The Potential of using Remote Sensing data to estimate Air-Sea CO2 exchange in the Baltic Sea

In this article, we present the first climatological map of air–sea CO2 flux over the Baltic Sea, based on remotesensing data: satellite imaging derived estimates of pCO2 using self-organizing maps classifications along with class-specific linear regressions (SOMLO methodology) and remote-sensed wind estimates. The estimates have a spatial resolution of 4-km both in latitude and longitude and a monthly temporal resolution from 1998 to 2011. The CO2 fluxes are estimated using two types of wind products, i.e. reanalysis winds and satellite wind products, the higher-resolution wind product generally leading 5 to higher-amplitude fluxes estimations. Furthermore, the CO2 fluxes were also estimated using two methods: the method of Wanninkhof et al. (2013) and the method of Rutgersson and Smedman (2009), i.e. reanalysis winds and satellite wind products, the higher-resolution wind product generally resulting in higher-amplitude fluxes. The seasonal variation in fluxes reflects the seasonal variation in pCO2 and is similar over the whole Baltic Sea, with high winter CO2 emissions and high cCO2 uptakes. All basins act as a source 10 for the atmosphere, with a higher degree of emission in the southern regions (mean source of 1.6 mmol m−2 d−1 for the South Basin and 0.9 for the Central Basin) than in the northern regions (mean source of 0.1 mmol m−2 d−1) and the coastal areas act as a larger sink (annual uptake of -4.2 mmol m−2 d−1) than does the open sea (-4 mmol m−2 d−1). In this study, we find that the Baltic Sea acts as a small source of 1.2 mmol m−2 d−1 on average and that annual uptake has increased from 1998 to 2012. 15 ir-sea CO2 flux, Baltic Sea, neural method, climatology.

water column, as well as light availability. In the Baltic sea, the former factors are affected by physical constraints such as the stratification of the water, the salinity and temperature profiles as well as the sea currents.
In recent years, the Baltic Sea has also been paid more attention as a coastal system affecting both the uptake/release of anthropogenic CO 2 and the natural CO 2 cycle ( ( Thomas and Schneider, 1999;Lansøet al., 2015)). Between 1994 and 2008 direct CO 2 measurements from a cargo ship has been mesured at monthly resolution The net annual air-sea exchange of CO 2 5 in the central Baltic Sea and the Kattegat varied both regionally and inter-annually. In the examined period, the Kattegat sea was, on average, a sink of CO 2 while the East Gotland and Bornholm seas were sources. The air-sea exchange of CO 2 and gas transfer velocity interannual variations were more pronounced in winter periods than in the summer periods. This indicates the interannual variability in the annual net flux is mainly controlled by the winter conditions (Wesslander et al., 2010).
The balance between mineralization and production, as well as the depth of the mixed-layer in the different oceanic zones 10 examined were shown to be the main drivers of their respective sink / source distributions (Wesslander et al., 2010). In the central Baltic Sea, where the mixed-layer depth is 60 m, CO 2 -enriched water mixes with water up to the surface in winter. The central Baltic Sea also receives large amounts of organic material from river water inflow; this may give rise to a heterotrophic system, making the central Baltic a net CO 2 source. This is not the case in the Kattegat, which is highly influenced by oceanic conditions. 15 In this study, the air sea CO 2 flux is estimated, with the ocean-surface pCO 2 in the Baltic Sea estimate from satellite-data derived products in (Parard et al., 2015(Parard et al., , 2016 where we used the self-organizing multiple linear output (SOMLO) method (Sasse et al., 2013). The outputs of the method have a horizontal resolution of 4 km and cover the period from 1998 to 2011.
Previous studies of the net uptake or release of CO 2 in the Baltic Sea have produced a wide range of results, with net exchange varying between -3.6 and +2.9 mol CO 2 m −2 y −1 in different time periods between 1994periods between and 2009periods between (Norman et al., 2013b. 20 The goal of the present study is to develop an air-sea CO 2 flux climatology based on remote-sensing products with a monthly time resolution and 4 • spatial resolution. In addition, we will further describe the processes and air-sea fluxes of CO 2 from 1998 to 2011 in the entire Baltic Sea. The study is structured in four sections. Section 2 presents the data and method used in this work. Section 3 presents the wind products used to estimate the exchange (based on satellite data and reanalysis data). In Section 4, we analyze the wind 25 products' quality, as well as various aspects of the estimated fluxes , and in Section 5 we present our conclusions.

Wind products
In this study We used wind products to calculate the transfer velocity, based on a meso-scale reanalysis product.
The wind product is based on a meso-scale modeling reanalysis product. A reanalysis is a combination of measurements 30 and a model in which the available data are assimilated into a high-quality modeling system. The reanalysis used here is from the Swedish Meteorological and Hydrological Institute (SMHI) with the High-Resolution-Limited Area Model (HIRLAM) geometry (22-km horizontal grid spacing and 60 levels in the vertical; the model top is at 10 hPa) (Soci et al., 2011) . HIRLAM is downscaled and dynamically adapted to a higher resolution (5-km grid) with a simplified HIRLAM called the Dynamic Adaptation Model (DYNAM). The observations of 10-m winds assimilated into the system are from four databases: the Integrated Surface Database Station History (ISH) database maintained by NOAA's National Climatic Data Center (NCDC), the MARS archive at ECMWF, the European Climate Assessment & Dataset (ECA&D) used as input for E-OBS version 6.0, and the national climate databases of SMHI and Météo France (MF). The temporal resolution is of 6 hours. In the following, this 5 product will be referred to as SMHIp. The method requires for the explicative data to stay coherent in terms of resolution, and as such we chose a temporal and spatial resolution of monthly, 4 x 4 km pCO 2 pixels.

Calculation of CO 2 flux
The flux of CO 2 (FCO 2 ) from sea to air (positive value) or air to sea (negative value) is often calculated using the difference in the partial pressure of CO 2 between the surface water and the atmosphere (∆pCO 2 ).

10
Here, the atmospheric pCO 2 was estimated using the method from Rutgersson et al. (2009)  The SOMLO methodology combines two statistical approaches: self-organizing maps (SOMs) (Kohonen, 1990) and linear regression.
In addition, the exchange efficiency was required, which was expressed in terms of a transfer velocity, k. The flux was then 15 calculated according to: where K 0 is the salinity-and temperature-dependent solubility constant (Weiss et al., 1982). The gas transfer velocity was computed using the parameterization from (Wanninkhof et al., 2009): 20 where U is the wind velocity at a reference height of 10 m and Sc is the solubility-dependent Schmidt number. Daily values of k were computed with a 6-h frequency for SMHIp; Eq. 2 is valid for all wind speed ranges. This method will be define as Method 1.
We compare the results with another method to compute the transfer velocity k from Rutgersson and Smedman (2009) where w is the water-side convection this is estimated from the model used in Norman et al. (2013b). This method will be defined as Method 2 .

Results
3.1 Analysis of the wind products 3.1.1 Validation of the wind product To validate our wind product, we compare the SMHI product with one based on remote-sensing data at daily scale 10 m wind data are reprocessed QuikSCAT (QSCAT) and ASCAT data (Bentamy and Croizé-Fillon, 2013) with a spatial resolution of 5 25x 25 km here called SATp. The two products are quite coherent when compared to all the station data used here, though SMHIp seems better, having a higher average correlation coefficient, i.e. R = 0.84 versus 0.67 for the remote sensing data wind (we chose not to show here). This is to be expected, as SATp has a much coarser spatial resolution (25 km) than SMHIp does (5 km). In the following we decided to used the SMHI product to compute the transfer velocity.
The wind product SMHIp used here to compute the air-sea CO 2 flux was compared with wind-tower data available from 10 24 stations in the Baltic Sea, including data from the Östergarnsholm measurement site Högström (2008) 1995-2002 and 2005-2009 periods. In addition, we validated the winds using synoptic station data from SMHI for 21 sites along the coast of Sweden.

15
The wind product SMHIp agree quite well with the station data ( Table 1). Most of the synoptic stations are very close to the coast, so there might be a bias due to land influence. The correlation coefficient (R) is quite high (0.66-0.91).
The root-mean-square differences (RMSDs) is given in Table 1.
The SMHIp have a quite high average correlation coefficient, i.e. R = 0.84 (Table 1). This is to be expected given that the spatial resolution is quite high for SMHIp (5 km). 20 We increase the resolution of the wind products by means of linear interpolation to compute the air-sea CO 2 flux. This was done to provide coherency between our datasets.

Wind variability over the Baltic Sea.
We examine the annual and monthly mean wind speeds and wind variability for the entire Baltic Sea ( (Figure 4). In all the basins the uptake is larger and April and May for the later period, the differences is particularly large in the basins most influenced by ice cover (GB and GF). There is also an indication in GB and 5 GF for a reduced outgassing in early winter. As the data is not entirely homogeneous (different satellite products are used in the beginning and ending of the studied period) one should not draw too far conclusions from the suggested trend. It could, however, be related to the higher pCO 2 concentrations in the atmosphere due to anthropogenic emissions, the corresponding increase in CO 2 concentration in the atmosphere during this period is 23.7 µatm. As the trend to a large extent is explained by an earlier onset of spring-time uptake differences in temperature and ice cover might be a more likely explanation.

Uncertainty analysis
The difference between the phase before 2003 and after 2007 could be explained by the repartition of the data used to calculate our results. In order to understand if this repartition of the initial data is responsible for the phase difference, we studied the representation of the data along the different years for each neuron of the SOM maps in each basin (Figure 7). For the three 30 first basins (Figure 7,a.,b.,c.), all the years are present at least in part, even if some classes seem to be solely composed from data measured before 2002, in particular in the Southern regions (the blue trend color classes). In the Gulf of Finland there is no data before 2008 so the results that we show can be affected by this lack of data, yet is coherent with the other basins. The distribution of the data is well spread (Figure 7,e.,f.,g.,h.) throughout the classes. Nevertheless, the seasonal cycle from air-sea CO 2 flux using SATp product is larger, with lower value in summer and higher in winter. we observe the maximum difference in January (when the flux using SMHIp winds is higher) and in September (when the flux using SATp winds is higher). The monthly variability of the flux using SMHIp winds is 8.7-11.4 mmol m −2 d −1 versus 3.4-13.4 mmol m −2 d −1 using SATp winds. High variability in January using the SATp wind product can be explained by the lack of satellite data during for this month. In addition, there are also interannual variations. In most years, the Baltic 10 Sea acts as a sink: using the SMHIp winds, the exchange ranges from -2.9 to 0.6 mmol m −2 d −1 with an average of -1.6 mmol m −2 d −1 ; using the SATp winds, the annual uptake is larger, being between -3.9 The two methods to compute the air-sea CO 2 flux have been used, one from (Wanninkhof et al., 2009)

Air-sea CO 2 flux climatology
The climatology of the flux displays high seasonal and spatial variability, ranging from -13. to 10 mmol m −2 d −1 . On average 25 from 1998 to 2011, the entire Baltic Sea acts is a source of 1.2 mmol m −2 d −1 ( 1.4 mmol m −2 y −1 using method from Rutgersson et al. (2009) and a source of -1.5 mmol m −2 y −1 using SATp winds) (Fig. 9). The values observed are in agreement with those from other studies, indicating that the Baltic Sea can be a small source on average or a small sink of CO 2 . Most previous research results concerning the carbon budget cover shorter periods, indicating a range between -1.16 and 2.9 mol m −2 y −1 )(e.g. Wesslander et al., 2010;Kulinski and Pempkowiak, 2012), though the maximum values reported in these studies 30 are all found in the same one or two years Algesten et al. (2006). Half of the studies demonstrate that Baltic Sea or certain basins of it act as sources, while the others demonstrate that it acts as a sink for the atmosphere (Norman et al., 2013a).

8
Earth Syst. Dynam. Discuss., doi:10.5194/esd-2017-33, 2017 Manuscript under review for journal Earth Syst. Dynam.  Canadell (2003) explain that it is really challenging to estimate precisely the variation of the pCO 2 over marginal seas. This is due to several aspects but mainly due to the lack of data in space and time. Remote sensing using applicable algorithms could certainly be an important approach, complementing ship-board observations as well as in situ buoy and wind-tower measurements. Using our method, we present the first estimated CO 2 flux climatology based on remote sensing for the Baltic In the Central Basin, Schneider et al. (2014) demonstrate that in four selected years (i.e. 2003,2004, 2009, and 2010), the surface water acts as a sink for the atmosphere, as found in our study, the value of the uptake rates ranging between -0.04 and -0.3 mol m −2 yr −1 . One study explain that the rates is the one which explain the enhance carbon in the sediments (Schneider and a smaller source in Bothnian Bay (0.14 mol m −2 yr −1 ) between 1999 and 2009; this finding could explain why the entire Gulf of Bothnia region is a small sink or small source on average.
Using remote sensing data to compute the FCO 2 gives good spatial and temporal resolutions compared with those of measurements from ships or wind-towers. The satellite data give information on pCO 2 variability and on FCO 2 . The first estimates 30 of Baltic Sea air-sea exchange based on remote-sensing products display reasonably good agreement with previous estimates and indicate a negative trend, with annual uptake changing from 0.6 to -2.8 mol m −2 yr −1 ) over the 1998-2007 period. After 2007, the decrease is smaller and the flux remains quite stable at around -2.8 mol m −2 yr −1 ). The pCO 2 flux product depends Thomas, H. and Schneider, B.: The seasonal cycle of carbon dioxide in Baltic Sea surface waters, Journal of Marine Systems, 22, 53-67, doi:10.1016/S0924-7963(99)00030-5, 1999.
Thomas, H., Pempkowiak, J., Wulff, F., and Nagel, K.: Autotrophy, nitrogen accumulation and nitrogen limitation in the Baltic Sea: A Table 1. RMS, bias, and correlation coefficients for in situ data from SMHI, Östergarnsholm wind-tower, and satellite products.