The GRENE-TEA Model Intercomparison Project (GTMIP) stage 1 forcing dataset

T. Sueyoshi, K. Saito, S. Miyazaki, J. Mori, T. Ise, H. Arakida, R. Suzuki, A. Sato, Y. Iijima, H. Yabuki, H. Ikawa, T. Ohta, A. Kotani, T. Hajima, H. Sato, T. Yamazaki, and A. Sugimoto National Institute of Polar Research, Tachikawa, Japan Japan Agency for Marine-Earth Science and Technology, Yokohama, Japan Kyoto University, Field Science Education and Research Center, Kyoto, Japan RIKEN Advanced Institute for Computational Science, Kobe, Japan National Research Institute for Earth Science and Disaster Prevention, Snow and Ice Research Center, Nagaoka, Japan National Institute for Agro-Environmental Sciences, Tsukuba, Japan Nagoya University, Graduate School of Bioagricultural Sciences, Nagoya, Japan Tohoku University, Graduate School of Science, Sendai, Japan Hokkaido University, Faculty of Environmental Earth Science, Sapporo, Japan now at: Sonic Corporation, Tachikawa, Japan


Introduction
GTMIP (GRENE-TEA Model Intercomparison Project, https://ads.nipr.ac.jp/gtmip/ 15 gtmip.html,  is a model intercomparison study in the GRENE-TEA project (Green Network of Excellence -Terrestrial Ecosystem of the Arctic), a sub-project of the 2011-2015 Japanese-funded Arctic research project GRENE-Arctic. GRENE-TEA aims to understand on-going changes to the Circum-Arctic terrestrial ecosystem and their effects on climate. In order to address this goal, GTMIP 20 attempts to assess inter-model and inter-site (climatic/vegetative) uncertainties, as well as variations in Land Surface Models (LSMs) including both physical and ecosystem processes. The project consists of two stages: stage 1 includes one-dimensional (1-D) model simulations for selected sites, while stage 2 is focused on distributed global/regional model simulation runs for selected climate scenarios. This paper de-It is very difficult, however, to create a complete dataset for driving a model solely from observations, which are prone to data gaps due to instrument problems, human mistakes, or other reasons (Nakai et al., 2013;Kodama et al., 2007;Watanabe et al., 2000;Iwahana et al., 2014;Ohta et al., 2001Ohta et al., , 2008Ohta et al., , 2014Kotani et al., 2013;Lopez et al., 2007;Nakai et al., 2008;Sato et al., 2001). Despite efforts to maintain and improve our observational network, data gaps are unavoidable, as the availability of instrumentation is often limited by logistical or funding issues. Many of the Arctic sites are located in remote areas with limited access and power supply (see Fig. 1 for GRENE-TEA observation sites; cf. site distributions of other national/international Arctic research projects, such as PAGE21 (http://portal.inter-map.com/#mapID=49&groupID=308&z=1.0&up= 15 -310027.6&left=2001105.4; accessed on 20 April 2015) or map of stations used for the Climate Research Unit (CRU) data set (http://climate.unur.com/cru-station-records/ cru-stations-map.html; accessed on 20 April 2015).
On the other hand, reanalysis data are often used to drive LSMs, as they are always continuous without missing values and physically consistent with regard to all 20 necessary elements. Although there are several good reanalysis datasets available now (e.g. NCEP, ERA, JRA, MERRA), these products, which are essentially model outputs at large horizontal scales (typically on the order of 100 km), are no match for a direct comparison with observational data. Previous efforts of the modelling and empirical research community have produced several datasets available as inputs into 25 LSMs such as the Global Meteorological Forcing Dataset for land surface modelling (Sheffield et al., 2006) and WATCH (Weedon et al., 2010), or FLUXNET observational data (Vuichar and Papale, 2015) for debiasing of reanalysis data (ERA-Interim). After consideration of the general problem, our particular circumstances, and methodologies 706 ESSDD 8,2015 The GRENE-TEA Model Intercomparison Project (GTMIP) stage 1 forcing dataset Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | used in creating the previous products, GTMIP took the approach of preparing a harmonized dataset, which relies on reanalysis data to create temporally continuous data (basis for level 0 data, see Sect. 3) while utilizing observational data in customizing the level 0 data to fit to each site. We produced a dataset with high consistency relative to the observational data and without gaps (level 1 data).

5
In this paper, after summarizing information about the study sites and observational period in Sect. 2, we describe the specification of the data in Sect. 3, and the details of the procedure used to produce the two datasets in Sect. 4. The procedure is explained for each variable in the dataset, and differences between the sites, when present, are described. Further implications of this work are described in Sect. 5, and access to the 10 data is documented in Sect. 6.

Study sites and observational periods
Stage 1 of GTMIP produced a harmonized forcing dataset for six sites in Arctic and sub-Arctic regions, which have been, at least partially, studied as part of the GRENE-TEA project ( Fig. 1): Fairbanks in Alaska (Nakai et al., 2013), Tiksi (Kodama et al., 2007;15 Watanabe et al., 2000) and Chokurdakh (Iwahana et al., 2014) in northeast Siberia, Yakutsk in central East Siberia (Ohta et al., 2001(Ohta et al., , 2008(Ohta et al., , 2014Kotani et al., 2013;Lopez et al., 2007), Tura in central Siberia (Nakai et al., 2008), and Kevo in northern Europe (Sato et al., 2001). The longitudes and latitudes of the sites are summarized in Table 1, as well as the observational periods for data collection. 20 The sites lie in different geographical zones with different ecoclimatic characteristics. Fairbanks is located in the discontinuous permafrost zone with open black spruce forest, Tiksi and Chokurdakh are in the continuous permafrost zone, the former with tundra vegetation and the latter in the transition between tundra and taiga, Yakutsk and Tura are in the continuous permafrost zone with larch forest, and Kevo is in the 25 non-permafrost (seasonal frozen ground) zone with mixed forest. 8,2015 The GRENE-TEA Model Intercomparison Project (GTMIP) stage 1 forcing dataset These sites were selected from "super sites" among the GRENE-TEA observation sites, where automatic weather stations were installed that measure a number of meteorological variables, and enough site information about soil and vegetation conditions were available. Geographical coverage was also considered, to ensure that the sites were distributed over the entire Arctic, and that they covered a combination of tundra 5 and forest.

ESSDD
Although we are planning to expand this network of "harmonized dataset sites", we believe that these six initial sites, already covering typical geographical conditions in the Arctic, are a good starting point for a model intercomparison aiming to improve our understanding of physical and ecosystem land surface processes in the Arctic and 10 sub-Arctic regions.

Dataset specifications
The data created in this study cover a 35 year period, from 1 September 1979 to 31 December 2013, local time, for all sites. Temporal resolution was set to 30 min, considering the requirements of a majority of the physical land surface models. The 15 start date was set to 1 September, when there should be no snow cover, in order to minimize any potential bias for snow depth at the initial condition.
The GTMIP dataset consists of three sets of data, each of which has eight meteorological elements so that a contemporary land surface model can be driven with the dataset. The first set is called level 0 (Lv0), consisting of reanalysis data with minimum 20 correction, corresponding to the existing observation-based global dataset. The second set is called level 1 (Lv1), with additional corrections applied to Lv0 data, based on the GRENE-TEA site observations. The third set is a detrended dataset (DT), which essentially consists of Lv1 data with long-term temporal trends removed. Both Lv0 and Lv1 are forcing datasets. The primary dataset for the GTMIP activity is Lv1, while Lv0 25 was prepared to enable an evaluation of the effect of the observation-based corrections. The DT is intended for use with ecosystem models with long spin-up times. We 708 ESSDD 8,2015 The GRENE-TEA Model Intercomparison Project (GTMIP) stage 1 forcing dataset also provided satellite-based photosynthetically active radiation (fPAR) and leaf area index (LAI) data for models that need these data as prescribed conditions. A schematic diagram for the creation of the Lv0 and Lv1 data is shown in Fig. 2

Methods of forcing data production
This section provides documentation for (1) selection of reanalysis data, (2) construc-10 tion procedure for Lv0 data, (3) construction procedure for Lv1 data, and (4) construction procedure for additional data (detrended data, and LAI data). The outline of the procedures for Lv0 and Lv1 is given in Table 2. Procedures are explained for each variable in Sect. 4.3 and 4.4. When data were handled differently for different sites, it was also noted.

Selection of reanalysis data
As the first step of the procedure, we selected the reanalysis dataset used as the base data. A detailed performance comparison was carried out for the following four datasets: Japanese Reanalysis JRA-55, The European centre for medium-range weather forecasts ReAnalysis ERA-Interim (Dee et al., 2011), National Centers for 20 Environmental Prediction (NCEP)/National Center for Atmospheric Research (NCAR); NCEP/NCAR, NCEP-Department of Energy (DOE). Climate Research Unit (CRU, East Anglia University) data for air temperature (Harris et al., 2014) and the Global Precipitation Climatology Project (GPCP) dataset for precipitation (Adler et al., 2003), based on weather station observations and/or satellite data, were used as a reference. Spe- 8,2015 The GRENE-TEA Model Intercomparison Project (GTMIP) stage 1 forcing dataset cial emphasis was placed on the reproducibility of air temperature and precipitation, as these two variables are known to have a large impact on the performance of LSMs. As a result of the comparison of monthly values, ERA-Interim was chosen as the base data set for GTMIP, reproducing these two variables reasonably with minimum root mean square errors (RMSEs) with respect to the CRU (air temperature) and GPCP 5 (precipitation) datasets (Table 3). Among the ERA-Interim variables, the following parameters were used in constructing the GTMIP forcing data: 10 m zonal (U) wind component, 10 m meridional (V ) wind component, 2 m dew point temperature, 2 m temperature (t 2 m ), surface pressure, surface solar radiation downward, surface thermal radiation downward, and total precipitation. The temporal resolution of the used variables is 6 h, and stamped at 00:00 Z, 00:06 Z, 00:12 Z, and 00:18 Z in coordinated universal time (UTC).

General
Construction of Lv0 data (Saito et al., 2014a) from reanalysis data contains the fol-15 lowing steps: (1) choose grid and time period (in this case September 1979 to December 2013) from the reanalysis data, (2) apply a common correction for all sites for temperature and precipitation, and (3) determine 30 min values using the 6 hourly data. Providing a template dataset for modellers to prepare their own read-in modules in advance is the primary purpose of this dataset, but as mentioned before, it also serves 20 to evaluate the effects of the site-fitting correction by comparing the performance of LSMs using these two datasets. Minimum corrections of temperature and precipitation using CRU and GPCP are applied so that problems with using unrealistic values can be avoided. Differences between sites were not considered for this correction.
For the selection of the data grid, the "nearest neighbour" principle was used. Table 4 shows the coordinates of grid centres of the ERA-Interim, CRU, and GPCP datasets for four sites. Temporal interpolation is a requisite as the temporal resolution of the target dataset is 30 min, while that of the ERA-Interim data is 6 h. A single strategy (e.g. linear interpolation or equal distribution) does not always work, depending on the characteristics of variables and sites. The process of temporal adjustment is described for each variable in the next sections.

2 m air temperature
For Lv0 air temperature (T Lv0 ), monthly mean values derived from 6 hourly ERA-Interim air temperature [K] was adjusted using the CRU TS3.22 dataset of the month, so that the monthly mean of Lv0 data coincides with the CRU temperature data, T month CRU [ • C]; (1)

10
Each value was set to the corresponding UTC time, and then linearly interpolated with the next 6 hourly data to produce data with a 30 min resolution (T

Specific humidity
Specific humidity for the Lv0 data (q Lv0 ) was determined by a two-step procedure. First, relative humidity was calculated using the 6 hourly ERA-Interim dew points (DewT 6 h ERA ) 15 and air temperatures. Second, 30 min specific humidity was calculated from the previously obtained 6 hourly relative humidity data and the 2 m air temperature calculated in Sect. 4.2.2. 20 where q s (T ) denotes saturation specific humidity at temperature T . 711 8,2015 The GRENE-TEA Model Intercomparison Project (GTMIP) stage 1 forcing dataset

10 m wind speed and direction
Six-hourly wind speed and direction, calculated from zonal and meridional wind of ERA-Interim, were set as the Lv0 data at the corresponding UTC time, and repeated 11 times to create 30 min interval data in the 6 h range.

Total precipitation 5
Six-hourly precipitation amounts (P Lv0 ) were calculated from 6 hourly ERA-Interim precipitation, and a correction was applied in such a way that the monthly amount was proportional to the GPCP precipitation data: To determine how to distribute the 6 h precipitation data among the 30 min time slots, 10 we conducted a preliminary analysis using the following three strategies and evaluated their influence on the simulated energy and water budget keeping other factors and variables fixed: (a) assign the entire amount to the first time slot (i.e. the entire 6 h precipitation amount falls in the first 30 min of the 6 h range), (b) assign the precipitation amounts in ratios of eight, four, two, and one to the first four time slots (i.e. the first two 15 hours), and (c) assign equal amounts to the 12 time slots in the range. The results showed a distorted energy and water cycle for (c), while results were reasonable for the first two cases (not shown). Based on this outcome, we used strategy (a) and assigned the entire precipitation amount to the first time slot of every 6 h range. 20 Shortwave radiation (SW), calculated as theoretical values for latitude, longitude, day, and time for a particular site and day at 30 min intervals, was corrected using daily Comparison with observed values shows that thus calculated SW coincides well with observations in terms of daily averages, implying that the total incident solar energy is well represented. Figure 3 shows an example of an annual time series for daily 5 maximum, average, and minimum of 30 min SW for Tiksi.

Downward longwave radiation
ERA-Interim 6 hourly longwave radiation data were set as the Lv0 data at the corresponding UTC time, and repeated 11 times to create 30 min interval data for each 6 h range.

Surface pressure
ERA-Interim 6 hourly pressure data were set as the Lv0 data at the corresponding UTC time, and repeated 11 times to create 30 min interval data.

15
For the Lv1 dataset (Saito et al., 2014b), we applied a site-specific correction using observational data. The principle of Lv1 data construction is to derive rules to modify base data (ERA-Interim) to make them as similar to the observed values as possible. The formulations for statistical correction were established through comparison between observations and climate data (that is, ERA-Interim, CRU, and GPCP) for the  Fig. 2). Observational periods are summarized in Table 1, and visualized by grey bars in the middle panel, below "GRENE-TEA site observation", of Fig. 2. The Lv0 data were readily available with no or minor modifications for some variables, whereas site-specific treatment and adjustment were needed for other variables. Comparisons between observed and pro-5 duced data, and the difference between Lv0 and Lv1 data are shown for selected sites and variables in Figs. 4-6. The overall quality of fit for all variables is shown in Fig. 7, which shows gradient a of the fitted line of daily mean observations y against Lv0, x (y = ax + b), and Pearson's correlation coefficient r for x and y. The more a and r approach unity, the better the fit. Note that the order of independent (abscissa, x) and dependent (ordinate, y) variables of the fitting line (exemplified in Figs. 3b and 4c-f) is unconventional because a choice was made to show the necessary modification or corrections made to Lv0 to construct site-corrected Lv1.

2 m air temperature
Comparison with observations shows that the Lv0 data agree well with daily averages, 15 but have a smaller daily temperature range than the observed data, mainly because Lv0 data were produced by interpolation of four points per day. Therefore, we had to employ two-step evaluations for the following targets (i.e. daily temperature range correction and monthly average values correction) to produce Lv1 temperature data. The climate data (CD: CRU and/or ERA-Interim) that were used for the diurnal correction 20 differed from site to site, and were chosen based on the reproducibility of the observed data (not shown). The height of near-surface air temperature measurements at each site is summarized in Table 1. We chose observations at the height closest to 2 m, but did not convert the observation values for height because the reference height for air temperature is different among models. 25 First, the daily temperature range (dtr) was corrected for the ERA-Interim 6 hourly data, employing the CRU diurnal monthly temperature range for Fairbanks (dtr CD = dtr month CRU ) but daily ERA-Interim data (namely, 2 m temperature maximum and minimum; 714 We did not adjust Tiksi air temperature with CRU data, because the CRU values showed poorer reproducibility of the Tiksi site observations than the ERA-Interim monthly mean values (see Table 3), mostly due to the sparse network of stations in the region. However, the ERA-Interim values at the closest grid point to Tiksi show modest continentality (warmer summers and cooler winters) than the observed values at the coastal site. Therefore, a seasonal correction for Lv1 data (L1 = T Finally, these 6 hourly values (L1) were linearly interpolated with the next 6 hourly corrected data to produce data at 30 min resolution for all sites. Improvement from fitting the produced values to the observational data ( Fig. 7a- Improvement at Tiksi is a reflection of improvement in air temperatures as illustrated in Fig. 7d, which shows the daily mean for specific humidity.

10 m wind speed and direction
We determined 6 hourly wind speed and direction data from U wind and V wind of 10 the ERA-Interim dataset and converted the data to the corresponding UTC time. We defined these data as L0. Then, we applied a statistical correction based on the relationship between the observed values and L0 at Kevo and Tiksi (see Fig. 4c). The observational periods used for the statistical correction were 1999 to 2001 and 1997 to 2012 for Kevo and Tiksi, respectively. The equation for correction of the Lv1 data (L1) 15 at each site can be written as follows: No correction of Lv0 wind speed was applied to Fairbanks and Yakutsk for the following reasons (Fig. 7k-m): Yakutsk already showed a good match between observed 20 and L0 data when outliers and data collected during equipment malfunctions were removed, while Fairbanks showed such a poor match between the two datasets that we decided that no statistic correction would have been appropriate.  Table 1. We chose the observations at the height closest to 10 m above ground (Kevo, Tiksi) or canopy (Fairbanks, Yakutsk), but did not convert the observation values for height because the reference height is different among different models.

5
Precipitation is the most challenging part of dataset construction because it can have the largest impact on the simulated results. Reanalysis data (in the case of this study, ERA-Interim), which are virtually a model output, generally contain large uncertainties in precipitation and produce highly averaged and smoothed values because of a sixhour output interval. On the other hand, site observation data are also not always 10 reliable, mainly due to instrumental problems. Analysed precipitation products (in the case of this study, GPCP) have difficulties in high-latitude regions due to a scarcity of available data and poor satellite coverage. With these shortcomings of the data sources in mind, we created site-fitted precipitation data by evaluating the reproducibility of precipitation amounts and variations on a monthly and annual basis. 15 For Fairbanks, we decided to use the Lv0 data as Lv1 data (L1) rather than attempting an unreliable correction, mostly because the two source datasets did not show any systematic relationship or good correlation with monthly observations, which were made at the international airport, located about 50 km from the site. For Kevo, comparison between the annual total of observed precipitation and that of the ERA-Interim 20 dataset showed a good inter-annual correlation, albeit with a quasi-constant underestimation. Since ERA-Interim data also showed a good correlation with intra-annual (monthly) variability, we used the ERA-Interim as the basic dataset L0 , that is, without any monthly correction by GPCP. The resulting correction equation is shown below. 25 For Tiksi, GPCP provided better monthly variations while ERA-Interim data showed better annual total and inter-annual variations. Therefore, we applied a correction to Lv0 717 using the ratio of annual precipitation of the ERA-Interim data to the annual precipitation of GPCP (Fig. 5a). In contrast, as the ERA-Interim 6 hourly total precipitation amounts agreed well with the observational data both on a monthly and annual scale for Yakutsk, they were directly used as the 6 hourly values without GPCP correction (Fig. 5b). Finally, the 6 hourly precipitation amounts were assigned to the first time slot of every 5 six-hour range to produce a dataset with 30 min intervals. Note that the correlation of daily values is not particularly high (Fig. 7f) relative to that of other variables shown in Fig. 7. This is a natural result since fitting is done only on a monthly and annual basis. The overall improvement is, however, already visible as Lv1 locates closer to the 1 : 1 gradient line (i.e. a = 1) than Lv0.

Downward shortwave radiation
As stated in Sect. 4.2.6, downward shortwave radiation is computed from the theoretical instantaneous radiation value decreased by daily average cloudiness, which corresponds well to the daily mean incident energy at the Earth's surface. Some discrepancies were found between Lv0 and the observational data with regard to daily 15 maximum (Fig. 7g) and/or minimum (Fig. 7i). Possible causes are short-time cloud cover or absence of clouds, shading by forested areas (as is the case for the minimum SW in Kevo), or equipment malfunction. As it is difficult to generalize these situations in a statistical formula, no correction of Lv0 was applied for downward shortwave radiation data for all four sites. Figure 7i does not show data for Fairbanks and Yakutsk since the 20 two stations lie south of the Arctic Circle, and hence the daily minimum is always zero.

Downward longwave radiation
Comparison between observed values and the ERA-interim output (i.e. Lv0 data, L0) revealed that the difference between the two datasets can largely be accounted for by emission differences of the air layer due to altitude differences between the observation 25 sites and the closest ERA-Interim grid points (see Fig. 6a Figure 7j shows a marked improvement for Tiksi. This improvement is mostly manifested by the location's gradient (0.85), which reflects a seasonal difference due to the continental character of the reanalysis value at the grid point closest to Tiksi against more coastal characteristics of the Tiksi observational site. Since diurnal cycles in downward longwave and surface pressure are not as prominent as in downward short-10 wave radiation or air temperature, only daily means are shown in Fig. 7.

Surface pressure
Similar to downward longwave radiation, differences in surface pressure between the observed and reanalysis data (i.e. Lv0 data, L0) are largely correlated with altitude differences (see Fig. 6b as an example for Fairbanks). The resulting correction equations 15 are as follows: Since changes in the intercept of the fitting lines in Eq. (11) do not change the gradient and correlation coefficient, Fig. 7e does not show any difference between Lv0 and Lv1.

Additional data
In addition to the two model driving datasets described in the previous section, we have prepared the following two datasets for standardization and convenience for model integration and spin-up: photosynthetically active radiation (fPAR) and leaf area index (LAI) datasets (Saito et al., 2014c), and a 20 year detrended meteorological driving 5 dataset . The former dataset was derived from Moderate Resolution Imaging Spectroradiometer (MODIS) satellite data (MOD15A2, MYD15A2) according to algorithms developed and validated by Sasai et al. (2011). Monthly averaged time series were processed for each of the six GRENE-TEA sites from the closest grid point values.
The latter detrended driving dataset, produced be removing monthly long-term trends from the Lv1 data for the period of 1980-1999, was provided for spin-up to include initial soil carbon conditions comparable to other biogeochemical models without being affected by warming trends . 15 Concepts and principles of a correction method for creating a forcing dataset from climate product data (e.g. ERA-Interim, CRU, and GPCP) and site observations were established through a comparison of the two data sources over the observational period as described in this article. The set of observation-based, site-specific data thus produced can be expected to produce a more realistic time series than the climate 20 product data themselves, and will allow a more meaningful comparison to observed target variables at a given site (e.g. evapotranspiration, snow pack, ground temperature and soil moisture, and gross primary products).

Discussion and outlook
Once correction rules (formulae) are defined, a time series can be extrapolated to a period of no observation. Not only can these datasets (Lv0 and Lv1, in this case) be 25 used as common driving data for model intercomparison projects to examine inter-site 720 ESSDD 8,2015 The GRENE-TEA Model Intercomparison Project (GTMIP) stage 1 forcing dataset and inter-model variations with regard to performance and uncertainty, they are also useful for investigating long-term ecoclimatic variations by analysing the time series per se that incorporates the knowledge of on-going observations, and by examining and testing the capability of reproduction and hindcast through model simulations driven by the time series. Naturally, the available time period is still limited by the time covered by 5 the product-based data, and the efficacy of the rules used to create the product-based data should be thoroughly tested when these datasets are applied. Another target of this MIP activity was to facilitate collaboration between modelling and field scientists. Such collaboration can be difficult and is often insufficient due to discrepancies between science targets and approaches, terminology, and different 10 prioritization during data acquisition among the two groups (for example, field scientists tend to seek for unusual data in their area of investigation, while modellers often need typical data for an area; cf. Saito et al., 2014d).
Considering the aforementioned discrepancies, the GTMIP study was conducted by a team of modelling and field scientists, which was especially important when the qual- 15 ity of observed data used for Lv1 creation was questioned (e.g. outliers, missing values, representativeness of the site and nearby station data). Similarly, when conducting model performance checks using observational data, which is an important and unavoidable process for model development, field scientists joined together to discuss issues with regard to modelled and observed data, capability and limits of models de-20 pending on the extent and complexity of the implemented processes, often concluding that a precise observation does not necessary coincide with a "correct" model output.
The detailed protocol of GTMIP stage 1 (site simulations) is described in our companion paper . 25 All data presented in this paper, as well as an overall outline of the GRENE-TEA Model Intercomparison Project (GTMIP), are available from the Arctic Data archive System ESSDD 8,2015 The GRENE-TEA Model Intercomparison Project (GTMIP) stage 1 forcing dataset Discussion Paper | Discussion Paper | Discussion Paper | Discussion Paper | Kodama, Y., Sato, N., Yabuki, H., Ishii, Y., Nomura, M., and Ohata, T.: Wind direction dependency of water and energy fluxes and synoptic conditions over a tundra near Tiksi, Siberia, Hydrol. Process., 21, 2028-2037, 2007