, from ice or sediment cores, are essential to put current changes in context, and to validate models. Collecting data on basal drag, thermal conditions and subglacial waterflow at the subglacial boundary will remain difficult, if not impossible. But recent advances in inverting at least some of these properties from surface data are promising (McNabb et al., 2012; Morlighem et al., 2014; Brinkerhoff et al., 2016; Farinotti et al., 2017).

Research on snow microstructure has exploded in recent years, because it is increasingly recognized that important mechanical, thermodynamical, and electromagnetic properties of snow cannot be predicted without a proper consideration of snow microstructure (Aoki et al., 2011; Lowe et al., 2013). Furthermore new observational technologies such as micro-computer-tomography (Schleef and Lowe, 2013), optical spectrometry (DUFISS) (Gallet et al., 2009) or snow micro-penetrometry (Lowe and van Herwijnen, 2012), have made it possible to quantify snow microstructural effects. What is missing is the development of a unifying theory, describing the physical properties and behavior of snow at a larger scale based on laws of snow microstructure development. While it is quite clear that such a theory will include—next to density—the specific surface area of the snow grains and curvature terms characterizing the snow—pore interface, it remains to be shown that a predictive microstructure model based on density, surface area and mean and/or Gaussian curvature is able to replace current more empirical microstructure models (Lehning et al., 2002). This step is currently being addressed in the snow microstructure community (e.g., Krol and Löwe, 2016) and will lead to the development of new snow models, which are able to describe mechanical and electromagnetic properties of snow in a more physical way.

Snow accumulation in mountainous regions, as well as on ice sheets and sea ice continues to be insufficiently understood (see also Section Surface Processes). The mountain snowpack is a major source of freshwater for many arid and semi-arid regions worldwide, but researchers still struggle to reliably estimate the amount of snow stored in the high alpine regions and how much water can be expected downstream. A major source of uncertainty comes from the fact that measuring (snow) precipitation in the high alpine zone is extremely difficult, if not impossible, given the windy conditions and limited infrastructure available in these areas (Sevruk and Mieglitz, 2002; Sevruk et al., 2009). Recent progress in remote sensing of snow depth distribution is helping to put more constraints on mountain snow cover and precipitation. The analysis of spatial snow depth maps has revealed reduced snow amounts at high elevations (Grunewald et al., 2014; Kirchner et al., 2014), but scientists are only starting to investigate the links to precipitation patterns (Mott et al., 2014). It is therefore critically important to intensify efforts to better quantify precipitation and related available snow and water resources in the high alpine zone.

The problems described above are aggravated in polar regions, particularly in Antarctica, due to continuous high winds and low temperatures. Smaller errors in snow accumulation lead to large errors in the total ice sheet mass budget due to the enormous area of the ice sheet (Zwally and Giovinetto, 2011). Errors in accumulation estimation stem from very sparse measurements of precipitation in both polar regions together with an unknown accuracy of precipitation predictions by weather models. Wind-blown snow adds uncertainty; it results either in redistribution or loss of snow (e.g., open water leads in sea ice). Drifting and blowing snow also results in an unknown amount of sublimation, reducing the total mass before final deposition.

Larger-scale models of ice sheets, glaciers, and sea ice typically treat snow processes at the surface of these ice masses in a rudimentary way, and at an inadequate resolution for calculating, for example, a flux divergence of blowing and drifting snow. While the most advanced models such as MAR (Gallee et al., 2013; Amory et al., 2015) or RACMO (Lenaerts et al., 2012) implement parameterizations of the complex snow processes, quantitative validation is challenging. The community agrees that potential calculation errors are large, and critically influence the modeled total mass balance. One of the most critical processes is drifting and blowing snow, but it is largely unknown how much of the snowfall is permanently deposited and ultimately incorporated in the ice mass (Zwaaftink et al., 2013). Only recently have regional climate models, applied to the ice sheets, started to include a more physical description of firn processes (Van Angelen et al., 2013) and snow redistribution by wind (Lenaerts et al., 2012).

Snow melt and energy balance have been investigated over decades and are generally well understood. However, snow melt models often rely on erroneous meteorological input data (Schlögl et al., 2016), either because the measurements are spatially not representative or because input is taken from meteorological models, which insufficiently resolve boundary layer feedbacks. These feedback effects have been studied over glaciers with respect to katabatic wind formation and associated temperature effects (Greuell and Böhm, 1998; Shea and Moore, 2010). Recent results show how local flow features can change the turbulent heat fluxes (Mott et al., 2015), and that terrain sheltering has significant small-scale effects on melt efficiency (Mott et al., 2016). It is generally accepted that regional melt and therefore also catchment runoff (Cullather et al., 2016) can be well described with calibrated models, often using empirical temperature index methods for snow melt. At the same time, physics-based energy balance models, which are preferable for their transferability to other environments and future climates, have shown a high sensitivity to meteorological input (Schlögl et al., 2016). This remains a challenge to be addressed in the future.

The discussion above already suggests some of the difficulties in dealing with the thin and highly dynamic yet important snow component in larger-scale models. Misrepresenting the massive influence that snow exerts on the atmosphere and the soils will strongly bias model results (Beer, 2016). Snow representation in larger-scale models generally does not correctly reproduce the effect of snow on the atmosphere or ground correctly (Van Angelen et al., 2012). To limit the detrimental effect of an insufficient snow representation, some meteorological models (e.g., COSMO, Baldauf et al., 2011) assimilate snow-covered area, which can be estimated from a combination of ground and satellite observations. This produces acceptable results for short-term forecasting, but it remains a challenge to have an adequate representation of snow dynamics in larger-scale models. Assimilation also does not solve the problem of insufficiently representing the effects of snow on other model components.

Global warming naturally has a large impact on snow dynamics impacting the duration of snow cover and timing of runoff production. Anticipated changes include a shift from snow to liquid precipitation and a potential increase in total precipitation (Lenaerts et al., 2013). More snow accumulation may offset increased ablation of ice sheets as long as it remains sufficiently cold, as has been shown to be important locally in certain regions (e.g., Schmucki et al., 2015). However, future snow changes are currently not modeled well on a global scale. Due to the difficulty to correctly represent snow dynamics in climate models, as discussed above, it is challenging to make reliable predictions on future snow and its feedback effects on the atmosphere and soils. Locally, water resources and tourist industries are affected by future snow changes. Globally, sea level rise and positive feedback mechanisms via release of carbon or methane following permafrost thaw are of major concern.

There has been over 50% loss of sea-ice volume from the Arctic since 1979 (Schweiger et al., 2011; updated to 2015), which is primarily manifested as a loss of perennial ice area (Maslanik et al., 2011) through enhanced transpolar ice drift events (Nghiem et al., 2007), export to the Greenland-Iceland-Norwegian Sea (Smedsrud et al., 2017), and increased melting within the Arctic (Kwok and Cunningham, 2010). The Antarctic has experienced a slight increase in sea-ice extent, but with large zonal variability, with increasing ice extent in the Ross Sea and decreasing extent in the Amundsen-Bellinghausen Sea (Turner et al., 2015). Less is known regarding variability in Antarctic ice thickness and volume.

The decrease in Arctic sea ice is happening at a time of increasing economic interest in mineral exploration, transport and tourism across the region (AMAP, 2004). There is increasing need for improved ice forecasting and accurate projection for strategic planning, as expected changes may increase hazards to human activity in the sea-ice pack (Stewart et al., 2007). The time scale needed to assist governmental and industry planning range from 10 to 50 years ahead. This presents one of the largest challenges to the sea-ice observers and modelers: non-linear dynamics in the climate system impede model predictions of the sea-ice change over timescales longer than 2–10 years (Blanchard-Wrigglesworth et al., 2011; Guemas et al., 2014).

Until recently, one of the most significant challenges of the sea-ice community has been global observation of ice thickness. With satellite monitoring of ice thickness now available (Laxon et al., 2003), the challenge has moved to understanding biases in the measurement method. This has led to a need to better characterize the snow cover on sea ice, as a lack of information about snow thickness is the largest cause of uncertainty in sea-ice thickness measurements (Kwok and Cunningham, 2008). The lack of information about ice thickness is no longer the leading cause for uncertainty in appropriate model parameterizations. Uncertainty in projection of sea-ice change over the coming century is mainly attributed to a number of physical processes, outlined in Section Observation and Model Needs, that are inadequately constrained by current observational data.

The observed recent changes in Arctic sea-ice pack are better represented in the climate models of IPCC's AR5 compared to AR4 due to improvements to sea-ice thermodynamic parameterizations in AR5 (Stroeve et al., 2012). However, these improvements have not decreased the spread in model realizations. Part of the uncertainty lies in natural variability in the Earth system, since predictions of sea ice are most uncertain on decadal timescales (Blanchard-Wrigglesworth et al., 2011).

Attribution of recent sea-ice loss in the Arctic has been called a “Grand Challenge” of climate science (Kattsov et al., 2010). To meet this challenge, we need to be able to accurately simulate the seasonality in the feedbacks controlling sea-ice mass balance. To do so we need better observational data constraining these feedbacks, which are related to coupled processes between the upper ocean, sea ice and atmosphere. In particular we do not know how heterogeneity in the sea-ice cover impacts the radiative, sensible and latent heat fluxes from the surfaces. Uncertainty in cloud properties results in significant uncertainty in downwelling longwave radiation flux at the ice-ocean surface.

Sea-ice has the largest seasonal cycle in mass of all cryosphere components. This seasonality is controlled by the balance of positive and negative feedbacks to ice thickness that change over the year as the surface energy balance goes from net positive in summer to negative in winter. Interannual variability in sea-ice thickness reflects variability in atmosphere and ocean, and indicates high sensitivity of the sea-ice cover to the surface energy balance. The average change in heat budget to explain the loss of ice over the last three decades is only about 1 Wm⁻² (Kwok and Untersteiner, 2011). Attribution of this additional heat is a matter of current debate.

Positive feedbacks on sea-ice loss are numerous, and are counteracted by the large negative feedback caused by ice thickening in winter. Representing the variability from these competing processes is important for correctly modeling climate sensitivity (Notz, 2015). Carmack et al. (2015) document the terms that have most uncertainty in the sea-ice surface energy budget. The atmospheric flux components introduce an order of 10 Wm⁻² uncertainty, which is more than enough to explain recent ice loss. Inability to constrain these fluxes in models contributes to the wide variability in sea-ice seasonality, interannual variability and trends in climate models. The uncertainty in ocean heat flux to the sea-ice pack is another, smaller, yet significant contributor.

A review of recent advances in parameterizing sea ice-atmosphere processes is presented by Vihma et al. (2014). Carmack et al. (2015) include a review of sea ice-ocean processes. Climate model sensitivity is related to the representation of sea-ice processes, and models present wide variability in their representation of the sensitivity of sea ice to greenhouse gas forcing (Stroeve and Notz, 2015). The processes that contribute to this uncertainty are outlined below.

In order to simulate observed variability in the sea-ice pack volume, the ice and ocean surface energy balance needs to be accurate. This requires accurate simulation of all the competing feedbacks in the system, which range from strongly negative to strongly positive. Parameterizing these processes will require improved observations. In particular there is need for: improved knowledge of the composition of Arctic clouds throughout the year; identification of melt pond fraction and the processes driving melt pond evolution on ice of varying topography and snow cover; improved representation of sea-ice deformation in models; and improved understanding of how heterogeneity in sea ice impacts atmospheric and oceanic stability and related fluxes across the ice surface.

At the heart of improving the representation of the sensitivity of sea ice to the surface energy balance is improving our ability to realistically simulate interannual variability in sea-ice mass. Climate models extending into the next century project ice loss events that occur on similar timescales as the reduction in Arctic ice volume observed in the last decade (Holland et al., 2006). Diagnosis of feedbacks controlling such events shows a variety of mechanisms that enhance ice loss (Holland et al., 2006; Vavrus et al., 2011). Climate models on average demonstrate lower sea-ice sensitivity to climate forcing than observations suggest. So while the IPCC model ensembles represent a large spread in future projection of sea-ice state, one can question whether this spread of realizations is realistic given that positive and negative feedbacks to ice mass in the system are likely to be larger than simulated.

Identifying the magnitudes of the leading order feedbacks in the climate system is crucial to improve climate model capabilities in projecting climate changes. The nature of sea-ice loss events, the rate of ice change and interannual variability is controlled by feedbacks that are poorly constrained in climate models. Rapid ice loss events are highly likely, and we are currently experiencing one. The time scale and mechanisms for these events are poorly resolved in current climate models.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.