Variation of Summer Oceanic p CO 2 and Carbon Sink in the Prydz Bay Using SOM Analysis Approach

This study applies a neural network technique to produce maps of oceanic surface pCO2 in the Prydz Bay in the Southern Ocean on a 0.1 longitude  0.1 latitude grid based on in-situ measurements during the 31 CHINARE cruise for February 2015. The study area was divided into three regions, Open-ocean region, Sea-ice region and Shelf region. The distribution of oceanic pCO2 was mainly affected by physical process in the Open-ocean region where mixing and upwelling became the main controls. While in the Sea-ice region, oceanic pCO2 changed sharply due to the strong change of seasonal ice. For the Shelf region, biological factor was the main control. The weekly oceanic pCO2 was estimated using a self-organizing map (SOM) by four proxy parameters (Sea Surface Temperature, Chlorophyll a concentration, Mixed Layer Depth, and Sea Surface Salinity) to resolve the nonlinear relationships under complicated biogeochemical conditions in Prydz Bay region. The reconstructed oceanic pCO2 coincides well with the in-situ investigated pCO2 from SOCAT, in the root-mean-square error of 22.14 μatm. Prydz Bay was mainly a strong CO2 sink in February 2015 with a monthly averaged uptake of 18.74.93 TgC. The oceanic CO2 sink is pronounced in the Shelf region due to its lowest oceanic pCO2 with peak biological production. Strong potential anthropogenic CO2 uptake in the Shelf region will enhance the acidification in the deep water of Prydz Bay and affect the deep ocean acidification in the long run since it contributes to the formation of Antarctic bottom water.


Introduction
The role of the ocean south of 60°S in the transport of CO2 to or from the atmosphere is still uncertain despite of its importance of reducing anthropogenic CO2 in the atmosphere (Sweeney et al., 2000(Sweeney et al., , 2002;;Morrison et al., 2001;Sabine et al., 2004;Metzl et al., 2006;Takahashi et al., 2012).This status derives from both the strong seasonal and spatial variations that occur around Antarctica, and the difficulty of field measurements in the region for its hostile weather and remoteness.
Following the Weddell and Ross seas, Prydz Bay, lying in the Indian Ocean section, is the third largest embayment in the Antarctic continent.With Cape Darnley to the west and the Zhongshan Station and Davis Station to the east, Prydz Bay is close to the Amery Ice Shelf to the southwest and the West Ice Shelf to the northeast (Fig. 1).Water depth varies sharply northward from 200 m to 3000 m.The inner continental shelf is dominated by the Amery Depression, which is mostly 600 to 700 m deep.The depression is bordered by two shallow banks (<200 m): Fram Bank and Four Ladies Bank, forming a spatial barrier to water exchange with the outer oceanic water (Smith and Trégure, 1994).The Antarctic Coastal Current (CoC) flows westward, bringing in cold waters from the east.When it reaches the shallow Fram Bank, it turns north and then part of it flows westward, while part of it turns eastward back to the inner shelf resulting in the clockwise rotating Prydz Gyre (see Fig. 1).The circulation to the north of the bay is characterized by a large cyclonic gyre, extending from within the bay to the Antarctic Divergence at about 63°S (Nunes Vaz and Lennon, 1996;Middleton and Humphries, 1989;Smith et al., 1984;Roden et al., 2013;Wu et al., 2017).The inflow of this large gyre hugs the eastern rim of the bay, and favors the onshore intrusions of warmer modified Circumpolar Deep Water (mCDW) across the continental shelf break (Heil et al., 1996).A westward flow along the shelf, that is part of the wind-driven Antarctic Slope Current (ASC), supplies water to Prydz Bay.In the austral summer, with longer daylight and increased solar radiation, sea surface temperature increases, ice shelf breaks and sea ice melts, resulting in stratification of the water column.Prydz Bay region is host to a marine ecosystem that interacts with the physical environment which makes it complicated to study the spatiotemporal variability and mechanism of oceanic pCO2.
Despite the importance of carbon cycle in the Southern Ocean, the observations are rather limited to analyze the spatiotemporal variation in the Prydz Bay.The analysis of temporal variability and the spatial distribution mechanism of oceanic pCO2 in Prydz Bay was limited to cruises or stations (Gibsonab and Trullb, 1999;Gao et al., 2008;Roden et al., 2013).To estimate regional sea-air CO2 fluxes, it is necessary to interpolate between in-situ measurements to obtain the maps of oceanic pCO2.Such an interpolation approach, however, is still a difficult task because observations are too sparse in both time and space to capture the high pCO2 variability.
Satellites do not measure pCO2 but they give access to parameters related to the processes that control its variability.The seasonal and geographical variability of surface water pCO2 is indeed much greater than that of atmospheric pCO2.Therefore, the direction of the sea-air CO2 transfer is mainly regulated by the oceanic pCO2.Linear regression extrapolation method has been applied to expand the cruise data to a big scale area to study the carbon cycle in the Southern Ocean (Rangama et al., 2005;Chen et al., 2011;Xu et al., 2016), however, the linear regression relied simply either on chlorophyll-a (CHL) or on sea surface temperature (SST) parameter.
Thus, this method is insufficient to represent all the controlling factors.In this study, we applied self-organizing map (SOM) analysis to expand our observed data sets and estimated the oceanic pCO2 in the Prydz Bay during the February 2015.
The SOM analysis, based on neural network (NN), a type of artificial neural network, has been proved to be a useful method for extracting and classifying features in geoscience (Gibson et al., 2017;Huang et al., 2017b).In oceanography, SOM has been applied for the analysis of various properties of the seawater such as sea surface temperature (Iskandar, 2010;Liu et al., 2006), chlorophyll concentration (Huang et al., 2017a;Silulwane et al., 2001).In the past decade, SOM has also been applied to produce basin-scale pCO2 maps mainly in the North  Laruelle et al., 2017).SOM has been proved to be useful to expand a spatial-temporal coverage of direct measurements or to estimate properties whose satellite observations are technically limited.One of the main benefits of the neural network method over the more traditional techniques is that there is more accurate representation of the highly variable system of interconnected water properties (Nakaoka et al., 2013).
During the 31 th CHINARE in Prydz Bay, we have conducted a survey on partial pressure of CO2 in oceanic water and atmosphere from the beginning of February to the early of March (data of the cruise track is shown in Fig. 2).This study is aimed to apply the SOM method to reconstruct the temporal and spatial variability of oceanic pCO2 distribution in Prydz Bay from 63°E to 83°E, 64°S to 70°S and discuss the capability of carbon absorption in February 2015.
The paper is organized as followed.Section 2 provides the descriptions of the in-situ measurements and the SOM methods.Section 3 presents the analysis and discussion of the results.Section 4 presents the summary.

in situ data
The in situ underway pCO2 of marine water and atmosphere was collected during the 31 th CHINARE when R/V Xuelong sailed from east to west at the beginning of February 2015 (see  2-a, b).Sea water at 5 meters underneath the sea surface was pumped continuously to the GO system (GO Flowing pCO2 system, General Oceanics Inc., Miami FL, USA), and the partial pressure of the sea surface water is measured by an infrared analyzer (LICOR, USA, Model 7000).The analyzer was calibrated every 2.5-3 h using four standard gases at pressures of 88.82 ppm，188.36ppm，399.47 ppm，528.92ppm supplied by NOAA's Global Monitoring Division.The accuracy of the measured pCO2 is within 2 μatm (Pierrot et al., 2009).The underway pCO2 in atmosphere was simultaneously collected by the GO system.
Sea ice melt has a significant impact on the local stratification and circulation in polar region.Salinity records the physical processes.During freezing, salt is excluded from ice, and thus increase the ocean surface salinity.This is so called brine rejection.When ice begins to melt, fresher water is added into the ocean to dilute the ocean water, i.e., reducing the salinity.In this study, we treat salinity as an index for the change of sea ice.The underway sea surface temperature SST and conductivity was recorded by a Conductivity-Temperature-Depth sensor (CTD, Seabird SBE 21) along the cruise track.Later sea surface salinity was calculated according to the recorded conductivity and temperature.The distributions of underway SST and SSS were shown in Fig. 2 c and d.
In previous studies it has been reported that the summer sink in Prydz Bay is clearly biologically driven and the pCO2 change is often well-correlated with surface chlorophyll-a concentration (Rubin et al., 1998;Gibsonab et al., 1999).When sea ice starts to melt, the active biological process affects oceanic pCO2 significantly (Chen et al., 2011;Xu et al., 2016).The chlorophyll-a value is regarded as an important controlling factor of pCO2.Daily Modis chlorophyll-a data of 4 km resolution (http://oceancolor.gsfc.nasa.gov)are interpolated to the observation section and time.The interpolated result along the cruise track is shown in Fig. 2e.
The ocean mixed layer is characterized as having nearly uniform physical properties throughout the layer with a gradient in properties at the bottom of the layer.The mixed layer links the atmosphere to the deep ocean and plays a critical role in climate variability.Very few studies have emphasized the importance of accounting for the vertical mixing through the mixed layer depth (MLD, Dandonneau, 1995;Lüger et al., 2004).The stability and stratification prevent the upward mixing of nutrients and limits the biological production and thus affect sea-

SOM method and input variables
We hypothesize that oceanic pCO2 can be reconstructed through the SOM based multiple non-linear regression with four proxy parameters (Eq.1): sea surface temperature (SST), the abundance of photo-synthesizing organisms in the surface ocean represented by the chlorophyll-a concentration (CHL), mixed layer depth (MLD), and sea surface salinity (SSS).
The SOM is trained using unsupervised learning to project the input space of training samples to a feature space (Kohonen, 1984), which is usually represented by grid points in twodimension space.Each grid point, also called neuron cell, is associated with a weight vector having the same number of components as the vector of input data (Zeng et al., 2017).During the SOM analysis three steps are taken to estimate oceanic pCO2 fields (see Fig.During the second process, each preconditioned SOM neurons is labeled with an observed oceanic pCO2 value.The labeling dataset consisting of the observed pCO2 and the normalized SST, CHL, MLD and SSS is presented to the neural network and then a winner neuron is found.
After the labeling process, neurons are represented by five-dimensional vectors.
Finally, during the mapping process, the labeled SOM neurons created by the second process and trained SOM neurons created by the first process are used to produce oceanic pCO2 of the winner neuron according to the geographical grid points of the study area.
All the daily datasets were first averaged to be 8-d fields regarded as weekly for this study.We compared the assimilated datasets of SST from AVHRR with in situ measurements obtained by CTD along the cruise.The relationship is 0.97 and the root-mean-square error (RMSE) is 0.2°C.The SSS and MLD fields from the Global Forecast system compare reasonably well with the in situ measurements, with relationships of 0.76 and 0.74, respectively and the RMSE of 0.41 and 5.15m.The uncertainty of the Modis CHL data in the Southern Ocean is about 35% (Xu et al., 2016).For the labeling procedure, the observed oceanic pCO2 together with corresponding in situ SST, SSS, MLD, and Modis CHL product in vector form are used as the input dataset.

Validation of SOM derived oceanic pCO2
To validate the oceanic pCO2 reconstructed by the SOM analysis, we used the fugacity of oceanic CO2 datasets (referred as "SOCAT" data hereinafter) from the Surface Ocean CO2 Atlas (SOCAT: http://www.socat.info)version 5 database (Bakker et al., 2016).In Pacific Ocean, the Atlantic Ocean or regions away from coast, datasets from different years can be assimilated to a reference year to have a good spatial coverage according to the equilibrium between sea surface and atmosphere (Takahashi et al., 2006;Wong et al., 2010;Nakaoka et al., 2013).However, the same approach should be applied carefully because the sea ice condition varies from year to year in the Southern Ocean.The sea ice cover has a great impact on the oceanic pCO2.SOCAT data in February from different years do have a good spatial coverage in Prydz Bay.However we could only select dataset for our study period in 2015 (see Fig. 4-a) although it covers limited area in study region.We recalculated pCO2 values from the obtained fCO2 offered in SOCAT data according to the fugacity correction (Pfeil et al., 2013).

Carbon uptake in the Prydz Bay
The flux of CO2 between the atmosphere and the ocean was determined by two items.
One is the difference in CO2 concentration across the sea-air interface and the other is the transfer velocity which is a function primarily of wind speed and temperature.The equation to calculate the sea-air carbon flux was simplified as a function of wind speed and delta pCO2 (from sea to air) in eq. 2, Xu et al. (2016).For the weekly estimation in this study, the scaling factor for the gas transfer rate is changed to 0.251 for a shorter time scale and at intermediate wind speed ranges (Wanninkhof, 2014).For each grid, weekly sea-air carbon flux in the Prydz Bay can be estimated by Eq. ( 2): where U represents wind speed 10 m above sea level, pCO2_sea and pCO2_air are partial pressure of CO2 in sea water and atmosphere.
We downloaded weekly ASCAT wind speed data (http://www.remss.com/,see Fig. S6) of 1/4 degree and then regridded the dataset to fit the 0.1 longitude  0.1 latitude spatial resolution of SOM derived oceanic pCO2.We regridded the atmospheric pCO2 collected along the cruise track to fit the spatial resolution of SOM derived oceanic pCO2 by linear method.The total carbon uptake was then obtained by accumulating the flux of each grid by each area according to

the distributions of underway measurements
From the beginning of February, R/V Xuelong sailed from east to west along the sea ice edge.Based on the water depth and the sea ice condition, the study area is robustly divided into three sectors, the Open-ocean region, Sea-ice region and the Shelf region.
The Open-ocean region was from 66°S northward to 64°S where locates the Antarctic Divergence Zone and with water depth greater than 3000 m.The AD zone was characterized by The Shelf region (from 67.25°S southward) is characterized of low depth below 700m, surrounding by the Amery Ice Shelf, West Ice Shelf, and the stretching permanent sea ice from the West Ice Shelf, formed by modification of low temperature and salinity shelf water (Smith et al., 1984).Two shallow banks (<200m): Fram Bank to the north-west and Four Ladies Bank to the north-east, forming a spatial barrier for the inner shelf to water exchange with the outer oceanic water (Smith and Tréguer, 1994).Prydz Bay coastal current flowed from east to west in the semi-close bay.There is always a fresher, warmer surface layer over the bay which is known as the Antarctic Surface Water (ASW).During our study period, the Shelf region was completely ice free, a large volume of freshwater was released into the bay, resulting in low sea surface temperature (an average of -0.61°C) and salinity (an average is 32.4).As shown in Fig. 2-f, the mixed layer depth is low in most of the inner shelf.Due to the vast shrink of sea ice and strong stratification in the upper water, algal bloomed and chlorophyll value was high with an average of 1.93 mg/m 3 .The oceanic pCO2 in this region turned out to be the lowest in three sectors.The average of oceanic pCO2 is 198.72 μatm with a range from 151.70 μatm to 277.78 μatm.
Chlorophyll-a value shows remarkably as high as 11.04 mg/m 3 from 72.3°E, 67.3°S to 72.7°E, 68°S when sea ice retreated eastwardly.In the bay mouth close to the Fram Bank, due to the local upwelling water salinity increased remarkably to around 33.2.Biological pump becomes the dominant factor of the distribution of oceanic pCO2.

Quality and maps of SOM derived oceanic pCO2
We selected SOM derived oceanic pCO2 to fit the cruise track of SOCAT for a same period using a nearest grid method.The slope of the scatter plot showed that SOM derived oceanic pCO2 is lower than the SOCAT data (see Fig. 4-b).The RMSE between the SOCAT data and SOM derived result is calculated as follows: RMSE= √ ∑ ( pCO 2 sea ( SOM )−pCO 2 sea (SOCAT )) where n is the number of the validation dataset.The RMSE could be used interpreted as an estimation of the uncertainty in SOM derived oceanic pCO2 in Prydz Bay.In this study, the RMSE is 22.14 μatm.This is consistent with the accuracy (6.9 μatm to 24.9 μatm) achieved in ).However, this precision of this study is not as good as most of the neuron methods.
Increasing the spatial coverage of the labeling data will help to increase the precision of SOM derived oceanic pCO2.

Spatial and temporal distributions of SOM derived oceanic pCO2
In austral winter, the entire Prydz Bay basin is fully covered by sea ice except for a few In the Open-ocean region, sea ice started to melt in the beginning of February.In most area of the Open-ocean region it was sea ice free while the average sea ice coverage is only 18.14%.The ice cover is mainly associated with the outstretching permanent sea ice.Affected by the upwelling CDW, the stability of water was weak and not suitable for the growth of phytoplankton.It is also evidence by, the observed biological productivity, which was below 0.5 mg/m 3 .From the distribution of SOM derived oceanic pCO2 as shown in Fig. 6, oceanic pCO2 value was the highest compared to the Sea-ice region and the Shelf region.From week-1 to week-4, oceanic pCO2 increased a little and reached 381.42 μatm which was equivalent to that of atmosphere.In the western part of Open-ocean region, oceanic pCO2 decreased due to mixing with low oceanic pCO2 flew from the high latitude by the large gyre.Over the whole study period, the averaged ocean-air pCO2 difference (△pCO2) is largest in the Shelf region, then follows the Sea-ice region and Open-ocean region.△pCO2 from Shelf region changed from -184.31μatm to -141.00 μatm as the chlorophyll decreased.The sea-air difference of pCO2 in the Sea-ice region and Open-ocean region showed the same pattern.It increased from week1 to week3 then decreased in week4.Based on the △pCO2 and wind speed, the uptake of CO2 (eq.2) in three regions is presented in Fig. 7.In Shelf region the low oceanic pCO2 levels drove relatively intensive CO2 uptake from the atmosphere.Carbon uptake in Shelf region increased from week-1 (2.13 TgC) to week-2 (2.24 TgC) due to increased wind speed.

Carbon uptake in Prydz Bay
While in week-3, wind speed slowed down, resulting in the uptake of CO2 in Shelf region decreased to 1.70 TgC.In week-4, even the △pCO2 was the lowest, the total absorption still increased to be 2.03 TgC due to the high wind speed (averaged value of 7.9 m/s).17 The total carbon uptake in Prydz Bay of three regions integrated over the whole February 2015 was 18.7 TgC.The uncertainty depended on the errors for the wind speed, the scaling factor and the accuracy of SOM derived pCO2 according to Eq.2.The scaling factor will yield a 20% uncertainty to regional flux estimation.The errors in wind speeds of Ascat dataset is assumed to be 6% (Xu et al., 2016) and will be 12% in quadratic wind speed.For the SOM derived pCO2 the RMSE is 22.14 μatm.Considering the errors above and an uncertainty occurred when the sea-air computation expression was simplified (1.39%, Xu et al., 2016), the total uncertainty of the final uptake is 4.93 TgC.

Summary
According to different controls factors of ocean pCO2, the Prydz Bay region was divided into three sectors for February 2015.In Shelf region biological factor was the main control for oceanic pCO2 while in Open-ocean region mixing and upwelling became the main controls.In Sea-ice region, due to the rapid sea ice changing, oceanic pCO2 was controlled by both the biological and physical processes.SOM is an important tool to do the quantitative assessment of

Fig. 1 3 4
Fig. 1 The circulations in the Prydz Bay derived from Roden et al. (2013), Sun et al. (2013), Wu et al. (2017).ASC: Antarctic Slope Current; CoC: Antarctic Coastal Current; ACC: Antarctic Circumpolar Current.The weekly sea ice extents for our study periods were overlapped on the cruise.the pink line is for week- Atlantic and Pacific Ocean by using different proxy parameters (Lafevre et al., 2005; Friedrich & Oschlies, 2009a, 2009b; Nakaoka et al., 2013; Telszewski et al., 2009; Hales et al., 2012; Zeng et al., 2015; Fig.2-a, b).Sea water at 5 meters underneath the sea surface was pumped continuously to the GO air CO2 exchange.The vertical profile of sea water including potential density was measured by a Seabird SBE 11.Comparison of MLD based on the difference and gradient criteria (Brainerd and Gregg, 1995; Thomson and Fine, 2003) suggested that MLD determined using a difference criterion is more stable in the Southern Ocean.Following Dong et al. (2008), we calculated the mixed layer depth (see Fig.2-f) based the difference criteria, of with sigma theta changed by 0.03 kg/m 3 .The MLD value at stations were later gridded linearly to match the spatial and temporal resolution of the in situ data along the cruise track.

Fig. 2
Fig.2 the distributions of underway oceanic and atmospheric pCO2, SST, SSS, and CHL gridded from MODIS, and MLD gridded from stations surveys.

3
). Input variables to estimate pCO2 are prepared in a vector form.The SOM analysis was carried out by the MATLAB SOM tool box 2.0 (Vesanto, 2002).It has been developed by the Laboratory of Computer and Information Science in the Helsinki University of Technology and is available from the following web page: http://www.cis.hut.fi/projects/somtoolbox.

Fig. 3 . 7 13 14 Biogeosciences
Fig. 3. Schematic scheme of the main three step involved in the SOM neural network calculations leading to weekly pCO2 maps for February 2015.More realistic pCO2 estimates were expected from the SOM analysis when the distribution and variation range of the labeling variables closely reflect the training data sets(Nakaoka et al., 4) # the skewness of common logarithm of each variable is shown in the parenthesis.* [number of training data within the labeling data range]/[total number of training data] + the percent labeling data coverage of normalized variables is shown in the parenthesis During the training process, a neuron's weight vectors are repeatedly trained by being presented with the input vectors, until the neural network sufficiently represents the nonlinear interdependence of proxy parameters used in training.This process results in clustering of similar neurons and self-organization of the map.The observed oceanic pCO2 is not needed at the first step.Biogeosciences Discuss., https://doi.org/10.5194/bg-2018-276Manuscript under review for journal Biogeosciences Discussion started: 19 July 2018 c Author(s) 2018.CC BY 4.0 License.

22 Biogeosciences
high nutrients and low chlorophyll (HNLC) concentrations with high pCO2 attributed to upwelling of deep waters, suggesting the importance of physical processes(Burkill et al., 1995;Edwards et al., 2004).Underway sea surface temperature in this region are relatively high with an average value of 0.36°C due to the upwelling Circumpolar Deep Water (CDW) while in the sea ice edge (73°E, 65.5°S to 72°E, 65.8°S) SST decreased below -1°C.From 67.5°E westward, affected by the large gyre, cold water from the high latitude lowered down SST to below 0°C.Near the sea ice edge, SSS decreased quickly to 31.7 due to the diluted water, while along the 65°S cruise it reached to 33.3, and then western from 67.5°E affected by the fresher and colder water brought by the large gyre it decreased to 32.5.The satellite chlorophyll-a image showed that it was of low value of 0.45 mg/m 3 except when the vessel near the sea ice edge CHL increased to be 2.26 mg/m 3 .The distribution of MLD varied along the cruise.Near the sea ice edge, because of the melting of ice and direct solar warming, it constituted a low-density cap over the water column the MLD was as shallow as 10.21 m.The maximum value of MLD in Open-ocean region is 31.67 m.In the Open-ocean region atmospheric pCO2 was stable from 374.6 μatm to 387.8 μatm.Oceanic pCO2 varied from 291.98 μatm to 379.31 μatm with an average value of 341.48 μatm.Along the 65°E cruise in the east part of the Open-ocean region, oceanic pCO2 was relative high reaching an equilibrium with atmospheric pCO2.The lowest value was found near the sea ice edge due to biological consume.For the western part, oceanic pCO2 decreased a little due to the mixture of low pCO2 from higher latitude by the large gyre.Mixing and upwelling were the dominant factors for oceanic pCO2 in this region.The seasonal Sea-ice region (from 66°S to 67.25°S) is between the Open-ocean region and the Shelf region.In this sector, sea ice changed strongly and water depth varied sharply from 700 m to 2000 m.Sea ice kept changing and reforming from the late of February to the beginning of 11 21 Discuss., https://doi.org/10.5194/bg-2018-276Manuscript under review for journal Biogeosciences Discussion started: 19 July 2018 c Author(s) 2018.CC BY 4.0 License.March.Sea surface temperature decreased slightly compared to that in the Open-ocean region and the average value was -0.72°C.With the rapid sea ice changing, sea surface temperature and salinity varied sharply from -1.3°C to 0.5°C and from 31.8 to 33.3 respectively.When sea ice melted, water temperature increased, biological activities became active and chlorophyll-a value increased by a small amount to an average of 0.51 mg/m 3 .Due to the rapid change of sea ice cover, the value of MLD varied from 12.8 m to 30.9 m.The average value of oceanic pCO2 was 276.48 μatm ranging from 190.46 μatm to 364.43 μatm.

Biogeosciences
Fig. 4 a)The cruise lines from SOCAT to validate the SOM derived oceanic pCO2 for the study period in 2015; b) Comparison between the SOM derived and observed SOCAT oceanic pCO2.

28 Biogeosciences
Fig. 5 a) Sea ice extent (unit: 10^4 km^2) in study area (gray line) and three sub-regions (blue: Openocean region; red: Sea-ice region; green: Shelf region); b) Averaged ice concentration in three sub regions from Feb. 2, 2015 to Mar. 5, 2015.As shown in Fig.5-a, Sea ice in Open-ocean region and Sea-ice region started to melt from Jan 13, 2015, during February it decreased to the lowest and then it began to reform from Mar.3, 2015.The average sea ice extent in Open-ocean region and Sea-ice region were 3.8510 4 km 2 and 3.5610 4 km 2 .During our study period, in the Sea-ice region, sea ice kept melting and reforming rapidly and the average value of sea ice coverage percent is 29.54%.Oceanic pCO2 changed sharply from 155.86 μatm to 365.11 μatm.

BiogeosciencesFig. 7 3 km 2 . 36 Biogeosciences
Fig. 7 Timeseries of weekly averaged △pCO2, wind speed and uptake of atmospheric CO2 in Open-ocean region (blue line, the negative value means the direction from sea to air), Sea-ice region (red line) and Shelf region (yellow line).

oceanic
pCO2 and succedent sea-air carbon flux especially in dynamic, high latitude, seasonally ice-covered region.The estimated results revealed that the SOM technique could reconstruct the variations of oceanic pCO2 associated with bio-geochemical processes expressed by the variabilities in four proxy parameters: SST, CHL, MLD and SSS.The RMSE of the SOM derived oceanic pCO2 is 22.14 μatm for the SOCAT dataset.Over February 2015, Prydz Bay region was a strong carbon sink with a carbon uptake of 18.74.93TgC.Strong potential uptake of anthropogenic CO2 in the Shelf region will enhance the acidification in the deep water of Prydz bay and then may influence the deep ocean acidification in the long run since it contributes to the formation of Antarctic bottom water.Biogeosciences Discuss., https://doi.org/10.5194/bg-2018-276Manuscript under review for journal Biogeosciences Discussion started: 19 July 2018 c Author(s) 2018.CC BY 4.0 License.

Table 1 .
Statistics of labeling and training data sets showing the distribution and coverage of each variable.