The meteorology system and thermosalinograph of the research vessel (RV) Nathaniel B. Palmer recorded the variables listed in Table  over 57 d. We use these observations to compute the bulk THF; the computation method is described in Sec. . The ship departed Punta Arenas (Chile) on 6 January 2022, reaching the Amundsen Sea (72° S, 117° W) on 15 January 2022. It then entered the polynya region and spent 31 d within 20 km of the Dotson or Getz ice shelves and finally left the polynya region on 25 February 2022 (Fig. b). To determine the THF and be consistent with ERA5's temporal resolution, we compute hourly means of the variables in Table . The initial resolution was 1 min. The hourly position of the ship is used to create a classification: Southern Ocean, open ocean in the polynya region (ship more than 20 km away from the coastline), in front of the Dotson or Getz ice shelves (ship within 20 km), and in a sea-ice-covered region (where the sea-ice concentration (SIC) is larger than 0.15). The RV Nathaniel B. Palmer presumably avoided regions of higher sea-ice concentration on its transit to the Amundsen Sea. Airflow distortion of the wind speed values caused by the superstructure of the RV Nathaniel B. Palmer was negligible (Appendix , Fig. ); we therefore did not perform any correction. Several conductivity, temperature, and depth (CTD) casts were taken during the research campaign. In the present study we use one, taken at 74.02° S, 113° W in front of Dotson Ice Shelf (Fig. 1a, blue square) to initialize the 1D PWP (Price–Weller–Pinkel) model (see the model description in Sect. 2.2.2).

We use Conservative Temperature, Absolute Salinity, and pressure from an ocean profiling Seaglider that was deployed during the ship campaign to compute the daily HC in the upper 40 m of the water column. The glider was deployed on 17 January 2022 in front of Dotson Ice Shelf (73.8° S, 112.6° W). It sampled to the seabed, with a maximum of 901 m, surfacing between each dive. The glider then headed south towards the ice shelf and returned north along the Dotson–Getz trough before being recovered on 4 February 2022 (red transect, Fig. a). A total of 286 profiles were collected. The data are gridded horizontally per profile and vertically with a resolution of 2 m.

In this study, we assess ERA5 by comparing its hourly mean THF with the THF computed from the observations. We also use some of ERA5's hourly mean meteorological and sea surface variables to recalculate the THF (Table ). ERA5 is a global reanalysis product that provides “maps without gap” of atmospheric and sea surface variables . It has a hourly temporal resolution and a regular 0.25° latitude–longitude grid. To co-locate its variables to the ship data, we use the nearest-neighbor grid cell and the corresponding hour (as the ship data have been hourly averaged).

We use sea-ice concentration to determine when the research vessel was surrounded by sea ice (15 % threshold) for the location classification (Fig. b). We select the satellite-based sea-ice product ARTIST sea ice (ASI) created by Bremen University data downloaded from https://data.seaice.uni-bremen.de/amsr2. The ASI algorithm takes input data from the Advanced Microwave Scanning Radiometer 2 (AMSR2, Level 1B) – a sensor operating on the JAXA satellite GCOM-W1 – and outputs gridded data (Level 3, grid space is 3.125 or 6.25 km). The temporal resolution is daily; the selected output grid space for this study is 3.125 km. The ASI algorithm has been validated against observations and shows good performance . It should be noted that during the research cruise, the sea ice gradually melted, so from February 2022 onwards, we can no longer really speak of a polynya, as only a tongue of ice attached to Thwaites Ice Shelf remains visible in the satellite-derived product (https://data.seaice.uni-bremen.de/databrowser/, last access: 13 January 2025). We therefore refer to the polynya region.

The THF is the sum of the latent heat flux (LHF) and the sensible heat flux (SHF). The LHF is related to the air–sea heat exchange originating from the sea surface evaporation, whereas the SHF arises from the air–sea temperature gradient. We use the Coupled Ocean–Atmosphere Response Experiment (COARE) 3.5 algorithm through AirSeaFluxCode to compute the THF from the observations. The COARE 3.5 algorithm relies on bulk parameterizations: the LHF and SHF are computed as a function of air density (ρair), wind speed measured at height zu (Uzu), the transfer coefficient corresponding to the measured height zm of humidity and temperature, and the measured height zu of the wind speed (Cq(zm,zu) or Ct(zm,zu)). For the SHF (Eq. ), the specific heat capacity (Cp) and air–sea temperature gradient (Tair,zm-Tskin) are calculated; for the LHF (Eq. ), the latent heat of vaporization (Lv) and air–sea humidity gradient (qair,zm-qsat) are calculated. 1SHF=ρairCpCt(zm,zu)Uzu(Tair,zm-Tskin)2LHF=ρairLvCq(zm,zu)Uzu(qair,zm-qsat)

The AirSeaFluxCode applies a logarithmic correction to the transfer coefficient definitions (Ct(zm,zu) and Cq(zm,zu)) to account for the height zu of wind speed and zm of air temperature and humidity measurements . Atmospheric stability is accounted for in the definition of the transfer coefficients Ct and Cq through stability functions. The measured relative humidity is converted to saturated humidity using the saturation vapor pressure function from . Warm-layer and cool-skin corrections are applied to convert the measured SST (Table ) to skin SST, as is required in Eq. (). These corrections follow . The COARE 3.5 algorithm requires wind speed relative to the ocean surface. Here, we assume that the ocean currents are low in comparison to the wind speed and neglect them. We find this reasonable, as have shown from recording current meters installed on a 2-year mooring that the coastal surface current in the Amundsen Sea is about 0.2 cm s⁻¹, which is 0.03 % of the mean wind speed (7.9 m s⁻¹; Fig. a) in our dataset. The flux convention is downward, which means that a negative (positive) flux corresponds to a heat loss (gain) for the ocean surface.

To account for the insulating effect of sea ice, we scale the turbulent fluxes by the sea-ice concentration (SIC) when SIC ≥ 15 % (Eqs. , , and ). We acknowledge that this is a simplified method to account for the sea-ice effect on turbulent fluxes but accept this, considering the small amount of time spent by the RV in a sea-ice-covered area (3 d out of 57 d) and the low importance of the flux variability in sea ice for the results of this study.

3LHF=(1-A)LHF,4SHF=(1-A)SHF, where 5A=SICifSIC≥0.150ifSIC<0.15. We perform a Reynolds decomposition to analyze the flux variability. We decompose SHF and LHF into the sum of their average (denoted by an overline) and their anomaly (denoted by ^′): SHF = SHF‾ + SHF^′ and LHF = LHF‾ + LHF^′. We follow the same method as in , , and except that we do not neglect the contributions of Cq′ and Ct′, as they are more than 5 % of their mean values not shown; criteria following. We replace the variables Uzu, ΔT=Tair,zm-Tskin, Δq=qair,zm-qsat, Cq, and Ct in Eqs. () and () with the sum of their mean and anomaly (details in Appendix ). Finally, we obtain 6SHF′=SHF-SHF‾=ρairCp[ΔT′U‾Ct‾︸t term+ΔT‾U′Ct‾︸u term+ΔT‾U‾Ct′︸ct term+Ct‾(ΔT′U′-ΔT′U′‾)︸t–u term+U‾(ΔT′Ct′-ΔT′Ct′‾)︸t–ct term+ΔT‾(U′Ct′-U′Ct′‾)︸u–ct term+ΔT′U′Ct′-ΔT′U′Ct′‾︸cov term]7LHF′=LHF-LHF‾=ρairLv[Δq′U‾Cq‾︸q term+Δq‾U′Cq‾︸u term+Δq‾U‾Cq′︸cq term+Cq‾(Δq′U′-Δq′U′‾)︸q–u term+U‾(Δq′Cq′-Δq′Cq′‾)︸q–cq term+Δq‾(U′Cq′-U′Cq′‾)︸u–cq term+Δq′U′Cq′-Δq′U′Cq′‾︸cov term].

LHF^′ and SHF^′ are the flux variability around the mean. We establish indices of contribution for each term in Eqs. () and (), following . To quantify the contribution of a term X (X has to be substituted by one of the terms defined in Eqs. () and ()) to the flux anomaly Y (Y= L for the LHF^′ and Y= S for the SHF^′), we compute the absolute value of the X term and divide it by the sum of the absolute values of the seven terms (Eq. ). 8CY(X)=|X term|∑|allterms|

Therefore, the contribution indices CY(X) have values between 0 and 1, and their sum equals 1. The closer to 1 CY(X) is, the larger the contribution of the term X is to the flux anomaly SHF^′ or LHF^′.

The 1D mixed-layer model PWP Price–Weller–Pinkel; is used to investigate the relative impact of the different THF estimates, produced using observations and ERA5, on the SST and HC. The model needs two input files: one contains the initial ocean state (temperature and salinity profiles), and the other contains a time series of the atmospheric forcing (radiative flux, turbulent flux, freshwater flux, and momentum flux). The initial ocean profile (Fig. ) comes from a CTD cast in front of Dotson Ice Shelf (74.0° S, 113° W; Fig. a, blue square). We carry out four simulations with forcings from the observations and ERA5 that differ only in the THF (Fig. e, g) in order to isolate its effect. The input freshwater flux only contains precipitation and evaporation; freshwater input from melting sea ice was not considered in this study. We consider this reasonable, as 2022 was a record-low sea-ice year , and most of the sea-ice melt had already occurred in the polynya region (https://data.seaice.uni-bremen.de/databrowser/, last access: 31 January 2025). The four simulations are further detailed in Sect. . We remove the Southern Ocean data to focus on the polynya region. The simulations are accomplished with the aim of (i) evaluating if the ERA5 flux co-location/computation method is important, (ii) verifying if any bias was induced by a moving ship, and (iii) retrieving daily changes in SST and HC due to THF and comparing them to the observations from the glider and the ship. We compute the ocean HC (Eq. ) across the upper 40 m of the water column because all four simulations converge below 40 m (Fig. g). ρ0 is the mean potential density in the first 40 m, computed from Absolute Salinity and Conservative Temperature; cp is the mean specific heat capacity in the first 40 m, computed from Absolute Salinity, in situ temperature, and sea pressure; and CT is the Conservative Temperature. 9HC=ρ0cp∫z=0z=40CTdz

First, we analyze the THF computed from the ship observations (Fig. ) to understand the heat flux magnitude and variability in the Amundsen Sea (Fig. ).

The Amundsen Sea (comprising the polynya region and along the ice shelves in the classification) lost on average more turbulent heat (-52 W m⁻², SD = 51 W m⁻²) than the Southern Ocean region (-30 W m⁻², SD = 26 W m⁻²; Fig. e). The largest instantaneous heat loss events were also observed in the Amundsen Sea (the maximum is -230 W m⁻² versus -145 W m⁻² in the Southern Ocean). In particular, more than 97 % of the turbulent heat loss events larger than 150 W m⁻² occurred when the ship was within 20 km of the Dotson or Getz ice shelves (purple classification, Fig. ).

Within the polynya region, the THF variations (Fig. e) were linked to short-scale SHF loss between -90 and -140 W m⁻² (Fig. c), while in the Southern Ocean, the largest turbulent heat loss events (Fig. e) were driven by the LHF loss (Fig. a).

Throughout the time series, the LHF contributed most to the net THF loss, accounting for an average of 57 % of the THF. This is evidenced by the mode of both the LHF and THF being between -30 and -10 W m⁻², whilst for the SHF, the mode is between -10 and 10 W m⁻² (Fig. b, d, f). Thus, while the LHF contributed the most to the total THF, the most significant short-term (days to weeks) THF loss events were imputed to large SHF losses.

Below, we investigate the key drivers of the variability in the THF. We decompose the flux anomalies into different terms (Eqs. , ). These terms indicate the contributions to SHF^′ and LHF^′ of the temperature gradient (t term), humidity gradient (q term), wind speed (u term), transfer coefficients (ct term and cq term), and the cross-contribution of the following variables: second-order terms (t–u term, q–u term, u–cq term, etc.) and a third-order term or residual (covariance term) (Fig. ).

The t term (associated with the anomalous air–sea temperature gradient) was larger than the sum of all the other terms (Fig. a, blue color bars dominate). The indices of contribution of the terms (Table ) were calculated for each hourly data point and range between 0 and 1 to show the relative importance of one term compared to another. The t term was the most frequent dominating factor: 37 % of CS(t) values were above 0.5 (Table ). This indicates that in 37 % of the data, air–sea temperature gradients were responsible for more than 50 % of SHF^′. Following the same arguments, the q term (the anomalous air–sea humidity gradient) contributed the most to LHF^′ (Fig. b, blue color bars) but to a slightly lesser extent: above 33 % of LHF^′ values had CL(q) as the strict dominant factor.

The decomposition indicates that the variability in the air–sea property gradient was the dominant factor impacting the variations in the THF. Further investigations show that the atmospheric variables (air temperature and humidity) control the air–sea property gradients. Indeed, the variability in the air temperature is higher (standard deviation (SD) = 2.62 °C) than the variability in the SST (SD = 1.13 °C; Fig. b, c). The same statement holds for the humidity: the air humidity has a higher variability (SD = 0.70 g kg⁻¹) than the saturated humidity (SD = 0.32 g kg⁻¹; Fig. d, e). The difference in SD is even larger when we remove the Southern Ocean data (not shown). These results indicate the importance of cold, dry air masses driving large heat loss events.

To further analyze the heat loss mechanism in the Amundsen Sea, we investigate the relationship between the turbulent heat loss and the broader-scale synoptic variability (Fig. ).

The large turbulent heat losses were associated with winds blowing from the south (Fig. a, b, c), with large temperature (Fig. a) and humidity (Fig. b) gradients. We focus on one heat loss event that lasted 6 h on 19 February 2022 from 08:00 to 13:00 GMT (stars in Fig. ). The THF remained below -170 W m⁻² (Fig. , red area), reaching its peak of -211 W m⁻² at 11:00 GMT. The mean value over the 6 h was -186 W m⁻². The wind was directed from the continent (Fig. d wind vectors and Fig. f) and brought cold (on average -9.4 °C; Fig. b) and dry (1.4 g kg⁻¹; Fig. d) air on top of the warmer (on average -0.1 °C; Fig. c) and moister (3.7 g kg⁻¹; Fig. e) sea, triggering the heat loss event. A low-pressure center was also visible on the map (Fig. d). It may have enhanced the heat loss event. This weather system (cold and dry continental winds) has been observed for all the major turbulent loss events occurring during this research cruise (not shown). On the contrary, the instances when the THF was significantly positive (> 30 W m⁻²) are consistent with warm and moist northerlies blowing over the Amundsen Sea (Figs. e and f).

This indicates the role of large-scale atmospheric variability on the local flux events. Next, we review the state-of-the-art reanalysis to understand the ability of numerical weather models to represent these key processes.

The research vessel spent 72 % of its time in the polynya region and along the Dotson and Getz ice shelves, where few validations of ERA5 have been conducted. First, we analyze the THF from ERA5 by comparing it to the observed fluxes, which were calculated using the COARE 3.5 algorithm (Fig. ).

The agreement between the THF product from ERA5 and the THF calculated from in situ data via COARE 3.5 was low: r2 = 0.186 for the SHF (Fig. a) and r2 = 0.291 for the LHF (Fig. c). Additionally, ERA5 was positively biased in comparison to the observations: the mean SHF and LHF were higher by 13 and 3 W m⁻² (Table , first and second rows). Similarly, hourly episodic heat loss events were not well represented by ERA5 (Fig. b, d): the difference between the two time series was up to 141 W m⁻² for the SHF and 81 W m⁻² for the LHF (Table , second row).

The low agreement between the two flux products is explained by the coarse resolution of ERA5 at land–sea boundaries. The research vessel was often stationed along the ice shelves where the closest ERA5 grid cell was considered ice shelf and as such does not have an SST value (Fig. ). The correlation between ERA5 and the COARE 3.5 fluxes improved when only comparing instances where the nearest ERA5 grid cell had an SST value not set to NaN (not a number) (r2 = 0.618 for the SHF and r2 = 0.691 for the LHF; Fig. a and c, blue points). For instances where there is no SST, the correlation is weak (r2 = 0.122 for the SHF and r2 = 0.094 for the LHF; Fig. a and c, yellow points). Thus, the ERA5 reanalysis THF product underestimates the turbulent losses at the land–sea boundary formed by the ice shelves due to missing SST values. These results illustrate the importance of careful analysis when investigating ice-shelf processes in reanalyses.

As shown above, the dominant mechanism impacting the THF was the variations in air temperature and humidity. As such, to make use of all available ship-based observations to compare with ERA5, we require a more suitable method to reduce the SST-based biases identified. We considered using the nearest THF that is an ocean point, but this method was not chosen, as this would introduce a bias into the THF magnitude caused by an overestimation of air temperature (not shown).

We create an ERA5 dataset with the atmospheric variables (wind, air temperature, dewpoint temperature, pressure) co-located using the closest ERA5 value and with the SST co-located using the closest cell that has an SST value. We found this reasonable, as the SST variability is less important (range = 1.9 °C) than the air temperature (range = 12.2 °C) in the polynya region during the research cruise (Fig. b, c). Therefore the gradient of temperature in Eq. () mainly depends on the air temperature. The COARE 3.5 algorithm was then applied to compute turbulent fluxes using the new dataset as input (Fig. ). We refer to the original ERA5 THF product as the nearest-neighbor product, which suffers from an inaccurate landmass, and the new product calculated from the ERA5 atmospheric variables and valid SST as the hybrid product (hybrid because of the differences in the co-location method between the SST and the other variables).

The SHF correlation between ERA5 and the observations was higher with the ERA5 hybrid dataset (r2 = 0.576; Fig. a) than the original nearest-neighbor dataset (r2 = 0.186; Fig. a). The same statement holds for the LHF: r2 = 0.658 for the hybrid dataset (Fig. c) versus r2 = 0.291 for the nearest-neighbor flux (Fig. c). The hybrid SHF was on average closer to the observation-based fluxes (lower by 3 W m⁻²; Table ) but with a negative bias because of colder air temperature (Fig. c). Regarding the mean LHF, the hybrid dataset did not bring a clear improvement: the mean hybrid SHF is lower by 4 W m⁻² than the SHF from observations, whereas the mean ERA5 SHF output was higher by 3 W m⁻². The instantaneous heat loss was slightly better represented (Fig. b, d in comparison to Fig. b, d), with a maximum difference between the time series of 107 W m⁻² for the SHF and 69 W m⁻² for the LHF (Table ).

The PWP model predicted a warming of the water column over the 35 d for all four simulations (Fig. g), which is in agreement with the seasonal warming expected during the austral summer. However, the SSTs of the four PWP simulations diverge (Fig. c). At the end of the 35 d run (the duration of the expedition in the polynya region), the nearest-neighbor SST (yellow line) was higher (0.76 °C; Fig. e) than in the other three simulations. The hybrid (blue line) and stationary (pink line) simulations were colder (-0.17 and -0.14 °C) than the observation simulation (dark-green line in Fig. c). This is in agreement with the ERA5 overestimation of heat loss for the hybrid dataset and underestimation for the nearest-neighbor dataset. The HC in the 40 m upper layer was higher by 8.2 × 10⁷ J at the end of the 35 d for the nearest-neighbor simulation and lower by -3.30 × 10⁷ and -2.98 × 10⁷ J for the hybrid and stationary simulations (Fig. f) in comparison to the observation-based simulation.

The observation and the nearest-neighbor simulations had large THF differences (on average 28 W m⁻² and instantaneously up to 200 W m⁻²; Fig. b, yellow points) that led to an increase in the slope of the difference in SST (Fig. e) and HC (Fig. f) between the two simulations. For example, from day 13.5 to day 15.5 (gray-shaded area), the nearest-neighbor THF was on average 82 W m⁻² warmer than the observed THF. This difference explained an SST increase of 0.20 °C and an HC increase of 1.4×10⁷ J (yellow line; Fig. e, f) in the nearest-neighbor simulation in comparison to the observation-based simulation. The cumulative effect of such events is critical to set the SST and HC differences between the simulations throughout the 35 d simulations (Fig. e, f).

We note that the daily change in SST (ΔSST; Table ) has the same order of magnitude between the stationary-dataset-based simulation (mean = 0.003 °C d⁻¹, SD = 0.042 °C d⁻¹) and the hybrid-dataset-based simulation (mean = 0.002 °C d⁻¹, SD = 0.040 °C d⁻¹), which gives us confidence in the credibility of using the PWP model with data that are not stationary (i.e., biases introduced when comparing datasets from moving vessels).

Thus, the four simulations diverge because of the different THF inputs (indeed all the other atmospheric forcings are identical). The resulting difference in evolution of SST (HC) of 0.76 °C (8.2 × 10⁷ J) over a month is not negligible for a coastal polynya region where sea-ice formation and melting, ice-shelf melting and primary production are dominant processes.

In this final subsection, we look at ship-based thermosalinograph (TSG) data (for the SST) and nearby glider data (for the HC) to understand how the temperature change in the PWP simulations compares to the change in temperature in the observations.

ΔSST from the PWP model output (solid lines, Fig. a) and from the TSG observations (dotted line) showed a similar average change (Table ; mean = O(10-3) °C d⁻¹), except for the nearest-neighbor PWP output, which has a higher mean (0.029 °C d⁻¹). Regarding the HC, the nearest-neighbor simulation exhibits a higher mean as well, even though the difference is smaller than for the SST (Table ). SDs of the daily change in SST and HC are 1 order of magnitude higher for the observations (0.140 °C d⁻¹, 2.0 × 10⁷ J d⁻¹) than for the model outputs (O(10-2) °C d⁻¹, O(106) J d⁻¹; Tables  and ). The larger spread in the observations can be explained by the horizontal processes that are assumed to be non-negligible in a sea-shelf environment and that were not represented in the 1D model.

It is thus reasonable to assume that the horizontal processes average out (as seen by the comparison of the means); the 35 d evolution of SST and HC in the observations was well represented by the PWP simulations when the nearest-neighbor simulation was set aside.

Coming back to the 2 d example of the previous section, according to the PWP model, we see a misrepresentation of the THF of about 80 W m⁻² (Fig. b) that leads to an SST increase of 0.2 °C (Fig. e) and an HC increase of 1.4 × 10⁷ J (Fig. f). Across this particular heat loss event, the SST and HC increases were on the same order of magnitude as the spatiotemporal variability scale observed in the 35 d ship-based SST observation (SD = 0.140 °C d⁻¹; Table , Fig. a) and 18 d glider data (SD = 2.0 × 10⁷ J d⁻¹; Table , Fig. b) that were imputed to horizontal processes. Therefore, the cumulative effect of small differences in the mean ΔSST and ΔHC due to a bias in the THF is critical for the temporal-scale variability in SST and HC in the Amundsen Sea. The 1D processes alone cannot explain the change in SST and HC that we see in the observations. However, our work highlights the fact that a consequent bias in the THF can lead to errors in the estimation of the 35 d evolution of the SST and HC that are on the same order of magnitude as the variability due to horizontal processes.

Antarctica is known for its strong surface winds. It might be plausible that the wind contributes significantly to the heat loss by increasing turbulent mixing, yet we did not observe any pattern of increasing wind intensity related to the heat loss (Figs.  and c). A systematic positive bias in the residuals was observed in Fig. , meaning that either ERA5 underestimates the wind speed or the measurements overestimate it. We imputed the residual bias to ERA5, as it has been shown that the wind speed along the Antarctic coastline is underestimated by ERA5 . The properties of the air (cold and dry) transported by the winds are more important for the THF variability than the wind speed itself. As shown, the atmospheric system is thought to be a key component in the Amundsen Sea. Indeed, used the MetUM model to perform episodic (2 to 3 d) high-heat-flux case studies in the eastern Amundsen Sea and found that strong easterlies and southeasterlies associated with cyclones are typically linked to heat loss. The importance of low-pressure systems in cold-air outbreaks is a result that has also been found in the Ronne Polynya (Weddell Sea) from aircraft observations . The Amundsen Sea is known for being a cyclogenesis region, and we indeed saw recurring low-pressure centers in our study area (Fig. d). In particular, the Amundsen Sea Low (ASL) is a quasistationary low-pressure center that oscillates between the Ross Sea (to the west) and the Bellingshausen Sea (to the east). showed that the ASL influences the meridional component of the large-scale atmospheric circulation. When the ASL is positioned to the west, it enhances the southerly flow. The ASL was located at the edge of the Ross and Amundsen seas (to the west) in January and March 2022 and in front of Thurston Island, in the Amundsen Sea, in February 2022 (the ASL index position was downloaded from https://github.com/scotthosking/amundsen-sea-low-index, last access: 31 January 2025). The ASL had, therefore, likely enhanced the southerly (continental) winds over the time span of the research vessel's presence in the Amundsen Sea and consequently increased the turbulent heat loss through enhanced air–sea temperature and humidity gradients. It would be interesting in future studies to investigate the longer-term air–sea THF variability associated with ASL location.

In this study, the largest heat losses occurred in front of the ice shelves. This result could indicate that the heat loss is larger along the ice shelves than in the open water in the polynya but could also be explained by the fact that the research vessel spent 27 d out of 41 d in front of the ice shelves when in the polynya region. However, it has been shown by modeling studies in other coastal polynyas (for the polynya that forms off Ronne Ice Shelf (Weddell Sea) and for the Pine Island Glacier Polynya, Pope Smith Kohler Polynya, and Thurston Polynya (eastern Amundsen Sea)) that the THF decreases with fetch. When continental cold (dry) air is advected on top of the sea, it gains heat (moisture) from the ocean as it travels offshore, which reduces the gradient of temperature (humidity) and, as a consequence, the SHF (LHF). This result gives us confidence regarding the spatial variability observed in our dataset and illustrates the key mechanisms for driving heat loss in regions of ice-shelf dynamics.