Local- and Large-Scale Drivers of Variability in the Coastal Freshwater Budget of the Western Antarctic Peninsula

The west Antarctic Peninsula (WAP) is a region of marked climatic variability, exhibiting strong changes in sea ice extent, retreat of most of its glaciers, and shifts in the amount and form of precipitation. These changes can have significant impacts on the oceanic freshwater budget and marine biogeochemical processes; it is thus important to ascertain the relative balance of the drivers and the spatial scales over which they operate. We present a novel 7-year summer-season (October to March; 2011 to 2018) series of oxygen isotopes in seawater (δ 18 O), augmented with some winter sampling, collected adjacent to Anvers Island at the WAP. These data are used to attribute oceanic freshwater changes to sea ice and meteoric sources, and to deduce information on the spatial scales over which the changes are driven. Sea ice melt shows significant seasonality ( ∼ 9% range) and marked interannual changes, with pronounced maxima in seasons 2013/14 and 2016/17. Both of these extrema are driven by anomalous winds, but reflect strongly contrasting dynamic and thermodynamic sea ice responses. Meteoric water also shows seasonality ( ∼ 7% range) with interannual variability reflecting changes in the input of accumulated precipitation and glacial melt to the ocean. Unlike sea ice melt, meteoric water extremes are especially pronounced in thin (<10 m) surface layers close to the proximate glacier, associated with enhanced ocean stratification. Isotopic tracers help to deconvolve the complex spatio-temporal scales inherent in the coastal freshwater budget, and hence improve our knowledge of the separate and cumulative physical and ecological impacts. cover west of the Peninsula, as well as in the speed of the melt and retreat of glaciers, and in the amount of snow that falls there. Upon melting, these freshwater sources affect the marine ecosystem in distinct ways, by injecting freshwater at different depths and supplying different nutrients, making it imperative to distinguish changes in these sources. Here, we use a new 7-year sequence of seawater measurements to calculate the amount of sea ice that melts into the ocean separately from the amount of snowmelt and glacier melt. We find very large changes in these sources, both on month-to-month and year-to-year timescales, and we identify causes of these changes. We find that some changes are driven by large-scale wind and precipitation variability, whereas other changes are caused much more locally. These findings are relevant to the marine life west of the Peninsula, and how it will respond to future climate change.

. At the WAP, winter sea ice extent and concentration showed moderate increases over the period since the turn of the century (Turner et al., 2016). This contrasts with the annual ice season duration, which showed marked increases in yearly variability post-2000, and which was punctuated by anomalously short annual ice seasons during 2007-2010, followed by anomalously long annual ice seasons during 2013-2016 (Henley et al., 2019). However, a marked retreat of sea ice to the west of the Peninsula and across the northern Weddell Sea was observed in spring 2016, coincident with an unprecedented circumpolar retreat of sea ice (Parkinson, 2019;Turner et al., 2016). Through 2019 and mid-way into 2020, circumpolar sea ice extent, including at the WAP, remained below the long-term mean, particularly during the summer-autumn period (Reid et al., 2020).
The multi-decade warming and climatic variability at the WAP has been concurrent with many other changes in the different components of its freshwater system. Atmospheric warming has been implicated as potentially important in glacial ice loss at the WAP, which is the main region in Antarctica with temperatures sufficiently high that significant surface melt and runoff may occur (Vaughan et al., 2003). The decadal increase in melt season duration at the WAP is consistent with the observed progression of ice shelf collapses (Scambos et al., 2000;van den Broeke, 2005), and while the majority of meltwater that percolates downwards will refreeze in the firn layer, there remains the possibility of significant runoff to the ocean (Barrand et al., 2013). Notwithstanding this, the observation that the majority of glaciers at the WAP are in retreat, with retreat rates accelerating (Cook et al., 2005;Pritchard & Vaughan, 2007), has more recently been ascribed to the influence of the heat from deep ocean waters. This is supplied in the form of warm Circumpolar Deep Water (CDW) from the Antarctic Circumpolar Current (ACC) penetrating onto and across the shelf to the WAP coastal regions, especially via deep glacially scoured canyons (Martinson & McKee, 2012;Venables et al., 2017). There is marked spatial variability in this process along the length of the WAP, resulting in marked latitudinal differences in retreat rates (Cook et al., 2016).
Precipitation at the WAP has increased in recent decades, with measurements from research stations indicating more precipitation events and a strengthening tendency for precipitation to fall as rain instead of snow (Kirchgäßner, 2011). A doubling of precipitation has been inferred for some WAP locations, using records from shallow ice cores (Thomas et al., 2008). Accumulation of snow on sea ice and its subsequent melt will contribute to the seasonal meteoric water budget at the WAP, although changes in snow accumulation on sea ice are currently not well constrained.
Each of these changes can exert profound influence on the physical, biogeochemical, and ecological functioning of the ocean. Freshwater injected at the ocean surface can create stably stratified layers, and hence can create environments more favorable for phytoplankton blooms (Mitchell & Holm-Hansen, 1991;Montes-Hugo et al., 2009;Saba et al., 2014;Venables et al., 2013;Vernet et al., 2008). Conversely, the destruction of stratification (e.g., by sea ice production) can have the opposite effect. The impacts of changes in stratification on phytoplankton dynamics can have consequences for the regional oceanic uptake of CO 2 . Solid-phase sea ice is mobile under the influences of wind forcing and advection by ocean currents, the spatial variability of which can lead to complexity in the distributions and impacts of sea ice melt. It also constitutes an important habitat for a range of species from across trophic levels, and thus changes in sea ice have the potential to induce pronounced ecosystem changes adjacent to Antarctica Thomas & Dieckmann, 2002).
Glacial discharge to the ocean also has distinct impacts on regional physical oceanography, marine biogeochemistry, and the ecosystem. While low levels of micronutrients, such as iron, tend to limit primary production across much of the open Southern Ocean, its supply in coastal waters relieves this limitation and can lead to intense and extensive plankton blooms. Shallow sediments in the coastal zone can be influential in this supply, as can glaciers that scour underlying rock and sediments prior to supplying freshwater to the coastal seas (Annett et al., 2017;Boyd & Ellwood, 2010;Raiswell, 2011;Sherrell et al., 2018). Stronger glacial discharge to the ocean has the potential to deliver increased micronutrients (Hawkings et al., 2014), and could impact phytoplankton assemblage composition and biomass (Moline et al., 2004;Sherman et al., 2020). For example, Dierssen et al. (2002) noted that >70% of the phytoplankton blooms observed offshore of Anvers Island (Figure 1) occurred soon after meltwater pulses from local glaciers. The depth at which glacial melt enters the ocean has the potential to be a key factor in its physical and ecological impact. Shallow injection from surface runoff or near-surface glacial melt will tend to increase stratification, and will supply any micronutrients or other biologically-active substances carried by the meltwater directly into the euphotic zone . Conversely, glacial melt that occurs at depth due to melting by warm deep waters (or surface runoff that has percolated down to the glacier bed before reaching the ocean) will typically result in rising buoyant plumes that ascend until they reach a level of neutral buoyancy, which can be close to the surface or below the euphotic zone (Cape et al., 2019).
A further variant of the impact of glacial discharge is via the calving/drifting of icebergs, and while the WAP is not typically influenced by large icebergs, it is impinged upon by a significant number of smaller ones (Tournadre et al., 2016). These icebergs can affect primary production by modifying the freshwater distribution and water column stability, and the ocean's geochemical properties (Duprat et al., 2016). They can also impact benthic ecosystems destructively by scouring when their keels extend to the seabed (Smale et al., 2008).
Climatic change at the WAP continues, and recently there have been observations of 32-year record-high surface melt rates at George VI Ice Shelf, concurrent with atmospheric temperatures over 20°C at the northern WAP that have been dubbed an Antarctic heatwave (Banwell et al., 2020). Given these extrema and their profound potential implications, combined with limited predictive skill concerning their future evolution, it is important that we improve our mechanistic understanding of the WAP system and its response to climate change.
In this study, we use a novel 7-year (2011-2018) series of freshwater tracer data from a coastal WAP site to quantify and attribute the changes in the freshwater input to the ocean, with a particular emphasis on deconvolving locally-driven changes and those driven on regional and larger scales. By doing so, we derive new information on how the WAP freshwater system has changed in recent years (including a period of rapid change in some of the key freshwater elements), and generate refined mechanistic understanding that will help improve knowledge of how it may change in future as climate change progresses.

Sample Acquisition and Processing
Water samples were collected quasi-weekly during summer (typically, from late October to March) from two sites near Palmer Research Station at Anvers Island at the WAP ( Figure 2). The sampling sites, Palmer Stations B and E (Figure 2), are within the boating limits at Palmer Station. Station B (64.7795°S; 64.0725°W) is less than 1 km from the nearest point on Anvers Island, and a short trip from Palmer Station. This location affords consistent access in light to moderate sea ice cover. Station B is protected by islands to the north, east, and west, but exposed to the open Bellingshausen Sea to the south. Station E (64.815°S; 64.0405°W) is close to the limit of permitting boating operations, and exposed to the continental shelf and open sea. Stations B and E are in water 75 and 200 m deep, respectively. The start of the sampling season, constrained mainly by sea ice preventing boat access, ranges from mid-October to late December. During winter, boat sampling is not conducted; instead, a program of sampling at Palmer Station's seawater inlet (SWI: 64.7738°S; 64.0545°W) was initiated. The intake for this system is at 5.8 m depth. The sampling sites reside at different levels of proximity to the shallow marine-terminating Marr Ice Piedmont and fringing land-based ice masses, including some on the smaller adjacent islands (Figure 2). Typical water depths adjacent to the Marr Ice Piedmont are just a few tens of meters, with a deeper channel running past Palmer Station and Station B to the more open shelf.
Water samples were obtained from these sites using Niskin bottles deployed from a rigid inflatable boat, and closed either at the surface or 10 m depth. From the Niskin bottles, samples were drawn into 50 ml glass vials, which were sealed with stoppers and aluminum crimps. These samples were transported by dark cool stow to the UK's National Environmental Isotope Facility at the British Geological Survey. They were then analyzed for their oxygen isotope composition (δ 18 O, the standardized ratio of 18 O to 16 O), using the CO 2 equilibration method with an Isoprime 100 mass spectrometer plus the Aquaprep device. Isotope measurements were calibrated against internal and international standards, including VSMOW2 and VSLAP2. Based on duplicate analysis, analytical reproducibility of around ±0.02‰ was obtained for these samples. Concurrent with the Station B and E sampling for δ 18 O, profiling of the water column was conducted using a SeaBird SBE19+ Conductivity Temperature Depth (CTD) profiler. From these casts, mixed layer depth (MLD) was derived using a criterion based on the vertical maximum of buoyancy, as per Carvalho et al. (2017). Ocean data used here are available at https://portal.edirepository.org/nis/mapbrowse?packageid=knb-lter-pal.309.1.

Mass Balance Calculations
Using the δ 18 O and salinity data, we quantitatively separated sea ice melt from meteoric water (glacial melt and precipitation) by solving a simple 3-endmember mass balance. This was originally developed for the Arctic (Östlund & Hut, 1984), and was adapted for freshwater studies at the WAP by Meredith et al. (2008). It imposes an assumption that each water sample is comprised of three constituent components, namely sea ice melt, meteoric water, and CDW. At the WAP, this balance can be expressed as follows: where F sim , F met , and F cdw are the fractions of sea ice melt, meteoric water, and CDW that are being solved for; S sim , S met , and S cdw are the salinities of the pure endmembers of the respective constituent waters; δ sim , δ met , and δ cdw are corresponding δ 18 O values for these endmembers; and S and δ are the measured salinity and δ 18 O of the water sample. Using this mass balance, negative values for sea ice melt are possible and are indicative of net sea ice production from the waters sampled, that is, a net salinification from brine rejection as opposed to a net freshening from sea ice melt. This decomposition has been used successfully at the WAP previously, in Marguerite Bay (Figure 1) (Meredith et al., , 2010, at King George Island in the South Shetlands , and across the width of the shelf between Anvers Island and as far south as Charcot Island . In each case, endmember values are chosen in order to reflect the local/regional freshwater and oceanic conditions as closely as possible. For the Palmer Station series, the endmembers chosen are given in Table 1. The CDW endmember values were based on measurements collected close to Anvers Island, as per Meredith et al. (2017). The sea ice melt values are appropriate values for the WAP based on previously published analyses . The mass balance (Equation 1) is most sensitive to the meteoric water endmember chosen; this is chosen here based on previous analyses and literature (Corbett et al., 2017;Meredith et al., 2013Meredith et al., , 2017, and values from sampling in meltwater streams immediately adjacent to Palmer Station (Mike Brown; Rutgers University; pers. comm.). Sensitivity analyses have shown that this set of equations yields absolute freshwater concentrations with an accuracy better than 1%, and that the series produced are internally consistent such that temporal changes in freshwater concentrations are more precisely resolved .

Palmer Station Meteorological Data
The meteorological instrumentation at Palmer Station is operated by the University of Wisconsin, and include sensors to measure air temperature, rainfall, snow precipitation, and snow depth. Full details and data are available at https://portal.edirepository.org/nis/metadataviewer?packageid=knb-lter-pal.214.3. Positive degree days (PDD) were calculated as the integral of positive surface air temperatures, starting on the 1st of August each year and extending to 31st of July of the following year. This comfortably encompasses the main period when surface air temperatures are persistently above zero: the PDD values remain low until the onset of spring, then rise rapidly before approaching an asymptote at their annual maximum in autumn/ winter.

Regional Atmospheric Climate Model 2
We used simulated precipitation from version 2.3p2 of the Regional Atmospheric Climate Model 2 (RAC-MO2; van Wessem et al., 2018), which has been configured for use in glaciated regions. It was run with a horizontal resolution of 5.5 km, and with 40 vertical levels. It was forced at the boundaries using ERA-Interim reanalyses (Dee et al., 2011). Model output used here extended from 2011 to 2019, thus encompassing the period of in situ observations used. We demonstrated previously that this model provides a realistic depiction of the WAP meteoric water system, including its response to changes in coupled modes of climate variability (van Wessem et al., 2017).

Climate Data
Several sea ice products, ranging in resolution, are used to characterize regional changes in sea ice at the WAP. The coarsest resolution data comprise the monthly sea ice concentrations from Version 3.1 of the Goddard Space Flight Center (GSFC) Bootstrap Scanning Multi-channel Microwave Radiometer-Special Sensor Microwave/Imager (SMMR-SSM/I) time series (Comiso, 2000;Comiso & Nishio, 2008), here spanning 1979 to 2018 (40 years). These sea ice concentration data are gridded to 25 km and are provided by the EOS Distributed Active Archive Center (DAAC) at the National Snow and Ice Data Center (NSIDC, University of Colorado at Boulder, https://nsidc.org/data/NSIDC-0051/versions/1). The advantage of these coarser-resolution sea ice concentration data is that they comprise the longest time series available, and are used here to characterize monthly anomalies based on their 40-year means.
In contrast, higher-resolution sea ice concentration data are used to characterize finer-scale features of the sea ice cover on key days of interest. These daily sea ice concentrations are from the ARTIST Sea Ice (ASI) MEREDITH ET AL. algorithm version 5 (Spreen et al., 2008) using Advanced Microwave Scanning Radiometer 2 (AMSR2) data. These sea ice concentration data are gridded to 3.125 km and are available from the Universität Bremen (https://www.seaice.uni-bremen.de) for the period since 2012.
Monthly winds and weekly ice motion data are used to infer dynamical changes in the sea ice cover. The numerically-analyzed monthly 10-m height winds are from the European Centre for Medium Range Weather Forecasts (ECMWF) Interim Reanalysis (ERA-I) (Bracegirdle, 2013;Dee et al., 2011), from which we determined monthly wind anomalies based on their 40-year means. The weekly ice motion data are from the Polar Pathfinder 25 km EASE-Grid sea ice motion vectors, version 4. Derived ice motion vectors are based on a combination of satellite (AVHRR, AMSR, SMMR, SSMI/S) and buoy data using methods described by Tschudi et al. (2020), and are provided by the EOS DAAC at the NSIDC.

Variability in Salinity and δ 18 O
While year-round ocean data from Palmer Station are not available for most of the period investigated here, clear seasonality and marked interannual variability are still discernible in each of the series produced (  Stn. E with summer seasons of 2015/16 and 2016/17 marking interannual cold and warm extremes, respectively. Salinity varies between around 33.0 and 34.0 on seasonal timescales, with some extreme fresh occurrences (down to 32.0) present during the summer months. δ 18 O constitutes the most complete record amongst our data, since sample collection was enabled ahead of salinity and temperature logging from the SWI. The general seasonal cycle for δ 18 O varies between around −0.6‰ and −0.2‰, with sporadic isotopically-light instances in summer, yielding values as low as −1.0‰. We have confidence that these instances are not errors in the data, since all spikes are duplicated and the data quality assured. The marked seasonality in δ 18 O is an immediate indicator that the seasonality in salinity is not caused solely by sea ice processes. Typically, the freshest, isotopically-lightest waters are sampled toward the end of the summer sampling period (March/April), when accumulated freshwater inputs have reached their maxima and ahead of the onset of winter. MLD is typically shallower than 50 m, with some instances approaching 90 m. It is noteworthy that the fresh, isotopically-light extrema are predominantly at Site B and occur in the surface layer sampling, and that these coincide with the low values for MLD.
Further insight into the processes underlying the temporal changes in freshwater composition can be gained by considering the data in salinity-δ 18 O space ( Figure 4). Addition of meteoric water moves the locus of points diagonally downwards toward fresher, isotopically-lighter values, whereas mixing with saline ocean water has the opposite effect. Sea ice processes move the locus of points almost horizontally, with sea ice melt adding freshwater but having minimal effect on δ 18 O, and sea ice formation raising the salinity via brine rejection but again having minimal impact on δ 18 O.

Variability in Sea Ice Melt and Meteoric Water
Changes in the relative freshwater contents can be quantified using the mass balance derived from Equation 1. This shows that sea ice melt varies seasonally between around −2% and +2% in general ( Figure 5), though full winter coverage is not available due to incomplete overlapping series of salinity and δ 18 O. There is distinct interannual variability in peak values superposed on the seasonality. The high seasonal values of F sim in January/February 2014 are consistent with the general shelf-wide conditions at that time, as reflected in January 2014 cruise data that showed elevated values of sea ice melt across the continental shelf to the shelf break and beyond, compared with the previous three Januaries .
Meteoric water shows seasonal ranges that typically span 2%-5% ( Figure 5). Values for F met are typically higher than those for F sim at any time during the sequence of measurements ( Figure 5, second panel cf. first panel). This echoes results from a time series site further south along the WAP (in northern Marguerite Bay; Figure 1), and indicates the general dominance of meteoric water in setting the overall mean freshwater levels. The WAP-wide cruise data reflect this general meteoric water dominance also, though with some notable counter examples (Meredith et al., 2010. As with sea ice melt, distinct interannual variability in meteoric water is present, with peak values varying from around 4% (2011/12; 2014/15) up to 5 or 6% ( Figure 5). Superposed on this are many shallow extrema with high values of F met ; these extend values up to 8%, and are exclusively from the surface water sampling MEREDITH ET AL.
10.1029/2021JC017172 9 of 22 at Station B ( Figure 5, lower panel). (Note that there are no surface F met values at Station E that exceed 6%; however, the surface sampling at Station E is markedly less frequent than that at Station B).
Within the sampling conducted here, a recurring feature is the general increase in F met during the first few months of each season ( Figure 5). This is the converse of salinity (Figure 3), which has a tendency to decline over this period as freshwater is accumulated in the upper layers of the ocean. The F met increase indicates that this freshening is normally dominated by meteoric water input (glacial melt, precipitation, runoff from melted snow), rather than sea ice melt. Melt of snow that accumulates on sea ice over winter could, in theory, contribute to the F met increases over this period, but this would likely be paralleled by an increase in F sim , which is mostly not a clear signal ( Figure 5, upper panel). The extreme high values of F met (>6%) are all in the later part of each sampling season (January to March; with none during October to December).
While the restriction of sampling to the summer window means that not all years have well-characterized peaks in F met , when such peaks are present it is apparent that they have a tendency to occur 1-2 months after the peak in F sim ( Figure 5). This likely reflects the overall reservoir size for the freshwater components and the timing of their release into the water column, with sea ice having a tendency to be depleted first during the melt season while meteoric water input (in the form of glacial ice, snow, and precipitation) can persist longer.

Controls on Sea Ice Melt: Local-and Large-Scale Influences
As noted above, the mass balance shows that sea ice melt varies generally between around −2% and +2%, suggesting that the region is in an approximate balance concerning sea ice formation and melt over the period 2012-2018. However, this does not inform on the spatial scales on which the processes affecting our derived F sim are driven, or how these might change. To investigate this, we interpret the freshwater series presented above alongside the climatological and local meteorological data.
The strong peak in F sim in 2013/14 was studied previously using the shelf-wide cruise data  and attributed to northward wind anomalies that persisted during September and December 2013 in particular, resulting in a long ice season duration and a very late spring retreat of the ice edge ( Figure 6). This tended to retain the sea ice in situ, allowing elevated levels of sea ice melt to become established as the melt period progressed. These findings were corroborated with time series data from adjacent to Rothera Research Station (Figure 1), emphasizing the large-scale relevance of the results. Significantly, Meredith et al. (2017) noted that the timing of the sea ice retreat (as opposed to the maximum winter sea ice extent preceding the melt period) was the key factor in controlling the elevation of F sim .
This overall concept of the cause of high F sim in 2013/14 is still consistent with the data from the more northward-located coastal sampling site used here, despite the broader WAP shelf region experiencing strong cyclonic winds in October, with southward winds causing an earlier ice edge retreat further offshore of this sampling site ( Figure 6). However, the ice edge retreat at the WAP was thereafter slowed by more eastward to northward winds during November and particularly December. Thus, the concurrence of the high F sim peak seen here with comparable anomalies seen in the shelf-wide cruise sampling and time series data at Rothera emphasizes that this high F sim anomaly reflects a large-(WAP-wide) scale modulation of the sea ice field associated with anomalous northward wind forcing, particularly in the key month of December.
In our Palmer Station series, there is a second occurrence (2016/17) of values for F sim that are as high as the peak values seen in 2013/14. This peak occurred earlier in the relevant season (December, as opposed to January; Figure 5), one consequence of which was that the shelf-wide cruise sampling that occurred in January 2017 did not capture it (not shown). Quasi-weekly time series data from Rothera Research Station (Figure 1) also do not show elevated F sim in this season, indicating that the peak at Palmer Station is more localized than the 2013/14 peak (Meredith, 2019).
Our interpretation of the elevated F sim in 2016/17 again relates to anomalous wind forcing, though of a profoundly different nature and scale. Figure 7 shows sea ice concentrations for the period September 2016 to February 2017, with wind anomaly vectors overlain. Of particular note are the strong south to southeastward wind anomalies that persisted in September and October 2016; these anomalous winds pushed a rapid ice edge retreat and forced its location well south of its climatological mean position, where it remained southward for the rest of the year, particularly in the southwest quadrant of the WAP. It may seem initially counterintuitive that these persistent southeastward winds, particularly in September, could create an anomaly in F sim equivalent to that seen in 2013/14, especially when the winds were from an opposite direction. However, the critical factor in 2016/17 is the geographical fate of the wind-driven sea ice, which appears to have been compacted rapidly into the coastal WAP regions where it was mechanically thickened. This created a large volume of solid-phase sea ice available to be melted locally over time as the melt period commenced and progressed.
This interpretation is supported by high-resolution sea ice imagery (Figure 8), which shows a narrow sea ice field in late 2016 with a remarkably linear sea ice edge. This is consistent with rapid wind-driven compaction of the sea ice field, such as the one noted previously at the WAP (Massom et al., 2008). By contrast, late 2013 shows a much more uneven ice edge, consistent with persistent wind anomalies that led to divergence and spreading of the sea ice field. While ice thickness measurements to corroborate the mechanical thickening in the coastal region are unavailable, further support from our interpretation derives from satellite-derived sea ice motion. This shows toward-land advection of ice throughout September 2016 (e.g., Figure 9), after which the band of ice becomes very narrow and almost motionless (not shown), that is, a thickened compacted ice cover becomes far less sensitive to wind-driven motion than an uncompacted thinner ice cover.
That the positive sea ice melt anomaly in 2016/17 is absent from the quasi-weekly time series data collected at Rothera (Figure 1; see also Meredith, 2019) may seem at odds with the linear ice edge extending across a large length of the WAP (Figure 8, lower panel) and wind-driven ice advection also spanning this scale (Figure 9, lower panel). However, the Rothera time series site is in northern Marguerite Bay and is shielded to the west by Adelaide Island. The high-resolution sea ice imagery (Figure 8) shows several patches of sea ice within Marguerite Bay at this time that had lower concentration than that of the general band of compacted sea ice along the WAP, likely partially a reflection of this more enclosed environment. The daily high-resolution sea ice images from November to January (not shown) also confirm that the more protected MEREDITH ET AL.
10.1029/2021JC017172 12 of 22 area adjacent to Rothera became ice-free much earlier than the southern portion of Marguerite Bay, the latter being more exposed to onshore winds, and thus more similar to the rest of the WAP coastal region which experienced anomalously late ice melt.
Overall, the difference in causality between the high F sim events of 2013/14 and 2016/17 can be characterized as fundamentally a different phasing of the dynamic and thermodynamic sea ice responses to anomalous forcing. In 2013/14, the sea ice showed a marked dynamic response concurrent with a thermodynamic one, with the ice being retained largely in situ during the melt period. By contrast, in 2016/17 there was a strong dynamic response and subsequently a strong thermodynamic response, with marked ice advection ahead of its melting in the coastal regions. The location of our sampling station, adjacent to Anvers Island, appears to be sensitive to both of these phasings of the dynamic and thermodynamic sea ice response, that is, to both divergent (2013/14) and convergent (2016/17) wind forcing. (Note that we do not have the surface ocean velocity information from which the role of the ocean in influencing the sea ice response can be assessed; nonetheless, the consistency between the wind fields, ice motion, and ice melt indicates that this latter influence acted in tandem with the atmospheric forcing or had little effect on the overall net response).
Previously, periods of anomalous winds at the WAP and the changes in freshwater composition that they induce have been related to coupled modes of climate variability including the Southern Annular Mode (SAM) and the El Niño/Southern Oscillation (ENSO) phenomenon (Marshall, 2003;Meredith et al., 2013Meredith et al., , 2017van Wessem et al., 2017). We note that the rapid retreat of sea ice in late 2016 coincided with a very strong negative SAM, which was at its lowest in November since 1968 (Turner et al., 2017). It also coincided with a deepening of the Amundsen Sea Low, a largescale atmospheric pattern that spans this part of Antarctica, and which modulates meridional winds west of the Antarctic Peninsula. This reinforces the concept that large-scale modes of climate variability exert pronounced influence on the distributions of freshwater at the WAP; our results here have demonstrated significant new complexity in the spatial scales of the response of sea ice melt.

Extreme Peaks in Surface Meteoric Water
The extrema in meteoric water (black dots in Figure 5) elevate the seasonal peaks of F met by up to an additional 2%. Notably, they occur almost exclusively in the surface sampling (not at 10 m or the SWI at 5.8 m) and at Station B. This is the sampling site closest to land and the Marr Ice Piedmont (Figure 2), indicating that this is a very local-scale freshwater input. Candidate processes are the melt of snow that accumulates on land over winter, or surface/facial runoff from the glacier. Other possibilities are less compelling as explanations; for example, direct precipitation would be spread over a much broader area and thus feature at both sampling sites, as would melt of snow that had accumulated on top of sea ice. Further, while glacial melt from depth driven by intrusions of warm CDW can reach the surface in Antarctic coastal regions, in the form of upwelling buoyant plumes, this would be associated with pronounced mixing and would not be constrained to the very surface layer.
MEREDITH ET AL.  Of the cluster of samples with δ 18 O less than −0.75‰ (Figure 3; third panel), all but one coincided with a mixed layer shallower than 10 m depth. This is unsurprising, since freshwater is the controlling term in the equation of state at low temperature, and hence fresh layers add greatly to stratification and upper-ocean stability. That these extrema are only ever seen in meteoric water (not sea ice melt; Figure 5) indicates that the freshwater injection in this form actively creates these layers, and the observations are not reflecting simple freshwater accumulation in an already-established stable stratification.
Also of relevance in the MLD data is that, while the general mixed layers are shallower than 50 m, there are occasional deep spikes that lower the mixed layer to 90 m or more (Figure 3). These occur predominantly at Station B, again closest to land and the proximate glacier. There are several processes that could cause these deepening events, including wind-induced coastal upwelling and the influence of internal tides, such as have been investigated in detail at Rothera . It is not our intention to explore these in detail here, but we note that such occurrences would disrupt the shallow surface layers when present.
Overall, we derive a picture of irregular pulses of meteoric water building stratified layers which can be disrupted periodically by mixing events. This has biogeochemical and ecological significance: for example, if the freshwater layers contain micronutrients, it is beneficial to their fertilization of the local phytoplankton blooms that this water is injected into the euphotic zone. Conversely, mixing events would distribute surface-injected freshwater over a greater vertical range and affect the fate of any geochemical substances associated with it. Upper water column stability is critical in initiating phytoplankton blooms in this region (Carvalho et al., 2016;Schofield et al., 2017) Against this background, it is worth noting that mixed layer deepening does not appear to be the same pronounced control on interannual variability in upper-ocean freshwater that is observed at the Rothera time series (Meredith et al., 2010. Close to Rothera, interannual changes in the seasonal upward entrainment of CDW is a dominant factor in controlling the upper-layer freshwater content. The deep mixed layers that can be observed at Rothera (down to 150 m in some winters) are a combination of buoyancy forcing in winter, especially related to sea ice processes, and wind-driven mixing. The latter process becomes markedly more effective in the absence of persistent fast ice during winter (Venables & Meredith, 2014). It is clear that the environs of Palmer Station are not exposed to the same level of extreme forcing by these processes as is Rothera.

Seasonal and Interannual Changes in Meteoric Water
Concurrent with the local-scale, intermittent injections of freshwater noted above are distinct seasonal and interannual changes in meteoric water content that occur at each of the Palmer Station sites and each of the depth levels sampled ( Figure 5). The seasonality of F met is such that it typically peaks in the January-March window, while the peak precipitation period at the WAP is typically during winter. This phase difference potentially reflects a combination of processes, including large-scale melt of accumulated snow from land along the length of the WAP, seasonality in glacial discharge, and the melt of snow on sea ice. We examine these possibilities here to gain insight into which is the most important in determining the seasonal and interannual variability of our meteoric water series.
With regard to the accumulation and melt of snow on land, snowstake measurements at Palmer Station show clearly the seasonal cycle in this, with pronounced accumulation occurring from around April/May and melt occurring typically during October to December (Figure 10). This reflects well the timing of the seasonal increase in F met ( Figure 5).
On interannual timescales, distinct changes in meteoric water concentration close to Palmer Station are present in our series ( Figure 5, third panel), with 2011/12 showing the lowest seasonal peak (around 4%) and 2016/17 the highest (around 6%). Previously, we explained interannual variability in meteoric water concentrations across the WAP shelf as seen in January cruise data by relating the changes seen in precipitation Depth at snowstake (cm) anomalies taken over the last 3 months of the preceding year . Such anomalies, derived from RACMO2, are shown here in Figure 11, and it is clear that the relationship previously established breaks down for the Palmer Station series, with the highest rates of precipitation occurring in late 2011, and comparatively low rates in 2016.
Instead, we demonstrate here that a better relationship is obtained by considering precipitation over the winter period (April to October) that precedes the seasonal peak in F met at Palmer Station ( Figure 12). The choice of this window is based on the snowstake measurements at Palmer Station, which inform on the period during which snow accumulates on land and subsequently melts ( Figure 10). When precipitation anomalies are averaged over this seasonal window (Figure 12), it is clear that 2011 featured moderately strong negative anomalies at Palmer Station, while 2016 showed distinct positive anomalies: these coincide MEREDITH ET AL.
10.1029/2021JC017172 16 of 22 Figure 11. Anomalies in precipitation at the Antarctic Peninsula in Regional Atmospheric Climate Model 2 (kg m −2 s −1 ), covering the period 2011 to 2018. Months October to December are used here, being the period used previously in explaining the shelf-wide meteoric water anomalies observed in January cruise data .
well with the interannual changes in meteoric water. Other years lie intermediate between these extremes, though with complexity due to combined processes; for example, 2012 showed anomalously low precipitation in RACMO2 ( Figure 12) which resulted in generally low F met at Palmer Station in 2012/13 ( Figure 5, third panel) that was elevated at the very end of the sampling season by a burst of direct precipitation ( Figure 10).
Overall, while correlations are not possible due to the small number of years of data, a greatly improved relationship is seen, indicating that changes in the winter accumulation of precipitation is generally a more important control on the peak level of F met seen at Palmer Station the following summer than is direct precipitation in spring. The precipitation anomalies are coherent in a band along much of the WAP coast ( Figure 12), suggesting that while local accumulation is a key factor, this process likely occurs similarly over larger scales and will influence freshwater concentrations coherently along much of the coastal ocean.
MEREDITH ET AL.
10.1029/2021JC017172 17 of 22 Figure 12. Anomalies in precipitation at the Antarctic Peninsula in Regional Atmospheric Climate Model 2 (kg m −2 s −1 ), covering the period 2011 to 2018. Months April to October are used here, being the general period over which precipitation tends to accumulate on land/sea ice during winter ahead of the spring melt season (see Figure 10).
General southward transport in the WAP coastal circulation will aggregate the inputs and transport them poleward (Moffat et al., 2008), thus integrating the local signals over much larger scales.
We thus find that a key factor influencing interannual changes in F met at Palmer Station site is that different amounts of accumulated snow build up over winter and are available to melt during spring and summer, thus injecting comparatively large amounts of meteoric water into the coastal ocean over relatively short periods. In addition, advection of water that has been similarly influenced from along the WAP coast will be an important factor, and while Anvers Island is typically thought to be close to the head of the southward-flowing Antarctic Peninsula Coastal Current (Moffat et al., 2008), the injection and integration of this freshwater will be an important influence on its transport and properties.
While the relationship between precipitation and F met succeeds in explaining some of the most pronounced anomalies in the series, it is inevitably not perfect. Some of this is due to the fact that accumulated precipitation represents the potential for seasonal meteoric water input to the ocean, but the actual input depends also on the evolution of local melt conditions at the sites of input. Local meteorological data recorded at Palmer Station give some insight into this. The snowstake measurements (accumulated snow; Figure 12) show strong interannual variations, including comparatively low accumulation in winter 2011, which preceded the lowest seasonal F met peak in the ocean time series, and much higher levels during the winter of 2016 prior to comparatively higher seasonal peak in F met in 2016/17. However, the highest accumulated snow occurred in 2015, and while F met in 2015/16 was higher than in 2011/12, it was not as high as in 2016/17. The potential for melt, as expressed by PDD, goes some way to explaining this difference: PDD was actually the lowest within our series in 2015/16, and the highest in 2016/17 ( Figure 10). Thus we observe that the accumulation and seasonal discharge of snow at the coastal WAP, with the latter modulated by interannual changes in melt potential, are key processes that determine interannual changes in meteoric water concentrations at Palmer Station.
We find that other processes considered are less influential in determining the seasonal and interannual changes in our meteoric water time series. Concerning glacial melt, and specifically that driven by atmospheric or surface ocean heat (Figure 3, upper panel; Figure 10), we note that the series of surface temperature or PDD by themselves do not adequately explain variations in F met ( Figure 5 cf. Figure 12). Further, glacial melt at depth driven by warm CDW is unlikely in the vicinity of our sampling site, since the local ice masses are notably shallow-or land-terminating, with maximum grounding depths of just a few tens of meters ( Figure 2). This does not negate the quantitative importance of glacial discharge in the temporal-mean freshwater balance of our study area, or the WAP more generally, but it does argue against it being the dominant process in controlling the variability in meteoric water as measured.
A different process that could potentially contribute to the changes in meteoric water is the seasonal accumulation and melt of snow on sea ice, and interannual changes therein. This is multi-faceted process that is not well constrained by data, however if it were to dominate we would expect a significant relationship to exist between the meteoric water and the sea ice melt as quantified in our series. Such a relationship is absent, with the seasonal phase of these series and their interannual changes not coinciding ( Figure 5). Significant work would be required to investigate this further; on the basis of existing information, however, this does not appear to be a leading-order process that determines our meteoric water series.

Summary and Conclusions
We have observed marked variability in the freshwater balance of the coastal WAP adjacent to Anvers Island. Interannual variability is driven by both sea ice and meteoric freshwater sources, and in both cases a range of spatial scales are inferred to be important in driving the variability. It is important to deconvolve this complexity, in order to better understand the totality of the causal factors and hence move toward better predictive skill.
In relation to sea ice melt, the results confirm the large-scale impact of anomalous northward winds in late 2013, which resulted in a late sea ice retreat and anomalously high sea ice melt concentrations across the WAP shelf in early 2014. Intriguingly, a similarly high level of sea ice melt is observed in late 2016, despite wind anomalies of a very different direction. We have shown that this second instance was caused by sea ice being compacted and mechanically thickened in the coastal regions of the WAP, and thus constituting a large reservoir of freshwater that was released locally when the melt season began. In essence, there is a complex balance and interplay between the dynamic and thermodynamic influences on the derived sea ice melt concentrations, each of which maps onto different spatial scales. This adds complexity and challenges the development of simple predictive relationships.
We infer that, despite having similar magnitudes of sea ice melt at Anvers Island in these two seasons, the spatial scale of anomalies of these events across the WAP will be markedly different. While this cannot be assessed using the data from the January cruises at the WAP (since the 2016 anomaly had passed by early 2017), this is consistent with the causal mechanisms deduced above and is lent support by the absence of the 2016 signal in the time series at Rothera (northern Marguerite Bay). We have shown that the δ 18 O sampling at Palmer Station is sensitive to both these scales, and correct determination of the sea ice melt changes is required in order to reliably attribute their local biogeochemical and ecological consequences.
The coastal meteoric water budget is seen to possess even more complexity across a range of spatial scales. Very locally, shallow layers of elevated meteoric water are seen close to the Marr Ice Piedmont adjacent to Palmer Station, which we deduce are caused by glacial runoff and/or snow melt in thin surface layers. These freshwater pulses cause intermittent enhancement in stable upper-ocean stratification, which is destroyed periodically, most likely by mixing events close to the coastal boundary.
We have observed that seasonal peaks in meteoric water have a tendency to occur 1-2 months after the peaks in sea ice melt. This is likely a consequence of differences in the overall reservoir size for these freshwater components, with sea ice having a tendency to be depleted first during the melt season while meteoric water input (in the form of glacial ice and snow, augmented with any direct precipitation) can persist longer. This has consequences for the physical and biogeochemical functioning of the coastal ocean; while the physical effects (e.g., stabilization of the upper ocean) may appear similar, the geochemical composition of these freshwater sources differs greatly.
Further, we have observed distinct interannual changes to the seasonal peaks of meteoric water concentration, and attributed them to changes in the amount of accumulated snow the preceding winter, combined with local melt potential from seasonal atmospheric warming. These mechanisms are dependent on the immediate proximity of a coastal boundary that can store freshwater for seasonal release during the melt season, and differ from the large-scale, WAP-wide mechanisms described previously. Input of freshwater to the coastal ocean at the WAP is important in driving a southward-flowing coastal current; this will integrate inputs along the length of its flow, and generate cross-shelf gradients in properties.
It has been noted previously that precipitation at the WAP is sensitive to large-scale modes of climate variability, including ENSO and SAM (van Wessem et al., 2017), and that this has the potential to build predictive skill if the future trajectories of these modes can be projected with some level of confidence. With our comparatively short time series, we cannot address directly the influence of such climate modes as forcing factors on our ocean meteoric water concentrations. However, to the extent that the precipitation anomalies observed here are known to be dependent on such coupled modes and have been shown to be causal in explaining our interannual changes in meteoric water, this connects our local-scale observations of change to very large scale (circumpolar and hemispheric) forcing. Further, SAM is known to be sensitive to anthropogenic influences, via greenhouse gas emissions and ozone depletion (Marshall, 2003;, creating a possible link whereby human-induced changes could influence the meteoric water component of the WAP coastal freshwater budget. However, our new results demonstrate notable complexity in the freshwater responses across a range of scales, and emphasize the need to carefully understand such responses if attribution is to be reliable and such predictive skill is to be enhanced.  Dee et al. (2011) and at https://www.ecmwf.int/en/ forecasts/datasets/reanalysis-datasets/era-interim.