2. DATA AND METHODS

2.3. Positive degree-day (PDD) sum

RACMO2 estimates of melting are based on the balance of energy fluxes at the snow/ice surface. Because we do not have these data for the AWS, we use PDD as a proxy for melting calculated from both modelled and observed temperatures. PDDs and melting are typically well correlated (e.g. Braithwaite and others, Reference Braithwaite1995). For example, we find a correlation of r = 0.72 between RACMO2 estimates of total annual PDDs and melt amount averaged over the now-missing portion of the Larsen-B ice shelf. We use a PDD sum that accounts for both the duration and intensity of above zero temperature conditions (Eqn (1), Vaughan, Reference Vaughan2006).

(1) $$\varphi _{\rm n} = \mathop \sum \limits_{i = 1{\rm st\; September,\; year}\; n - 1}^{i = 31{\rm st\; August,\; year}\; n} T_i^{} \left( {t_{i + 1} - t_i} \right)\alpha \left( {T_i} \right),$$

where φ is annual total PDDs, T is observed/modelled 2 m temperature in °C, t is the time expressed in fractions of days and α(T _i ) is 1 if T is positive and 0 if T is negative. Prior to calculating PDDs we interpolate/aggregate (as appropriate) all seven sets of observations and both model versions onto a commonly sampled hourly time series (we do not gap-fill missing data), thus T in Eqn (1) is hourly temperature.

2.4. Estimating radiation errors

Unaspirated temperature sensors are prone to radiation errors under high-insolation, low wind speed conditions, which result in a temperature excess to the actual air temperature (Erell and others, Reference Erell, Leal and Maldonado2005, Huwald and others, Reference Huwald2009). Smeets (Reference Smeets2006) showed that for a Vaisala HMP45C platinum resistance thermometer this excess temperature can be estimated as a function of total shortwave radiation at-sensor and wind speed (Eqns (2) and (3)).

(2) $$\eqalign{\Delta T &= \; \displaystyle{k \over {(9.6\; \times \; {10}^{ - 3}k + 6.3){(12U)}^{1.25}}}\; \;{\rm for}\;\; \cr & \quad U \ge 2\,{\rm ms}^{ - 1}}.$$

(3) $$\Delta T = 4.14 \times 10^{ - 3}k - 0.15\; \;{\rm for}\;\; U \lt 2\,{\rm m}{\rm s}^{ - 1},$$

where ΔT is the excess temperature, U is wind speed and k is the total shortwave radiation at sensor. In the absence of observed radiation, we estimate k to be:

(4) $$(1 + \alpha )\tau SW_{{\rm TOA}},$$

where SW _TOA is the top of the atmosphere incoming shortwave radiation, τ is the atmospheric transmissivity (0.79 for clear sky conditions in Antarctica, van den Broeke and others, Reference van den Broeke, Reijmer and van de Wal2004) and α is the surface albedo. We estimate the impact of this effect on our AWS derived PDD totals at Larsen C AWS (Fig. 2) which uses an instrument and radiation shield comparable with that described in Smeets (Reference Smeets2006). We take α to be 0.75 since this AWS is situated on snow; this can be considered a conservative estimate (Kuipers Munneke and others, Reference Kuipers Munneke, van den Broeke, King, Gray and Reijmer2012). We are unable to estimate radiation errors at Matienzo and the UNAVCO stations using this method; while Matienzo also uses a HMP45C temperature sensor it has a much larger old-style radiation shield and the UNAVCO AWS use a different type of thermometer entirely.

Fig. 2. RACMO2.3/5.5 simulated temperatures for Matienzo and Larsen C AWS. (a, c) daily mean 2 m temperature with (blue) and without (red) the Larsen B ice shelf. (b, d) daily mean 2 m simulation with Larsen B vs simulation without Larsen B.

3. RESULTS AND DISCUSSION

We compare RACMO2.3/5.5 and RACMO2.3/27 simulated values for annual mean 2 m temperature and surface atmospheric pressure with observations made at the AWS (Fig. 1). Both versions of the model reproduce spatial and interannual variability in annual mean surface pressure almost exactly (r = 0.99 in both cases), which suggests that the model captures synoptic scale atmospheric conditions across the region well. Generally, both versions of the model capture interannual variability in mean annual temperature (r = 0.90 and 0.87, respectively). A slight deviation from the one-to-one line is apparent in the slope of the linear fit to these data in both model versions and suggests that RACMO is less well able to capture temperatures that are either particularly high or particularly low. Specifically, RACMO overestimates mean annual temperature in particularly cold years, and underestimates mean annual temperature in particularly warm years; this is also the case at the sub-annual timescale (Fig. S2) and in the mean, summertime only, temperatures (Fig. S3). This is potentially due to differences in the surface type; RACMO assumes that all surfaces are ice/snow covered when in fact four of the seven AWS are on bare rock, including Matienzo and Robertson Island. Due to its low albedo, relative to ice/snow, exposed rock is likely to heat up considerably in summer, potentially allowing for higher 2 m air temperatures than over snow surfaces. A paucity of observations precludes a robust assessment of precipitation, but previous evaluations show that RACMO2.3/5.5 can reproduce interannual variability in accumulation at James Ross Island (the nearest available accumulation record to Larsen B) during our study period (r = 0.75 for 1998–2007, from figure in van Wessem and others, Reference van Wessem2016).

While the simulation with higher spatial resolution (RACMO2.3/5.5) offers a slight improvement in model-observation agreement overall, the change in resolution introduces a low bias in mean annual temperature at Matienzo (~−1.8°C) and Robertson Island (~−1.6°C), which are both located on the Seal Nunataks at the interface between the Larsen B and Larsen A embayments. This is somewhat consistent with a previous study, which has shown that RACMO2.3/5.5 simulates temperatures which are too low on average (van Wessem and others Reference van Wessem2015). However, it is possible that this bias is partly due to the presence in the RACMO domain of the Larsen A (~four grid points) and B (>20 grid points) ice shelves, the former of which collapsed in 1995. We investigate the degree to which the presence/absence of Larsen B affects simulated temperatures at Larsen C and Matienzo by comparing the data with an additional RACMO2.3/5.5 simulation for the period 2003–13 in which we removed the Larsen B ice shelf from the model ice-sheet mask (Fig. 2). At Matienzo, which is closest to the Larsen B ice shelf, there is a bias of ~3°C between the two simulations; the removal of the ice shelf results in higher simulated temperatures here, presumably due to heat exchange from the (relatively) warm ocean. This is particularly evident during the winter; the bias diminishes by an order of magnitude as temperatures increase. In fact, we calculate a bias of just 0.3°C if we only include above-zero temperatures in the calculation. This can be attributed to the fact that ocean-atmosphere temperature gradients are much greater during the winter. The interannual variability in mean annual temperature is consistent between the simulations with and without the ice shelf (r = 0.99, n = 9) and the effect is very localized; we see no difference in simulated temperatures with/without the ice shelf at Larsen C.

Interannual variability in total PDDs (Fig. 3) is particularly well captured by RACMO at Matienzo (r = 0.97, r = 0.91 for RACMO2.3/5.5 and RACMO2.3/27 respectively) and at Robertson Island (r = 0.98, r = 0.98), Foyn Point (r = 0.98, 0.97) and Cape Framnes (r = 0.95, r = 0.98), but less so at Larsen C AWS (r = 0.72, r = 0.74). In particular, both versions of RACMO2 ‘miss’ the particularly high PDD year in 2001/02 observed at the Larsen C AWS. The high melting observed on Larsen C in this year has been attributed to anomalous atmospheric circulation conditions (van den Broeke, Reference van den Broeke2005); predominant northeasterly winds and prolonged transport of air over open water towards the Larsen ice shelf. Presumably, this was not fully captured by the model. Both versions of RACMO2 underestimate total annual PDDs by more than half at all sites considered here except for Cape Framnes (Fig. 3). We investigate the degree to which this can be attributed to radiation errors under low wind, high insolation conditions at the AWS sensors. We subtract the calculated excess temperature (order of 1°C, see Section 2) from the observations when (1) it is daylight, (2) wind speed is <4 ms⁻¹ and (3) wind direction is from the North West (so likely to be cloud-free, Kuipers Munneke and others, Reference Kuipers Munneke, van den Broeke, King, Gray and Reijmer2012). We find that this reduces the number of total annual PDDs by up to 30% at Larsen C (8% on average), that the effect is particularly notable in high melt years and that comparing with the corrected time series improved model performance (r = 0.82, r = 0.79). RACMO2.3/5.5 again offers an improvement against RACMO2.3/27 in predicting both PDD amount and interannual variability at Matienzo, however we note that in this case, increasing the horizontal resolution does not have a major impact on model-observation agreement. When we correct for the model bias at Matienzo (by adding 1.82°C to all modelled data) we see excellent agreement with the observations (Fig. S4). PDDs in very high melt years are underestimated slightly even with this correction; this is likely due to radiation errors at sensor. Note that while annual PDD totals do not always reflect true annual values due to sampling of observation data (e.g. Fig. S1), with the exception of 84 days between March and May in 2004, the Matienzo AWS record is complete during the period reported.

Fig. 3. Annual PDDs modelled and observed at each AWS. Note: model has been sampled to available observations (Fig. S1). Observed values are given in black, observed (corrected) values are given in grey. Simulated values are given in blue (RACMO2.3/27) and red (RACMO2.3/5.5). Panels (a, c, e–i) show cumulative PDDs from the 1^st September each year. Panels (b, d, j–n) show modelled vs observed total annual PDDs, Pearson's correlation co-efficient (r) and mean bias (b) between the observed and modelled annual values are annotated. Vertical orange line in (a) and (c) indicates timing of ice shelf collapse.

The timing of the Larsen B ice shelf collapse has been attributed to the draining of melt ponds on the ice shelf surface (Scambos and others, Reference Scambos, Hulbe, Fahnestock, Domack, Leventer, Burnett, Bindshadler, Convey and Kirby2003) coinciding with anomalously strong melt conditions observed at the Larsen C AWS (van den Broeke, Reference van den Broeke2005). However, the AWS data presented here show that while the number of PDDs at Matienzo in the year of collapse was above average (145 in 2001/02, 7% greater than the 1999–2010 mean value), the total PDD amount 2 years prior to collapse (1999/2000) was 30% higher. Previous work based on these data has reported that, of the 2 years, 2001/02 had more days where the temperature rose above zero (Skvarca and others, Reference Skvarca, De Angelis and Zakrajsek2004). Our analysis agrees with this (Table 1), but we find that there were significantly more hours with above zero temperatures in 1999/2000 than in 2001/02 and that the average above-zero temperature was almost half a degree higher. This results in more PDDs in 1999/2000 than 2001/02, when calculated using our method, which takes both duration and magnitude of above-zero temperature events into account. It has also been noted previously that 2001/02 had a warmer summer in the Larsen B region than 1999/2000 (Cape and others, Reference Cape2015). We also see this in our analysis (Fig. 4) but note that the autumn of 1999/2000 was particularly warm. This accounts for the fact that we calculate a continued accumulation of PDDs after the end of the summer in 1999/2000 but not in 2001/02. These signals appear in both the Matienzo AWS data and the RACMO simulation, which can be considered independent.

Fig. 4. Temperature during 1999/2000 (black) and 2001/02 (pink) observed at the Matienzo AWS and simulated by the RACMO2.3/5.5 model. Bottom: full temperature range, top: above zero temperatures only. Solid lines show data smoothed with a 7-day window.

Table 1. Summary of the above zero conditions as observed at the Matienzo AWS and simulated by RACMO2.3/5.5 in 1999/2000 and 2001/02

Of the 2 years, although 1999/2000 had more PDDs than 2001/02 at Matienzo, we agree with previous findings that the latter was, in fact, the higher melt year of the two at Larsen C AWS (60 PDDs in 2001/202, 100% greater than the 1996–2012 average with 2004–06 and 2012 excluded due to missing data, Fig. S1). The Larsen C AWS is ~220 km further south than the now-missing portion of the Larsen B ice shelf. However, it is unlikely that this distance is large enough to explain differences in the PDD signal between the two sites. The anomalous atmospheric circulation conditions promoting high melting on Larsen C (van den Broeke, Reference van den Broeke2005) would likely have also affected Larsen B. We attribute this high degree of spatial variability to differences in the relationship between melting and föhn wind events. Larsen C AWS is further away from the Antarctic Peninsula Mountains than Matienzo and so is unlikely to be as strongly affected by föhn warming. Total melt in the northern part of the Larsen B embayment (particularly over the Seal Nunataks) is more highly correlated with föhn activity than on Larsen C (r > 0.5 and r < 0.5, respectively during 1999–2010, Cape and others, Reference Cape2015). In fact, total annual PDDs and number of föhn days observed at Matienzo (ibid) are highly correlated (r = 0.72). Our finding that the RACMO2.3/5.5 model reproduces temporal variability in total annual PDDs at Matienzo with a high degree of fidelity (especially when temperatures are bias corrected) suggests that RACMO2.3/5.5 is able to represent local scale processes controlling melt such as the frequency of föhn winds. However, further work is needed to confirm that this is generally the case, and not just confined to this location.

In the longer term context (1980–2014, Fig. 5, Table 2); we see from RACMO2.3/5.5 that the 1990s were a climatologically anomalous decade. Although mean annual temperature generally declined between 1980 and 2014, summer (DJF) temperatures in the 1990s (−2.35°C on average, σ = 0.88) were higher than in the 1980s (−2.68°C, σ = 0.51) and in the 2000s (−2.61°C, σ = 0.77). Similarly, melting was stronger and precipitation lower in the 1990s than in either the decade before or after. We use multivariate changepoint analysis (Killick and others, Reference Killick, Fearnhead and Eckley2012, see supplementary methods) to assess the degree to which this difference is statistically sound and detect changepoints at 1992, 1996 and 1999, in general showing that melting and snowfall in the 1990s was statistically different to that before and since (Fig. S5). This is consistent with the findings of Turner and others (Reference Turner2016) who show that air temperature increased and then decreased during this period with the transition from ‘warming’ to ‘cooling’ occurring ~1998–99. Turner and others (Reference Turner2016) attribute the pre-1999 warming to a positive trend in the Southern Annular Mode increasing the flow of mild, northwesterly air, which is likely to have been amplified through the föhn effect in the Larsen B region. Melting can contribute to ice shelf break-up if supraglacial lakes are able to form and drain, however a condition of supraglacial lake formation is the depletion of firn pore space (KuipersMunneke and others, Reference Kuipers Munneke, Ligtenberg, van den Broeke and Vaughan2014), which is achieved when snowfall is low and/or melting is high. A high degree of melting in the decade before the collapse of the Larsen B ice shelf in 2002 is therefore consistent with previous findings that multiple strong melt years are required to deplete firn pore space and weaken an ice shelf before it can fail (Trusel and others, Reference Trusel2015). The apparent ‘return’ to 1980s levels of melting after break up is also consistent with the wider pattern of decadal-scale atmospheric cooling observed in the region since then (Turner and others, Reference Turner2016, Oliva and others, Reference Oliva2017) and may have provided for a period of stability for the remnants of Larsen B, including Scar Inlet. The RACMO2 simulation also shows that the year before collapse – 2000/01 – was characterised by the lowest amount of melting (0.004 Gt, one third of the average value) and the highest amount of snowfall (0.0145 Gt, 60% higher than average) of the 35 year period. Further investigation is required to determine whether these conditions played a role in the collapse event, but potential avenues for enquiry include the impact of higher than usual snow loading on the weakened ice shelf, and the impact of refreezing/latent heat release within the thicker than usual snowpack.

Fig. 5. RACMO2.3/5.5 estimates of annual PDDs (based on 3-hourly 2 m air temperature, red), melting (purple) and snowfall (pink), averaged over the now-missing portion of the Larsen B ice shelf. Number of föhn days at Matienzo (blue, Cape and others, Reference Cape2015). Solid gray vertical lines refer to the Larsen A and Larsen B ice shelf collapse events. Dashed grey horizontal lines represent decadal mean values.

Table 2. Decadal average 2 m temperature, PDDs, melt and precipitation over the now-missing portion of the Larsen B ice shelf, as simulated by RACMO2.3/5.5

5. ACCESS TO DATA AND ACKNOWLEDGEMENTS

Larsen C AWS data are available via request from Steve Colwell ([email protected]), Matienzo AWS data are available from Pedro Skvarca ([email protected]) and Sebastian Marinsek ([email protected]) and UNAVCO AWS data can be accessed here: https://www.unavco.org/data/tropospheric/surface-met/surface-met.html. RACMO2 model output is available on request from J. M. van Wessem. The R package used to perform multivariate changepoint analysis (changepoint.mv) is soon to be released on (the comprehensive R network) CRAN, and is available from Rebecca Killick ([email protected]). MRvdB, SRML and JMvW acknowledge funding from the Polar Program of the Netherlands Organization for Scientific Research (NWO/NPP) and the Netherlands Earth System Science Centre (NESSC).