The effect of assimilating satellite derived soil moisture in 1 SiBCASA on simulated carbon fluxes in Boreal Eurasia 2

Boreal Eurasia is a region where the interaction between droughts and the carbon cycle may 2 have significant impacts on the global carbon cycle. Yet the region is extremely data sparse 3 with respect to meteorology, soil moisture and carbon fluxes as compared to e.g. Europe. To 4 better constrain our vegetation model SiBCASA, we increase data usage by assimilating two 5 streams of satellite derived soil moisture. We study if the assimilation improved SiBCASA’s 6 soil moisture and its effect on the simulated carbon fluxes. By comparing to unique in-situ 7 soil moisture observations, we show that the passive microwave soil moisture product did not 8 improve the soil moisture simulated by SiBCASA, but the active data seem promising in 9 some aspects. The match between SiBCASA and ASCAT soil moisture is best in the summer 10 months over low vegetation. Nevertheless, ASCAT failed to detect the major droughts 11 occurring between 2007 and 2013. The performance of ASCAT soil moisture seems to be 12 particularly sensitive to ponding, rather than to biomass. The effect on the simulated carbon 13 fluxes is large, 5-10 % on annual GPP and TER, and tens of percent on local NEE, and 2% on 14 area-integrated NEE, which is the same order of magnitude as the inter-annual variations. 15 Consequently, this study shows that assimilation of satellite derived soil moisture has 16 potentially large impacts, while at the same time further research is needed to understand 17 under which conditions the satellite derived soil moisture improves the simulated soil 18 moisture.


Introduction 21
The interest of this publication is to explore the potential of assimilating satellite derived soil 22 moisture in the vegetation model SiBCASA over Boreal Eurasia with particular focus on the 23 impact on simulated carbon fluxes. In remote regions as Boreal Eurasia meteorological driver 24 data for vegetation models (temperature, precipitation, etc.) are poorly constrained by surface 25 observations, leaving room for improvement in the soil moisture content simulated in 26 vegetation models. Boreal Eurasia is also a region with large carbon stocks in biomass and 27 vegetation (Schepaschenko et al., 2013;McGuire et al., 2009;Tarnocai et al., 2009), which are 28 subject to the fastest climatic change rates on Earth (Goetz et al., 2007), making it a relevant 29 region in the context of ecosystem carbon sequestration. Furthermore, large parts of the 30 region are located in continuous and discontinuous permafrost soils. In the spring, melt water 31 from the accumulated winter precipitation cannot percolate into the still frozen soil and runs 32 off in hilly terrain, and forms ponds on the soil in flat terrain. This causes a bi-modal spatial 1 distribution in soil moisture, with dry hills and wet plains. This process is probably hard to 2 catch by land surface models and satellite observations alike. Satellite derived soil moisture is 3 observed to have large variability in the Northern regions, both within and between different 4 approaches (Mladenova et al., 2014). The derivation of satellite soil moisture is difficult in 5 snow, ice and surface water rich areas (Högström et al., 2014;Naeimi et al., 2012a). These 6 aspects make a working soil moisture data assimilation system very relevant for vegetation 7 modelling in the region, and challenging. The few tower sites running in the region now have 8 fairly long measurement records, so that they can be used for validation of inter-annual 9 variation (e.g. the 2010 drought). An effort specifically targeted at Boreal Eurasia and carbon 10 fluxes has not been tried before . 11 Earlier efforts to assimilate satellite derived soil moisture in vegetation models were often 12 focussed on the improvement of soil moisture itself and/or validated with in-situ observations 13 over short vegetation in temperate and Mediterranean climate zones (Reichle and Koster, 14 2005;Reichle et al., 2007;Draper et al., 2009a;Draper et al., 2009b;Miralles et al., 2011b). 15 Other studies focus on the effect on crop yield and carbon fluxes in Europe (Verstraeten et al., 16 2010;de Wit and van Diepen, 2007;Han et al., 2014). The Global Land Data Assimilation 17 Systems (GLDAS, http://ldas.gsfc.nasa.gov/GLDAS/, cf. (cf. Chen et al., 2013)) is worth 18 mentioning here too. Our study is thus the first to assimilate satellite soil moisture in Boreal 19 Eurasia and to evaluate the impact on simulated soil moisture and carbon fluxes with in-situ 20 data. 21 Soil moisture affects vegetation carbon fluxes through photosynthesis and respiration rates . 22 Photosynthesis rates depend on the stomatal conductance, which the plants regulate according 23 to the water potential in the leaf and the atmospheric vapour pressure deficit. The water 24 potential in the leaf is a function of water supply by the roots and the water use by 25 transpiration (Katul et al., 2010). Heterotrophic respiration depends on the soil moisture 26 content, which is the substrate in which microbes and bacteria consume organic matter and 27 release CO 2 . The dependence of photosynthesis, transpiration and respiration fluxes on soil 28 moisture is implemented in virtually all contemporary vegetation models via similar types of 29 drought sensitivity functions (Verhoef and Egea, 2014;Sellers et al., 1996;Vetter et al., 2008). 30 Based on simulations with global climate models, it is expected that global warming is 31 associated with more extreme precipitation regimes, resulting in more frequent and more 32 FIGURE 1 about here 23 Soil moisture also affects the turnover times of organic matter in the soil. The turnover times 24 are shortest at an optimal soil moisture saturation fraction (which varies around 60% of the 25 pore space) and from there increases towards dryer and wetter soils ( Fig. 1), (Raich et al., 26 1991). The respiration rates are a function of the carbon pools and turnover times, which are 27 temperature and soil moisture dependent. 28 The net effect of soil moisture on Net Ecosystem Exchange (NEE) is the sum of the effect on 29 photosynthesis (GPP, Gross Primary Production) and on respiration rates (TER, Total 30 Ecosystem Respiration). 31 reliable for sparse and moderately vegetated areas, and less for bare soils (Liu et al., 2012). 23 We use METOP ASCAT25, version WARP5.5, release 2.1, with a 0.25 degree spatial 24 resolution, and a 1-day temporal resolution) (Wagner et al., 1999;Naeimi et al., 2012b;Naeimi 25 et al., 2009). The period of record of the ASCAT data constrains the study period to  2013. 27

Assimilation method 28
The objective of this paper is to attempt an improvement of soil moisture dynamics in 29 SiBCASA by assimilating the passive microwave and/or ASCAT satellite derived data 30 described above. We use the same assimilation method as used in GLEAM (Miralles et al., 1 2011a;Miralles, 2011) with w t x the soil moisture content of the top soil layer, as satellite-observed (x=obs) or 4 simulated (x=sim), dw the change in w after assimilation. The index t indicates the time in 5 steps of days. K t , the Kalman gain, describes how much of the difference (w t obsw t sim ) is 6 applied to dw to update w sim , and depends on the error in the model soil moisture σ sim and the 7 error in the satellite observed soil moisture σ obs : 8 The error in model soil moisture depends on dσ mod , the uncertainty associated with model 10 integration over a time step of one day: 11 We use a constant dσ mod = 0.01 m 3 m -3 day -1 . The model soil moisture is updated according to: 13 When observations were available, the model error σ t mod is reduced to σ t + after the 15 assimilation step: 16 The error or noise in the satellite observed soil moisture, σ obs , depends on the vegetation 18 optical depth, land surface heterogeneity of the pixel and snow or ice in or on top of the soil, 19 and is output by the retrieval algorithm. The noise is typically in the order of 0.1 m 3 m -3 20 (standard deviation σ, e.g. Fig. 5 -8). 21 Since the satellite observed soil moisture data essentially carry information only about the 22 temporal variations in soil moisture, and not about the absolute value of mean and the 23 amplitude of the variations, we normalise the satellite data (see below) before assimilating the 24 satellite data in SiBCASA. The entire assimilation procedure consists of the following steps: 25 Step 1: Run SiBCASA without data assimilation to equilibrium in 2007 and then run until 26 2013, storing daily model results. 27 Step 2: Take the spatial average of satellite derived soil moisture within the 1° × 1° SiBCASA 1 grid boxes. Normalise the satellite observed soil moisture's mean, standard deviation and 2 higher moments towards the SiBCASA's soil moisture using the CDF matching technique 3 (Liu et al., 2009;Reichle and Koster, 2004;Liu et al., 2012). We matched the distribution 4 function at every 10 th percentile between 10 and 90. Because the retrieval algorithms do not 5 work under (partially) frozen and snow-covered conditions, we discarded periods with frozen 6 soil in SiBCASA from building the CDF transformation coefficients. 7 Step 3: Run SiBCASA from equilibrium in 2007 until 2013 with assimilation of the satellite 8 derived soil moisture. 9 Step 4: Evaluate the simulated soil moisture, water and carbon fluxes with in-situ flux tower 10 data described in section 2.4. 11

In-situ flux tower data 12
For evaluation of the model results, particularly the carbon fluxes, we use 25 site-years of 13 data from 4 flux tower sites in Boreal Eurasia available in the period 2007-2013 (table 1). The 14 sites vary in vegetation type, continentality and permafrost. Flux data were taken in Siberia at 15 more locations, although predominantly in the period 1997-2005, when the ASCAT satellite 16 was not yet launched (Dolman et al., 2012). The sites are in Hyytiälä, Finland (Ilvesniemi et 17 al., 2010;Kolari et al., 2009;Mammarella et al., 2009;Rannik et al., 2004), Tver, European 18 Russia (Kurbatova et al., 2008;Milyukova et al., 2002), Yakutsk, East Siberia (Dolman et al., 19 2004;Ohta et al., 2008) and Elgeeii, East Siberia (Kotani et al., 2014). The eddy covariance 20 data have been processed according to the harmonized LaThuille FLUXNET database 21 (Baldocchi et al., 2001;Reichstein et al., 2005;Papale et al., 2006;Moffat et al., 2007) We will first evaluate the spatial coherence between the simulated and satellite observed soil 28 moisture, then the temporal coherence. All satellite and in-situ data are CDF-matched to the 29 SiBCASA soil moisture. Next, we will compare model and satellite soil moisture data with in-30 situ observations. Finally, we will evaluate the impact of satellite soil moisture assimilation 1 on the simulated carbon fluxes. 2

Eurasia 4
The spatial coherence between SiBCASA and satellite observed soil moisture is studied by 5 comparing maps of monthly soil moisture anomalies and by quantifying the spatial correlation 6 between the anomalies. We compute the anomalies for each month with respect to the average 7 soil moisture in that month in the years 2007 to 2011. We use the reference period until 2011 8 (and not 2013), because the AMSR-E satellite became dysfunctional in November 2011, and 9 we have no passive microwave soil moisture after that date. Fig. 2 shows an example of the 10 spatial coherence in August 2009. This is the month with the largest spatial correlation 11 between SiBCASA and ASCAT soil moisture in the period that AMSR-E data are available 12 (r=0.60). A dry anomaly in North-central Siberia is apparent in both SiBCASA and ASCAT 13 and a wet anomaly in South and West Siberia, showing that there is coherence in the spatial 14 structure of both data sources. But there are also striking differences. The drought is more 15 intensive and confined to a smaller area in SiBCASA as compared to ASCAT. In addition, 16 east from the Lena river, SiBCASA tends to have a wet anomaly, where ASCAT has a light 17 dry anomaly. Nevertheless, the spatial correlation coefficient is good with r=0.60 (second 18 panel in Fig. 2). The correlation appears to be better for wet anomalies than for dry ones. If 19 we compare the soil moisture from the passive microwave data to SiBCASA and ASCAT, no 20 coherent pattern emerges, neither in August 2009 nor in other months. This is reflected in the 21 low spatial correlation coefficient (r=0.03). 22

FIGURE 2 about here 23
These findings are also quite typical for August months in other years. Only in August 2013 24 the spatial correlation coefficient between SiBCASA and ASCAT soil moisture was slightly 25 larger, r=0.62 (figure not shown), but the corresponding AMSR-E data were no longer 26 available. For other months, the spatial patterns are usually less pronounced, and the 27 correlation coefficients smaller. Fig. 3 shows the seasonal evolution of the spatial correlation 28 coefficients, also for the land cover types tundra, forest and steppe (i.e. grasslands and 29 croplands) separately. The correlation coefficients are generally small outside the summer 30 months. Steppe regions have larger correlation coefficients, and tundra regions smaller ones. 31 The overall correlation coefficient is strongly dominated by the forests, because forests cover 1 by far the largest part of the study region (66%), versus 24% for tundra and 9% for steppe. In 2 the discussion we will provide potential explanations for the variation of the correlation 3 coefficients over the seasons and over land cover types. 4 FIGURE 3 about here 5 Considering the seasonal evolution of the correlation coefficients between SiBCASA and 6 passive microwave soil moisture, there is no coherence between the two, except perhaps for 7 the steppe regions, for which the correlation coefficients reach to 0.50 in Septembers. 8 However, the slope of the regression curve is only about 1:3 (third panel of Fig. 2), whereas it 9 is near 1:1 for ASCAT soil moisture (second panel of Fig. 2). Because the prior agreement 10 between SiBCASA and passive microwave soil moisture is too low for Boreal Eurasia, we do 11 not consider it meaningful to proceed with assimilation in SiBCASA. We will focus on the 12 ASCAT soil moisture alone in the remainder of this publication. This decision will be further 13 addressed in the discussion section. 14 The spatial correlation is a measure of how well satellite and SiBCASA agree on the location 15 of drought and wet regions. For assimilation purposes it is also interesting to investigate the 16 temporal correlation at each location. The temporal correlation coefficient is a measure of 17 how well satellite and SiBCASA agree on the timing of dry and wet periods. Central Siberian Plateau, where the correlations are much smaller (0 < r < 0.2). In Eastern 25 Siberia, East of the Lena river, the correlations are variable, but generally small and 26 sometimes even negative (-0.3 < r < 0.3). This pattern is somewhat representative for July, 27 August and September (see Figs. S1 and S2), except that in September the area North East of 28 60°N, 90°E is masked out for the lack of good quality satellite data. Since we do not calculate 29 correlations when SiBCASA simulates frozen soil or when good quality data are lacking, an 30 apparent 'winter mask' advances from the North East in September to cover all of the region 31 by December. In April this winter mask regresses into Siberia and disappears in June. As 32 expected, at the front of the winter mask, which is generally 5 ° wide, the correlations are low 1 and patchy. We will consider potential underlying reasons for these patterns in the discussion. 2 FIGURE 4 about here 3 The temporal correlations shown in Fig. 4 were computed with soil moisture on a daily basis. 4 When computed on a monthly average basis, the correlation coefficients generally improve 5 considerably, but the variability in soil moisture is smaller accordingly. This shows that day-6 to-day noise in particularly the satellite soil moisture is responsible for the low correlation 7 coefficients, and that the remaining correlation is dominated by inter-annual variations. This 8 suggests that it may be worthwhile to investigate assimilation of low-pass filtered satellite soil 9 moisture instead of instantaneous measurements. 10 To better understand what these large-scale spatial and temporal correlation coefficients imply 11 for the use of satellite soil moisture for assimilation in SiBCASA, time series of soil moisture 12 were compared for 4 stations across Eurasia. In these time series we show the original 13 SiBCASA soil moisture, the ASCAT soil moisture (CDF-matched), and in-situ soil moisture. 14 In addition, we show the SiBCASA soil moisture after assimilation of ASCAT soil moisture 15 according to section 2.3. The first station is Hyytiälä in Finland. The time series are shown in 16 larger than SiBCASA soil moisture in the spring period (May to early June). This causes the 20 assimilation procedure to increase soil moisture in SiBCASA (red line is higher than the blue 21 line). This increase of soil moisture in the spring period improves the match with in-situ 22 observed soil moisture to the degree specified by the uncertainties (Eq. (2)). 23

FIGURE 5 about here 24
We loosely define a drought as a period when in-situ observed soil moisture is more below 25 the average soil moisture in that period (see bottom panels of Fig simulation also 'saw' the drought, but ASCAT did not. The assimilation therefore decreased 29 the match with in-situ soil moisture. In 2013 the in-situ observed drought in Hyytiälä was 30 picked up by neither SiBCASA nor ASCAT. 31 The second site we analyse is 'Tver wet forest', for which the time series are shown in Fig. 6. 1 In Tver, a similar spring time behaviour emerges. ASCAT is generally larger than SiBCASA 2 soil moisture in April to early May, and the assimilation improves the match with in-situ 3 observed soil moisture. Field workers confirm that the water table is generally high or even 4 above the soil surface after snow melt (April) and decreases quickly in May. In the summers, 5 soil moisture is generally quite constant, except during the 2010 drought, which caused 6 extensive fires in European Russia. The drought was picked up by in-situ observations and 7 SiBCASA, but not by ASCAT. As a consequence, the assimilation decreased the match of 8 SiBCASA with in-situ soil moisture. 9 FIGURE 6 about here 10 The third site is Yakutsk Larix. At this site, ASCAT soil moisture is noisier than for the other 11 sites (Fig. 7). There is a tendency that in-situ soil moisture is high in spring, due to melted water logging had severe impact on the ecosystem, with reduced photosynthesis rates and tree 19 browning and mortality. This water logging may be a larger-scale process in eastern Siberia 20 (Muskett and Romanovsky, 2009;Vey et al., 2013). Therefore the term 'drought' is relative to 21 the studied period. SiBCASA sees the in-situ observed droughts in 2008 and 2012 to some The fourth site is Elgeeii, for which the time series are shown in Fig. 8. SiBCASA soil 30 moisture is again low in the spring-time (May), and ASCAT soil moisture is larger. The 31 assimilation increases the soil moisture in SiBCASA, and this seems to improve the match 32 with in-situ observed soil moisture, although the early spring-time in-situ observations were 1 unreliable. In 2012 a drought occurred in July and August. SiBCASA sees the drought too, 2 although with a much earlier development. ASCAT does not see the drought, and as a 3 consequence, the assimilation moves the soil moisture in SiBCASA away from the in-situ 4 observations. It is interesting to see that on 5 August 2012 the soil moisture in SiBCASA 5 increases due to a precipitation event, but this is not seen in the in-situ observations. However, 6 an increase is seen at that time in the in-situ observations in Yakutsk, some 340 km to the 7 North East, possibly suggesting a displacement of the precipitation event in the SiBCASA 8 driver data from ECMWF ERA-interim. 9 Siberia, which was caused by a strong blocking situation. The drought was accompanied with 20 many wildfires (Miralles et al., 2014;Krol et al., 2013). The drought was also apparent in the 21 in-situ measurements performed in Hyytiälä and Tver. Fig. 9 shows that SiBCASA simulates 22 a drought extending from Scandinavia to Novosibirsk (55°N, 80°E), with the Hyytiälä and 23 Tver sites on the western rim of the drought region. ASCAT locates a drought in roughly the 24 same region, although less intense and with a smaller geographical extent. The ASCAT wet 25 anomaly over Europe expands further into Russia and Scandinavia. As a result the sites 26 Hyytiälä and Tver are just outside of the drought region as observed by ASCAT and this is 27 most likely attributable to the ASCAT soil moisture retrieval skills. Fig. 4 shows that the 28 ASCAT performance is low around those sites. In section 5, Fig. 12 we will discuss the 29 performance at the sites in more detail. 30 FIGURE 9 about here 31 In July 2012 SiBCASA simulates an intense drought that was located around the city 1 Yakutsk, extending eastward to the region between the Lena and Aldan rivers (Fig. S3). The 2 Elgeeii site was located just on the Eastern border of the simulated drought region. ASCAT 3 does not observe this drought region, not in July, nor in earlier or later months. Where Elgeeii 4 was on the perimeter of the 2012 drought region, as were Hyytiälä and Tver in 2010, Yakutsk 5 was in the centre of the drought region, which ASCAT does not observe at all. Therefore a 6 site's location on the rim of a drought does not explain why ASCAT does not observe the 7 drought. Rather it appears that ASCAT has limited capability to observe droughts in the 8 forested zone where the in-situ observations were made. The changes in soil moisture in springtime and during drought, induced by the assimilation of 12 ASCAT observed soil moisture in SiBCASA (section 3.1 Fig. 5 -8), may have substantial 13 effects on the representation of the carbon fluxes, which we will look at next. We will show 14 separately how GPP, TER and NEE depend on the change in soil moisture and the season. 15 An interesting case is presented at the Yakutsk Larix site (Fig. 10). At this location, the 16 change in soil moisture due to assimilation of ASCAT soil moisture was large relative to the 17 other sites (Fig. 7). However, although the absolute value of the change in soil moisture was 18 more or less constant throughout the years, the change in GPP shows a distinct seasonal cycle, 19 with large changes in spring, small changes in summer and hardly any change in fall. This is 20 because of two reasons: i) the sensitivity of GPP to soil moisture is simulated as a function of 21 the plant available water fraction (section 2.1, Fig. 1a). In Yakutsk in spring, the permafrost 22 soil has only thawed for a couple of centimetres, resulting in a very small plant available 23 water fraction and very strong soil moisture sensitivity (Fig. 1a: the soil moisture sensitivity 24 curve is steepest on the low plant available water fraction side). This results in a strong GPP 25 effect of assimilating ASCAT soil moisture in SiBCASA. Note that soil thawing does not 26 automatically mean that more soil moisture becomes available for root uptake. The soils in 27 Yakutsk often freeze after a relatively dry summer, so that the frozen soil may be quite dry. In 28 the spring, the snow melt water cannot penetrate the soil, which is still frozen, and may run 29 off; ii) the Yakutian spring is almost simultaneous with the solar maximum on 21 June, so 30 that the potential GPP is large. In the course of the growing season, the permafrost active 31 layer thaws deeper, resulting in a larger plant available water fraction, reducing the drought 32 sensitivity. This explains the smaller change in GPP in the summer. In the fall, GPP is limited 1 more by the lack of available sunlight than by water stress, explaining the absence of change 2 in GPP with assimilation of satellite soil moisture. 3 FIGURE 10 about here 4 In a similar way, the change in TER (Fig. 10) does not only depend on the change in soil 5 moisture with satellite soil moisture assimilation, but also in the absolute value of soil 6 moisture (Fig. 1b) and temperature limitation on TER. In June, when the soil is still cold, the 7 changes in TER are small. In July and August the changes in TER are larger than in GPP, 8 because the soil is warm and TER is a function of absolute soil moisture change. In this 9 example, the changes in GPP and TER are into the same direction. Fig. 1 shows that this is 10 always the case when the soil moisture saturation fraction is below its minimum value of ca. 11 60 percent. Consequently, the changes in GPP and TER compensate each other partly in the 12

NEE. 13
Accumulated over a year (table 3), the changes in GPP, TER and NEE are in the order of tens 14 of gC m -2 yr -1 , amounting to a few percent of GPP and TER. For NEE however, the changes 15 can amount to tens of percent and a 7-year mean of -34 percent. We note that the changes in 16 GPP and TER are larger in Yakutsk than in Hyytiälä, Tver and Elgeeii. This is because the 17 plant available water fraction is smaller in Yakutsk than for the other sites, creating a strong 18 drought sensitivity, and because the change in soil moisture is larger. While the relative 19 changes in GPP and TER for these sites is generally small, and they partly compensate, the 7-20 year mean changes in NEE are +52% at Hyytiälä, -105% at Tver and -38% at Elgeeii. 21 The effects of ASCAT soil moisture assimilation in SiBCASA are also significant when 22 integrated over the entire study domain (27.8×10 6 km 2 ) and the year (Fig. 11). The mean 23 simulated NEE is -1.91 PgC yr -1 with an inter-annual variations of 0.12 PgC yr -1 (RMSD). 24 Assimilation of ASCAT soil moisture in SiBCASA causes a change of 0.045 PgC yr -1 25 (RMSD). This is 41% of the normal inter-annual variation of 0.11 PgC yr -1 (RMSD), and 26 2.4% of the mean NEE. The effect of satellite soil moisture assimilation is negligible until 27 May, it then grows in the months June and July. After August, the net effect does not change 28 much. This is in line with the observation that the effect of assimilation on soil moisture and 29 carbon fluxes is largest in spring time (section 3.1 and 3.2, Fig. 5 -8 and 10). The effect of 30 assimilation was largest in 2010 with an extra anomaly in NEE of + 0.08 Pg C yr -1 (less 31 uptake). This anomaly grew between May and September. In this year, a widespread drought 1 occurred in European Russia and West Siberia, which ASCAT captures quite well. The 2 assimilation effect could have been even larger if ASCAT had not wrongfully detected a wet 3 anomaly over far eastern Siberia, where SiBCASA simulates a second drought region (Fig. 9). 4 FIGURE 11 about here 5 The second largest effect of soil moisture assimilation occurred in 2012, with an extra 6 anomaly of -0.07 Pg C yr -1 (more uptake). This anomaly grew mostly in June and July, when 7 ASCAT soil moisture was much higher in June over large parts of Siberia, and the July 8 drought in Central Siberia was confined to a smaller region in ASCAT data. 9 10 4 Discussion 11

Soil moisture 12
The spatial and temporal correlation coefficients between SiBCASA and satellite observed 13 soil moisture shown in section 3 suggest that ASCAT and passive microwave satellite signals 14 have a certain skill in observing land surface soil moisture. The absence of perfect 15 correlations implies that assimilating the satellite observed soil moisture in SiBCASA will 16 have an effect. The question is whether that effect is an improvement. 17 The performance of passive microwave data was low over the entire study region and in all 18 months (Fig. 2, section 3.1). Only in steppe regions the temporal correlations were large (r = 19 0.8). The spatial correlation is smaller than that, (r ~ 0.5) and with a smaller sensitivity (a 20 slope of ca. 1:3, Fig. 2), probably because of the absence of significant spatial patterns in the 21 small extent of the steppe zone. The poor performance of the microwave soil moisture in 22 Boreal Eurasia is not entirely surprising: the passive microwave radiation emitted by the soil 23 moisture is known to be disturbed by vegetation, surface water, snow and ice (de Jeu et al., 24 2008;Mladenova et al., 2014;Champagne et al., 2010), which are abundant in Boreal Eurasia. 25 The microwave soil moisture product has been validated extensively (Miralles, 2011;Miralles 26 et al., 2011b;de Jeu et al., 2008;Liu et al., 2011;Liu et al., 2012;Owe et al., 2008;Griesfeller et 27 al., 2015;Champagne et al., 2010). However, the vast majority of the validation sites were 28 located on grasslands and croplands, and in temperate and (semi)arid climate zones. 29 Therefore, the poor performance of microwave soil moisture in Boreal Eurasia, except 30 perhaps the steppe zone, is probably related to the canopy, which is too dense, as well as to 1 the presence of snow, ice and surface water. Our results are therefore specific to our region, 2 and cannot be simply extrapolated to other climate zones and land covers. 3 The spatial and temporal correlation coefficients vary with the months and with land cover. 4 The spatial correlation between SiBCASA and ASCAT soil moisture is largest in August and 5 quickly decreases towards the spring and fall. What processes may cause this? Ecologically 6 there are large differences between the seasons in Siberia. Large parts of Siberia are snow 7 covered and particularly the region North of Mongolia and East of the Yenissei river is 8 subject to continuous permafrost. This hampers a correct retrieval of soil moisture from 9 satellite observed signals (Naeimi et al., 2012a;Högström et al., 2014), while correctly 10 simulating soil moisture under snow conditions is also difficult in vegetation models . 11 However, even in the Northern tundra regions most snow and ice have disappeared by June. 12 Considering that the grid cells with frozen top soil in SiBCASA and snow/ice detection in 13 ASCAT have been excluded from the statistical analysis, the lower correlations in June, July 14 and September (Fig. 3) are probably not only caused by the presence of snow and ice on the 15 land surface. 16 Other important changes from May to July are the expansion of leafs, the drying out of the 17 topsoil after snow melt on frozen ground, and the deeper thawing of the permafrost active 18 layer. The increase of the leaf area index (LAI) does not seem beneficial for better satellite 19 soil moisture retrievals, as is also suggest by the smaller correlation coefficients for forests 20 than for steppe zone (Fig. 3). The decrease in the ponded area fraction after snow melt on 21 frozen ground is a potential explanation for the improving correlation coefficients (Högström 22 et al., 2014), since they occur particularly in the forest and the tundra zones, which contain the 23 wettest parts of the region, and not for the steppe zone, which is drier and outside the 24 permafrost zone. 25 With the same arguments the increasing depth of the permafrost ice front may also be a 26 potential explanation of the improving spatial correlation coefficients towards August. Indeed, 27 ice and frozen soil at some depth may disturb the satellite signal (Way et al., 28 1997;Wegmüller, 1990). Maximum active layer thicknesses of a mere 10-20 cm are not 29 uncommon in the Northern tundra, although the penetration depth of microwave radiation in 30 the soil is in the order of one to a few centimetres. 31 It is interesting that the spatial correlation coefficients for steppe zones are larger and for 1 tundra zones smaller than average. Both steppe and tundra vegetation are characterised by 2 short vegetation, but tundra regions are generally much wetter than steppe regions and with 3 continuous permafrost. This implies that the presence of short vegetation alone is not the only 4 prerequisite to obtain a good match between SiBCASA and ASCAT soil moisture. 5 On the site level, Fig. 5 -8 show that ASCAT soil moisture has much more day-to-day 6 variability than SiBCASA soil moisture. While SiBCASA soil moisture has a significant, 7 physically meaningful auto-correlation with lag times up to 10-17 days (r > 0.3), ASCAT 8 observations and associated errors are independent in time, which indicates that the signal is 9 compromised by measurement noise. On top of this, ASCAT was not able to detect the 8 10 large drought occurrences observed in in-situ soil moisture time series, nor the pronounced 11 inter-annual variation associated with recovery after water logging in Yakutsk. This is 12 reflected in small site-level temporal correlation coefficients between in-situ soil moisture and 13 ASCAT soil moisture (r < 0.06 at all sites), while the June-September correlation between in-14 situ soil moisture and SiBCASA soil moisture is much larger (0.49 at Hyytiälä, 0.63 at Tver, 15 0.74 at Yakutsk and 0.76 at Elgeeii). The applicability of in-situ soil moisture observations for 16 this purpose is supported by Robock et al. (2000) and Mittelbach and Seneviratne (2012). 17 This suggests that SiBCASA soil moisture is more reliable than ASCAT soil moisture at these 18 sites. This is not entirely surprising, because Fig. 4 shows that the in-situ observations were 19 made at locations outside the area of high temporal correlations between SiBCASA and 20 ASCAT. However, it suggests that the low correlations outside the steppe zone are more 21 likely to be due to poor performance of ASCAT soil moisture than to SiBCASA soil moisture. indicates that ASCAT soil moisture was larger than SiBCASA soil moisture, and assimilation 28 seemed to improve the match with in-situ observed soil moisture. Is this a realistic pattern? 29 Experimentalists confirm that ponding after snow melt occurs on the sites. However, it is 30 known that ASCAT soil moisture is unreliable when the footprint of the observation is 31 (partially) covered with snow, ice or surface water, which is likely to happen in springtime. At 32 the same time, SiBCASA soil moisture in spring depends on the amount of snow accumulated 1 in the winter, the time of snowmelt, the fate of the meltwater on frozen ground (runoff or 2 ponding). Since it is hard to simulate these processes correctly, also considering the coarse 3 resolution of SiBCASA relative to dependency of these processes on topography, springtime 4 soil moisture in SiBCASA may also be questioned. Nevertheless this springtime 5 underestimation pattern is also observed at other steppe and forest grid cells where the 6 temporal correlations are large. Thus there are indications that the spring wetting with 7 assimilation of ASCAT data in SiBCASA improves the soil moisture. Field workers (see 8 author contributions) confirm the spring-time water logging and ponding at the four sites. 9 High water tables during spring-time are succeeded by drying out of the soil, depending on 10 the weather conditions. The low soil moisture in SiBCASA could be caused by 11 overestimation of the evapo-transpiration rates in the spring. 12 In an attempt to explain the variation in temporal correlation coefficients over the region, Fig.  13 12 shows the temporal correlation coefficient of SiBCASA and ASCAT soil moisture in 14 August 2013 as a function of several variables. Each dot in the figures represents a grid point. 15 With increasing LAI the correlation coefficient r indeed decreases, which is physically 16 logical, because water in leafs disturbs the soil moisture signal. Similarly the aboveground 17 carbon in biomass has a negative relationship with r for steppe, but not for forests and tundra 18 zones. For forests, the relationship is, counterintuitively, positive. This may be explained by a 19 cross correlation between carbon in biomass and temperature: the forest biomass decreases 20 towards the northern treeline, where temperatures are lower. Apparently, aboveground 21 biomass itself does not necessarily disturb the satellite signal. Soil temperature has a positive 22 relation with r, and there is no indication that the relationship saturates at higher temperatures. 23 This is a somewhat puzzling observation. We would have expected low correlation 24 coefficients at low temperatures, due to the presence of snow and ice, but at temperatures 25 higher than 10 °C the ice would have disappeared, and we would not have expected an 26 increase in r with temperature. Possibly, higher temperatures are indicative of a longer period 27 into the local growing season, when soil ponding has diminished after snowmelt and the 28 performance of SiBCASA is consequently better. This is confirmed by the negative relation 29 between top soil moisture in SiBCASA with r. At large soil moisture contents, the chance of 30 (partial) ponding is larger, with subsequent disturbances of the satellite signal (See Naeimi et 31 al., 2012b;Högström et al., 2014;but also Griesfeller et al., 2015). The correlation coefficients 32 between SiBCASA and ASCAT are best when the error estimate of the retrieved ASCAT soil 33 moisture is smaller than 10 -1 m 3 m -3 . Finally, the top most soil layer which contains ice is a 1 poor predictor of r. Where the first layer is frozen, the r's are indeed near 0, but all other grid 2 points have ice only much deeper than the 8 th soil layer, and there is no relation with r. This 3 essentially means that permafrost does not disturb the satellite signal in August in Siberia. 4 The characteristics of the four field sites are indicated by black marks in Fig. 12. This shows 5 that the performance at the Yakutsk and Elgeeii sites may be expected to be low, because of 6 the large LAI, low temperatures and relatively large soil moisture. At the Tver and Hyytiälä 7 sites, the expected performance is better, although the Tver site performs below average. We 8 can only guess what might explain this difference. The region around the Tver site is quite 9 heterogeneous, with a mixture of Spruce and deciduous forests and peat bogs, rivers and lake 10 Seliger. Perhaps the LAI is in reality larger than SiBCASA predicts, and the satellite retrieval 11 is hampered by surface water. 12 FIGURE 12 about here 13 In conclusion, (partial) ponding of the soil appears to be a good potential explanation of why 14 the poor performance of ASCAT soil moisture improves into the summer months in Boreal 15 Eurasia. The presence of dense leafs rather than aboveground biomass disturbs the satellite 16 signal. 17

Carbon effects 18
It has been shown in section 3.2 that assimilation of ASCAT soil moisture in SiBCASA has 19 an effect of 5 to 10 percent on GPP and TER, and of a few tens of percent on NEE, at the site 20 of Yakutsk, over the entire year. This represents the higher end of the range, since the effect 21 of assimilation on soil moisture and carbon fluxes was relatively large in Yakutsk. The reason 22 why Yakutsk is so sensitive is because the plant available water fraction is small there, so that 23 the drought sensitivity is large (Fig. 1a). Integrated over the entire region, assimilation causes 24 changes in the order of half the inter-annual variability, or 2% of the mean annual NEE. We 25 consider this quite large, given the fact that we only applied the assimilation to the top soil 26 moisture. However, the temporal correlation coefficients were quite low in large parts of the 27 region (Fig. 4). This implies that simulated and observed soil moisture are quite different. 28 Assimilation will thus have a large effect when the observational errors are small. A 29 comparison between observed and simulated NEE is made and discussed in the supplement 30 (Fig. S4). 31 The effect of changing soil moisture on GPP is largest in SiBCASA when the plant available 1 water fraction is smaller than 0.3. The area where this occurs is confined to the steppe zone in 2 South West Siberia in South European Russia, where it is dry and in the North East Siberian 3 forest zone, where water availability is limited by permafrost. If the drought stress function in 4 Fig. 1a would be defined more linearly, the effects of soil moisture would be spread more 5 evenly over the study domain. Note that it may not be realistic to prescribe identical water 6 stress formulations for all biome types, as SiBCASA does. Furthermore, Fig.e 10 shows that 7 the drought sensitivity in Fig. 1 only represents the potential drought sensitivity. The actual 8 sensitivity of GPP to change in soil moisture also depends on the temperature, radiation and 9 vapour pressure deficit (Fig. 10). This applies to TER in a similar way too. As a result, 10 changes in NEE are not linearly dependent on the change in soil moisture due to assimilation 11 of satellite observed soil moisture. Consequently, local effects may be much larger than 2% of 12 the mean annual NEE. Furthermore, Ohta et al. (2014) show that in reality, water logging at 13 high plant available water fractions may also reduce photosynthesis rates and affect the water 14 use efficiency. 15 16

Conclusions 17
The spatial and temporal correlation between SiBCASA soil moisture and ASCAT soil 18 moisture are considerable in the summer period and the steppe zone. However, ASCAT 19 derived soil moisture fails to detect the 8 major droughts observed in-situ at 4 sites during 7 20 years, while SiBCASA reproduces half of those droughts. At site-level, temporal correlations 21 between SiBCASA and in-situ observed soil moisture are larger than between SiBCASA and 22 ASCAT soil moisture. These facts suggest that SiBCASA soil moisture is more reliable than 23 ASCAT soil moisture at those 4 locations and that assimilation of ASCAT soil moisture does 24 not improve SiBCASA soil moisture. 25 The temporal correlation between SiBCASA and ASCAT soil moisture is best in the steppe 26 zone, and in a selection of forest locations where LAI is low, soil temperature is high, and soil 27 moisture is low (Fig. 12). Unfortunately, we do not have ground observations to proof 28 whether assimilation in such conditions would lead to improved soil moisture in SiBCASA. 29 There is evidence that assimilation of ASCAT soil moisture improves the match of SiBCASA 30 soil moisture with in-situ observations in spring time (Fig. 5 -8). However, these results 31 should be taken carefully, because ice and ponding occur often in the spring. Irrespective of 32 the question whether assimilation improves soil moisture in SiBCASA, assimilation of 1 ASCAT soil moisture causes considerable changes in GPP, TER and NEE. At individual 2 locations these changes may reach up to 5 to 10 % of annual GPP and TER, and tens of 3 percent of annual NEE, and integrated over the entire region, the changes cause changes in the 4 order of half the inter-annual variability in NEE or 2 % of annual NEE. 5 Ultimately, this study shows that assimilation of satellite observed soil moisture in vegetation 6 models potentially has large impacts on the simulated carbon fluxes, but that further research 7 is needed to clarify when, where and in which conditions assimilation leads to more reliable 8 soil moisture simulations. In the near-future important improvements in the quality and spatial 9 resolution of soil moisture are expected to be realised with the SMAP L-band instrument and 10 Sentinel-1. Additionally, the benefit of more advanced assimilation techniques, e.g. by 11 assimilating low-pass filtered satellite signals, may be investigated. Acknowledgements 20 The research on which this publication is based has been partly funded by the Dutch 21 Organisation for Scientific research (NWO) under grant number 864.08.012 (VIDI WP "A 22 multiple constraint data assimilation system for the carbon cycle"). JK and AV, and the in-situ 23 observations in Tver were funded by the Russian Science Foundation, Grant 14-27-00065. 24