Spaceborne infrared imagery for early detection and cause of Weddell Polynya openings

Abstract. When will sea ice open is a crucial information for navigation and scientific deployments. This became painfully obvious when the Weddell Polynya, a large hole in the winter Southern Ocean sea ice, unexpectedly re-opened in 2016 for the first time in forty years. With no early warning, observations were limited to chance autonomous sensors, so the much-debated cause of the opening still cannot be determined accurately. We aim here to create such an early warning system. From the full historical sea ice concentration record, we find in fact 30 polynyas since 1980. Then, using the full time series of the 5 spaceborne infrared Advanced Very High Resolution Radiometer, we determine that these events can be detected in the two weeks before the polynya opens. Area-average median brightness temperature larger than 253 K in all three bands and areamaximum larger than 269 K along with a footprint at least larger than 4000 km successfully forecasts the polynyas and does not return any false positive. Or rather, it returned false positives that were in fact events that the sea ice concentration threshold had missed. Moreover, we find temporal oscillations in brightness temperature that could indicate upwelling of warm water, 10 but also changes of sign in T45 (band 4 – band 5) which could indicate a lead. We hence combine the spaceborne infrared data with atmospheric reanalysis, hydrographic mooring data and Sentinel-1 radar imagery and find that all events, including the 2017 Weddell Polynya, are caused by both atmospheric divergence and oceanic upwelling. That is, the debate is closed: both parties are correct; the Weddell Polynya is a hybrid.


Data
In this study, we first determine the dates of past polynya events since 1980 using spaceborne microwave-based sea ice products as reference. We then study these events using spaceborne infrared data, validated against in-situ hydrographic data and atmospheric reanalyses, but also SAR imagery. Our region of interest (red contours on Fig. 1), hereafter referred to as "the polynya-prone region", lies over the topographic feature Maud Rise, in the eastern Weddell Sea sector of the Southern Ocean 70 (longitude 6 • W to 12 • E; latitude 68 • S to 60 • S).
For reference, we use daily sea ice concentration at 25 km resolution from the National Snow and Ice Data Center, available continuously since 1978 (Cavalieri et al., 1996). We also briefly use in section 3.2 the higher resolution sea ice concentration product from the University of Bremen (Spreen et al., 2008), available since 2002.
The spaceborne infrared data come from the Advanced Very High Resolution Radiometer (AVHRR) Polar Pathfinder or 75 APP, provided by the National Oceanographic and Atmospheric Administation (Key et al., 2019). It provides twice-daily, 5 km gridded composites of all available AVHRR brightness temperature data since 1982. We use only the ones acquired at 2 AM to ensure that we work year-round with dark images. The three bands that we use are commonly referred to as T3b (wavelength of 3740 nm), T4 (10800 nm) and T5 (12000 nm).
For validation, we use the hourly 2 m temperature and 10 m horizontal wind components u and v from the European 80 Centre for Medium-Range Weather Forecasts ERA5 hourly reanalysis, provided on a 0.25 • grid (doi: 10.24381/cds.adbb2d47).
Hydrographic data come from three moorings deployed since 1996 along the Prime Meridian by the Alfred Wegener Institute, named AWI229 (Fahrbach and Rohardt, 2012a;Rohardt and Boebel, 2019), AWI 230 (Fahrbach and Rohardt, 2012b, c) and AWI 231 (Fahrbach and Rohardt, 2012d). Finally, we also use the backscatter information from Sentinel-1 SAR images in the extra-wide swath mode (approx. 40 m resolution) provided by the European Space Agency / Copernicus. We use only the HH 85 polarisation, as we do not perform a detailed analysis but rather a qualitative assessment of sea ice conditions.

Cloud masking of APP data
Clouds are a known issue for AVHRR data, especially in polar regions (e.g. Drinkwater, 1998). The first cloud filters adapted to the polar regions were designed by Yamanouchi et al. (1987), which imposed criteria on T4, T34 (T3b minus T4) and T45 (T4 minus T5) to detect thick, high and thin clouds, respectively. Saunders and Kriebel (1988) added a geographical/texture 90 perspective, imposing criteria on 3 by 3 pixel areas, while Key and Barry (1989) added a temporal perspective, comparing each pixel from day to day. But these filters did not perform as well as expected, and we had to wait until Vincent et al. (2008) for extra criteria on T45 that can detect ice fog.  Yamanouchi et al. (1987, indigo and green) and Vincent et al. (2008, yellow).
Insert indicates the location of the two images. Red contours on panel a, the location of the so-called polynya-prone region that we use in this study.
We aim to create a system that can work on individual images independently, so the approach of Key and Barry (1989) is not adapted. Likewise, Saunders and Kriebel (1988) is mostly based on day-time images, so not for us. We hence use the following 95 three criteria to detect clouds ( Fig. 1): 4 https://doi.org/10.5194/tc-2020-123 Preprint. Discussion started: 7 May 2020 c Author(s) 2020. CC BY 4.0 License.
-T4 < 245 K (indigo, Yamanouchi et al., 1987); or |T34| > 2 K (green, Yamanouchi et al., 1987); or |T45| > 2 K (yellow, Vincent et al., 2008) As shown on Fig. 1, these criteria are not perfect, but visual verification (not shown) proves that they are powerful enough 100 to detect most of the clouds. Moreover, leads and polynyas generate a large heat and moisture flux (e.g. Cheon et al., 2014), which we do not want to see masked as a cloud or else we will not detect the polynyas. Any pixel that meets any of the three criteria listed above is set to NaN in all three bands T3b, T4 and T5 for our calculations.

Methods
We first detect polynyas by applying sea ice concentration thresholds on each pixel, or on the area-average. We show only three 105 different thresholds based on the literature in section 3.1 (Gloersen et al., 1992;Gordon et al., 2007;Campbell et al., 2019), but we tested and visually assessed each option between 15 and 100%. We calculate the area of the polynya by simply multiplying the number of contiguous pixels with sea ice concentration lower than the specific thresholds by the pixel area. An event is defined as an uninterrupted series of consecutive days with sea ice under that threshold (60% for most of the study); if there is a day with higher sea ice concentration in between, a new event is created.

110
The aim of our work is to eventually produce an automatic system that would be scanning the polynya-prone region, so we do not track individual polynyas and instead analyse the brightness temperature time series over the whole region in section 3.2. For robustness though, we did produce brightness temperature composites centered on each polynya, which yielded similar results (not shown). For each day from 1982 to 2018, we produced time series of the geographical median, standard deviation, minimum and maximum brightness temperature over the polynya-prone region for each band, but also anomalies of these 115 brightness temperatures relative to daily climatological values over the 40 years. Note that we produced a separate climatology using only the years with no polynya, but found no significant difference (not shown).
Finally, the atmospheric and hydrographic data used in section 3.3 are directly studied without further processing, except for the wind components. We produced over each polynya a time series of so-called curl of the wind. We cannot compute the wind stress curl per se, as we lack the drag coefficient, which will anyway change depending on whether the polynya is open 120 or close. So instead, we use a similar method as e.g. Petty et al. (2016) and work with the curl of the wind components u and v:

Number of days matching SIC criterion
Year Figure 2. For each year of the NSIDC sea ice time series , number of days between 1st July and 31st October with a polynya according to the three most common criteria in the literature: black, minimum sea ice concentration (SIC) lower than 15% (e.g. Gloersen et al., 1992); blue, minimum sea ice concentration lower than 60% ; red, average sea ice concentration over the polynya-prone region lower than 92% (after Gordon et al., 2007).
We start by determining the dates of the polynya events that we want to further study, using the traditional criterion sea ice minimum over the region lower than 15% (black, Fig. 2) or lower than 60% (blue), and that of average sea ice concentration (SIC) lower than 92% (red). We limit ourselves to the period 1st July to 31st October. As noted by Campbell et al. (2019) 130 already, these methods return qualitatively similar results; it is only the number of days with each criterion that differ. They all agree there was some polynya activity in the late 1980s -early 1990s and early 2000s, the so-called Maud Rise halo already studied notably by de Steur et al. (2007). Then, the region was rather quiet until the widely reported 2016-2017 return of the polynya (Swart et al., 2018). for their criterion too. We then visually checked the individual images to separate such late freeze-up / early melt from the actual polynya events. This visual validation also showed that the 60% criterion performed best (not shown).
The characteristics of the events thus detected are given in appendix table A1. We have 24 events over 11 years, which yield

Infrared-based early detection criteria
In the previous section, we determined the dates of 24 polynya events (giving 30 actual polynyas) from sea ice data dating back to 1978. We now investigate, in the timeseries of brightness temperature from APP, whether all these events share something in common, especially in the two weeks leading up to the event. We present the 30 days prior to the events in section 3.3 but 155 we found that for the current purpose, 15 days are enough. As explained in the Methods section, this "something in common" needs to be easily detectable by a crude automatised system, hence we computed basic single-image properties over the entire polynya-prone region ( We want to find a criterion that would not only robustly detect a polynya, but also not flag any false positive. Fig. 4 shows 160 that such criterion will have to be band-dependent. For T3b (cyan diamonds) and T5 (red stars) for example, the geographical minimum is of no use owing to its large spread across the events, whereas it is very specific for T4 (black circles). For T4 and T5, it is the standard deviation that is of no use: the 15-day minimum is always at 0, regardless of the event, and the maximum can also be 0. The geographical median and maximum look most promising, so we verify how many false positives would be returned by using their smallest values as a threshold (i.e. for the median for T5, either 243.1 or 252.6 K, see Fig. 4). There are 165 3443 days with infrared data that are between 1st July and 31st October in the years with no polynya event. For the individual bands, only the temporal maximum of the geographical median and the temporal maximum of the geographical maximum, in absolute values (left, top and bottom, y-axis, Fig. 4), that wrongly detect fewer than 1000 days. That is not good enough.
What Fig. 4 reveals is that we need sacrifice only one or two polynya events to be much more restrictive. More specifically, few polynya points have a geographical median lower than 253 K, and a geographical maximum that does not exceed 269 criteria, we obtain only 36 false positives. That is a remarkable improvement. We can also create an extra criterion based on the size of the signal detected: in our 24 polynya events, across the three bands, they all had on average more than 4000 km 2 of pixels affected. Note that we are not talking about the size of the polynya here, only the footprint on the brightness temperature data. If we add this final size restriction, we are left with 12 false positives (  and maximum (bottom). One point per polynya event: cyan diamonds for the temperature band T3b; black circle for T4; red star for T5.
Shown both the actual brightness temperature value (left) and the daily anomaly (right), both in degree K.
9 https://doi.org/10.5194/tc-2020-123 Preprint. Discussion started: 7 May 2020 c Author(s) 2020. CC BY 4.0 License. After visual examination, 4 of these false positives were late freeze-up. They all occurred in the first half of July, so a more restrictive winter date definition could solve this, potentially involving sea ice thresholds as in Campbell et al. (2019). Of the remaining 8, the majority had the typical halo appearance, with sea ice concentration between 60 and 90%. Some had even been mentioned in past publications (e.g. Smedsrud, 2005;Campbell et al., 2019), albeit with large uncertainties. The most obvious results were for the two latest dates, in 2010 and 2011, where we could verify the sea ice concentration at a much 180 higher resolution in the University of Bremen datasets (Spreen et al., 2008). There, leads and small polynyas were directly apparent, proving the necessity of high resolution data for climate research. In summary, the criteria on the infrared brightness temperature did not return any false positive, or at least any that cannot easily be removed by simply checking the date. If anything, it successfully detected 8 extra events that the sea ice concentration criterion had missed.
We have completed our objective and determined brightness temperature thresholds to detect a polynya before it opens.

185
Instead of stopping here, can we extract more information out of the infrared data? For example, an answer to the much debated "why did the polynya open"? One option is the difference between bands T4 (most adapted to ice) and T5 (most adapted to open ocean) or T45, where T45 < 0 means lead (Vincent et al., 2008). Unfortunately, out of the 24 events, all of them ( Fig. 5a and supplementary Fig. A1) had a minimum T45 lower than 0 in the 14 days before the event. Which would mean that all of them had a lead somewhere in the polynya-prone region. The other option is that, as pointed out by Heuzé and 190 Aldenhoff (2018), oscillations in the brightness temperature in the days before the polynya might reflect oceanic convective movements. The argument is that as the warm water is being upwelled, more heat is going through the ice; same as the ice thins, melted from below by that warm water. But likewise, all polynya events exhibit such oscillations, albeit with a frequency https://doi.org/10.5194/tc-2020-123 Preprint. Discussion started: 7 May 2020 c Author(s) 2020. CC BY 4.0 License.
varying from 1 to 3 days depending on the band and event (Fig. 5b and c and supplementary Figs A2 and A3). In conclusion, we cannot determine the cause of the polynya opening from the infrared images alone. In the next section, we investigate this 195 further, using in-situ validation data. Days before polynya Days before polynya Days before polynya From the infrared images, both mechanisms are possible: the oscillations in brightness temperature suggest an upwelling of 205 warm water, but simultaneously a negative T45 suggests that leads open. We hence now augment our spaceborne infrared data with atmospheric and hydrographic data for clarification.

b) All bands, median T c) All bands, maximum T a) Minimum T45
The data coverage for the atmosphere is much better than for the hydrography, so we start with the atmosphere. The curl of the wind at the location of the individual 30 polynyas (listed in Table A1) alternates between negative and positive values for all polynyas (supp Fig. A4, negative curl / upwelling is blue, positive curl / divergence is red). That is, all polynyas have been 210 preceded by both upwelling-prone and divergent winds. Note that the required timing is unclear: for divergence, e.g. Out of the 30 polynyas, 15 reach either of these freezing temperatures (supplementary Fig. A5). The main caveat here though is that the air temperatures are not given at the surface of the ice, but at 2 m height where it probably is colder. Moreover, these are reanalysis data, not actual observations at the polynya location, and hence are somewhat uncertain. All that we can say then is that three events remain colder than or around -10 • C, but for all the others it might have been warm enough to melt. Our 220 conclusion for now is that the atmospheric reanalysis data did not provide a satisfactory discrimination.

Latitude (°S)
Depth ( We hence finish this paper with a more in-depth analysis of the three individual polynya events for which we have mooring data in the vicinity. Five mooring deployments coincide with a polynya event: one mooring (230-2) for 1999, three (229-5, 230-4, 231-5) for 2004, and one (229-13) for 2017. We want to determine if 1) we see in the mooring data upwelling of the comparatively warm and salty Circumpolar Deep Water (CDW), and 2) whether such upwelling is in sync with the brightness 225 temperature oscillations. Unfortunately, for four out of five deployments, we are limited in the top 500 m to one sensor at the surface and all the others in the CDW (Fig. 6). We lack sensors in the top 100 m. Moreover, the surface sensor often had only temperature data, not salinity, and in the case of 229-13 for the 2017 polynya, there was no surface sensor data available; luckily, there were Sentinel-1 images. All we can study are variations in the CDW, both the suspected heaving of its core (Dufour, 2017) or its cooling as convection begins (Gordon, 1978). As sensors at the same depth returned different values but 230 similar variability (not shown), we will not comment on the mooring values per se but only on their changes.
In 1999 (Fig. 7), the most notable variation is the dip in all infrared bands (panel a) and in T45 (panel b) 18 days before the polynya, mere hours after a warming and salinification of both deep sensors (panels e and f). The surface sensor (panel d) unfortunately shows variations that are of the level of the sensitivity of the instrument, and hence is not very useful. Then for two days, temperature and salinity at 480 m depth oscillate with a near 12 hour frequency, only to stop 13 days before the 235 polynya, at the onset of the continuous increase in all infrared bands until 6 days before the polynya. Over that same period, the CDW depths cool down while the surface might be warming. All this seems to have started after a few hours of negative curl of the wind (panel c), suggesting that we might be witnessing an upwelling event from 13 to 7 days before the polynya. Finally, 4 days before the polynya, we see oscillations in all infrared bands and a strong negative T45. This immediately follows a sudden warming and salinification at 380 m depth (and to a lesser extent 480 m) before an intense cooling. In conclusion, in 240 1999, it seems that we first had upwelling 13 days before the polynya, convection starting 8 days before, and maybe leads 4 days before.
The 2004 event (Fig. 8) starts with a surprising decrease by 2 • C of the surface temperature 24 days before the event (panel d). All the other depths are rather stable until 18 days before the polynya, where they start having high frequency variability, especially at 170 and 220 m (panels e and f, respectively), correlated between salinity and temperature (warmer and 245 saltier alternating with cooler and fresher). At the same time, the infrared brightness temperature in all bands oscillates while increasing to 260 K as average (panel a). Then starting 15 days before the polynya, the curl of the wind (panel c) becomes and stays negative for three days while the CDW oscillations continue and reach the deeper levels; to us, this obviously suggests upwelling. Following the lowest T45 of this series (-0.6 K, panel b) and divergent winds 11 days before the polynya, both the temperature at depth and the brightness temperatures increase again but without a change in salinity, which might indicate 250 a lead. The final period in the two days before the polynya is more convincing: strong divergence in the wind, brightness temperature increases, but all depths see a cooling. We suspect that this indicates that a lead opened following the divergence, that the ocean is losing heat to the atmosphere, and that the infrared data detects that heat (and moisture) exchange to the atmosphere. Note that we showed only the mooring that had the most data, but all three moorings agree. For this event as well then, there might have been both leads and upwelling. Luckily for the last event that we will look at, radar images are available. Finally, the 2017 event (Fig. 9) shows the same patterns again: an increase in variability along with a low T45 and divergent winds 15-14 days before the polynya (lead?), but also a synchronous increase in infrared brightness temperature and temperature and salinity at depth (upwelling?). But this time, we have high resolution Sentinel-1 SAR images to verify our hypothesis.
What we see first is a sudden warming and salinification at depth from 19 to 16 days before the polynya (panels d and e).
Straightafter, we again observe a decrease in T and S along with high frequency variability, just before the usual T45 lead 260 signature (panel b); the ice is closed then on the SAR image. Five days later, or 11 days before the event, there are leads all around the area on the SAR image. One can imagine that the surface cooling resulting from the lead opening may have been strong enough to initiate convection, which would explain the apparent warming at 310 and 410 m. We see that three days later, i.e. 8 days before the polynya, the leads have refrozen, so it is more likely is that the convection started as the ice formed again and ejected brine. The convection would bring down the cooled water from the surface to 310, 410 m depth, hence their cooling 265 (as observed in reality and in models, Gordon, 1978;Smith Jr et al., 2007;Martin et al., 2013;Cheon and Gordon, 2019). And although we lack images between 8 and 4 days before the polynya, the fact that new freshly-refrozen leads appeared closer to the centre of the polynya area suggest that the ice opened again. In conclusion, the ice was very dynamic in 2017, with multiple openings and closing of leads potentially initiating the oceanic convection.    (Francis et al., 2019;Campbell et al., 2019) arguments. All three events that we could study in details show both upwelling and wind-induced leads. That is, the Maud Rise polynya presents the characteristics of both a sensible and latent heat polynya. Such a phenomenon has in fact been observed recently in a polynya off Alaska, and dubbed "hybrid polynya" (Hirano et al., 2016). Maybe it is time to admit that there has been no consensus in forty years because the Weddell Polynya is a hybrid polynya too.

Conclusions
The aim of this paper was to determine criteria on spaceborne infrared imagery (AVHRR / APP) to detect an upcoming reopening of the Weddell Polynya. Using the NSIDC sea ice concentration, we first generated a time series of past polynya events over Maud Rise and obtained 24 events since 1980, or 30 polynyas as some days had several polynyas opened simultaneously (Fig 2). The widely accepted narrative is that there had been no polynya in the Weddell Sea since "the" Weddell Polynya of 280 1974-1976 when the polynya unexpectedly re-opened in 2016 (e.g. Swart et al., 2018). Yet, our study is but one of many that found once again that there has in fact been many polynyas in the region in the forty years in between (e.g. Lindsay et al., 2004;Smedsrud, 2005;de Steur et al., 2007;Campbell et al., 2019).
Investigating the weeks leading to each of these 24 events, we found four criteria based on the area-average and areamaximum daily infrared brightness temperature in all bands that, when combined, successfully detected the events without 285 finding false positives. Or rather, we thought it had returned 12 false positives, but a closer inspection (Table 1)  Finally, we investigated whether spaceborne infrared data could be used to answer the forty-year old debate: is it the ocean or the atmosphere that causes the Weddell Polynya to open? Using the infrared data combined with atmospheric reanalysis, mooring hydrographic time series and even Sentinel-1 SAR imagery, we found evidence of both (Figs. 7-9). We found obvious signatures of warm water upwelling and deep convection (e.g. Holland, 2001;Martin et al., 2013;Dufour, 2017;Cheon and 295 Gordon, 2019) but also of wind-driven lead opening (e.g. Gordon et al., 2007;Cheon et al., 2014;Campbell et al., 2019;Francis et al., 2019). Our results suggest that there should not be any debate in fact: both parties are correct, and the Weddell Polynya probably is a hybrid one, as happens for some Arctic polynyas (Hirano et al., 2016).
This study also proved the crucial role of high resolution products. SAR imagery is widely used for lead detection in the Arctic (Murashkin et al., 2018), but as Sentinel satellites pass over the same spot only every 4 days at best, they are of limited 300 use for operational purposes. We need more high resolution instruments in space, onboard more satellites. Moreover, this study would not have been possible without continuous observation programs, both in space (AVHRR) and at sea (moorings). These must be maintained, as long time series are crucial not just for statistical exercises like here, but also to monitor climate change in the fast-changing polar regions (Stocker et al., 2014).     (229-5, in 2002-2005, and 230-2, in 1999-2000) and cyan, the more recent one (229-13, in 2016-2019, and 230-4, 2002-2005); for mooring 231, only deployment 231-5 (2002-2005) could be used here.