Uncertainty analysis of eddy covariance CO 2 flux measurements for different EC tower distances using an extended two-tower approach

The use of eddy covariance (EC) CO2 flux measurements in data assimilation and other applications requires an estimate of the random uncertainty. In previous studies, the (classical) two-tower approach has yielded robust uncertainty estimates, but care must be taken to meet the often competing requirements of statistical independence (nonoverlapping footprints) and ecosystem homogeneity when choosing an appropriate tower distance. The role of the tower distance was investigated with help of a roving station separated between 8 m and 34 km from a permanent EC grassland station. Random uncertainty was estimated for five separation distances with the classical two-tower approach and an extended approach which removed systematic differences of CO2 fluxes measured at two EC towers. This analysis was made for a data set where (i) only similar weather conditions at the two sites were included, and (ii) an unfiltered one. The extended approach, applied to weather-filtered data for separation distances of 95 and 173 m gave uncertainty estimates in best correspondence with an independent reference method. The introduced correction for systematic flux differences considerably reduced the overestimation of the two-tower based uncertainty of net CO2 flux measurements and decreased the sensitivity of results to tower distance. We therefore conclude that corrections for systematic flux differences (e.g., caused by different environmental conditions at both EC towers) can help to apply the two-tower approach to more site pairs with less ideal conditions.


Introduction 28
The net ecosystem exchange of CO 2 between the land surface and the atmosphere (NEE) can 29 be determined with the eddy covariance (EC) method. NEE is positive if the amount of CO 2 30 released to the atmosphere via respiration is higher than the amount of CO 2 assimilated 31 during photosynthesis. In contrast, negative NEE values denote a higher CO 2 uptake and a net 32 flux from the atmosphere into the ecosystem. During night-time, NEE is mainly a function of 33 respiration and therefore positive fluxes predominate, whereas during (summer) daytime 34 negative NEE values predominate because more CO 2 is assimilated than respired. Eddy 35 covariance CO 2 flux measurements are commonly used to analyze the interactions between 36 terrestrial ecosystems and the atmosphere which is crucial for the understanding of climate-37 ecosystem feedbacks. In this regard reliable EC data with appropriate uncertainty estimates 38 are crucial for many application fields, such as the evaluation and improvement of land 39 surface models (e.g. Braswell  When using the term 'uncertainty', we here focus on the random error following the 41 definition in Dragoni et al. (2007). It differs from the systematic error in that it is 42 unpredictable and impossible to correct (but can be quantified). Uncertainty doesn't 43 accumulate linearly but "averages out" and can be characterized by probability distribution 44 functions (Richardson et al., 2012). Systematic errors are considered to remain constant for a 45 longer time period (> several hours). Ideally they can be corrected, but in case of EC 46 measurements this is still limited by either our understanding of various error sources or 47 insufficient background data. Systematic errors arise not only from instrumental calibration 48 and data processing deficits, but also from unmet underlying assumptions about the 49 meteorological conditions (Richardson et al., 2012). A main assumption is that turbulence is 50 always well developed in the lowest atmospheric boundary layer and responsible for the mass 51 sites do not have a nearby second EC tower to provide nearly identical environmental 76 conditions. Therefore, Richardson et al. (2006) introduced the "one-tower" or "24-h 77 differencing" method which is based on the two-tower approach. The main difference is that 78 the uncertainty estimate is based on differences between fluxes measured on subsequent days 79 if environmental conditions were similar on both days. Because most often environmental 80 conditions are not the same on two subsequent days (Liu et al., 2006), the applicability of this 81 method suffers from a lack of data and the random error is overestimated (Dragoni et al., 82 2007  error. The method is based on the assumption that the model error is negligible, which is 86 however a very questionable assumption. Alternatively, if the high-frequency raw-data of an 87 EC tower are available, uncertainty can be estimated directly from their statistical properties 88 (Billesbach, 2011). Finkelstein and Sims (2001) introduced an operational quantification of 89 the instrumental noise and the stochastic error by calculating the auto-and cross-covariances 90 of the measured fluxes. This method was implemented into a standard EC data processing 91 scheme by Mauder et al. (2013). The advantage is that a second tower or the utilization of 92 additional tools such as a simple model to estimate the EC measurement uncertainty is no 93 longer required. However, many data users do not have access to the raw-data but to 94 processed EC data only. Moreover, a large amount of solid metadata about the setup of the 95 EC measurement devices is required (but often not provided at second hand) to obtain 96 reliable raw-data based uncertainty estimates adequately. Therefore a two-tower based 97 approach has still a large group of users. In particular with regard to pairs of nearby towers 98 from local clusters which play an increasing role in the monitoring strategies of e.g. ICOS 99 and NEON, and have already been employed in case studies (e.g. Ammann et al., 2007). 100 Important advantages of the two-tower approach are (1) its simplicity and user friendliness, 101 (2) its usability for relatively short non gap-filled time series of several months, and (3) the 102 independence of a model. 103 The classical two-tower approach (Hollinger et al., 2004;Hollinger and Richardson, 2005; 104 Richardson et al., 2006) is based on the assumption that environmental conditions for both 105 EC towers are identical and flux footprints should not overlap to guarantee statistical 106 independence. Hollinger and Richardson (2005)  Given the fact that site specific, adequate uncertainty estimates for eddy covariance data are 128 very important but still often neglected due to a lack of resources, we are aiming to advance 129 the two-tower approach so that it can also be applied if environmental conditions at both eddy 130 covariance towers are not very similar. 131 The main objectives of this study were (1) to analyze the effect of the EC tower distance on 132 the two-tower based CO 2 flux measurement uncertainty estimate and (2) to extend the two-133 tower approach with a simple correction term that removes systematic differences in CO 2 134 fluxes measured at the two sites. This extension follows the idea of the extended two-tower 135 approach for the uncertainty estimation of energy fluxes presented in Kessomkiat et al. 136 (2013). The correction step is important for providing a more reliable random error estimate. 137 In correspondence with these objectives we analyzed the following questions: What is an 138 appropriate EC tower distance to get a reliable two-tower based uncertainty estimate? Can the 139 random error be quantified in reasonable manner with the extended two-tower approach, even 140 though environmental conditions at both EC towers are clearly not identical? The total 141 random error estimated with the raw-data based method (Mauder et al., 2013) was used as a 142 reference to evaluate our extended two-tower approach based results. 143

Test sites and EC Tower setup 144
The Rollesbroich test site is an extensively used grassland site, located in the Eifel region of 145 western Germany (Fig.1) Sistig" 20.5 km north-east of Rollesbroich is another grassland site with similar 158 environmental conditions as Rollesbroich. The vegetation in Kall-Sistig is extensively 159 managed C3 grass, the same as for Rollesbroich. However, the average plant height measured 160 between Aug. 14 th and Oct.

Eq. 1
Based on Eq.1 we calculated the two-tower based uncertainty estimates using the NEE 1 data 217 measured at the permanent EC tower in Rollesbroich (EC1) and the NEE 2 data of a second 218 tower which was either the roving station (EC2) or -in case of the 34 km EC tower distance 219 -another permanent EC tower (EC3, Tab.1). 220 For comparison, the measurement uncertainty σ( ) was calculated separately for each EC conditions and non-overlapping footprints, we applied the classical approach for all EC tower 268 distances, even if these basic assumptions were not fulfilled, to allow for a comparison of the 269 results before and after the usage of the weather-filter and the sfd-correction (extended two-270 tower approach). 271

Correction for systematic flux differences (sfd-correction) 272
Different environmental conditions and other factors such as instrumental calibration errors 273 can cause systematic flux differences between two towers. Because these flux differences are 274 not inherent to the actual random error of the measured NEE at one EC tower station they 275 lead to an overestimation of the two-tower approach based uncertainty. Therefore, we 276 extended the classical two-tower approach with a simple correction step for systematic flux 277 differences (sfd-correction). The reason why systematic flux differences can statistically be 278 separated quite easily from random differences of the EC flux measurements is their 279 fundamentally different behavior in time: random differences fluctuate highly in time 280 whereas systematic differences tend to be constant over time or vary slowly. The sfd-281 correction introduced is similar to the second correction step in Kessomkiat  for averaging in that particular window. Due to the frequent occurrence of gaps in the data 294 series the amount of available NEE corr values considerably decreased by applying stricter 295 criteria like 70% or 90% data availability (Tab. A2). We assume a 12 hour averaging period 296 to be long enough to exclude most of the random error part but short enough to consider daily 297 changes of systematic flux differences. For a six hour interval for instance the uncertainty of 298 the mean NEE is usually higher. For larger window sizes (24 or 48 hours) further analysis 299 was hampered by too many data gaps, i.e. the 50% criterion was hardly ever fulfilled and not 300 enough averages remained to allow for the two-tower based uncertainty estimation (Tab. A2). The final sfd-corrected NEE1 corr values for EC1 and NEE2 corr values for EC2 should not be 306 understood as corrected NEE flux data. They were used only to enhance the two-tower based 307 uncertainty estimation in a way that systematic flux differences which cause an 308 overestimation of the uncertainty are filtered out. Moreover, systematic flux differences at 309 two EC towers are not to be confused with systematic errors, which are independent of the 310 uncertainty estimation method and optimally corrected before the random error is estimated. 311

Filter for weather conditions 312
For larger distances of two EC towers, such as the 20.5 km and 34 km distance in this study, 313 different weather conditions can cause differences of the measured fluxes in addition to the 314 different land surface properties. Some weather variables (e.g. temperature) are following a 315 clear diurnal and annual course and differences in e.g. temperature at two EC towers are 316 therefore relatively constant. This is expected to cause rather systematic differences in the 317 measured NEE which can be captured with the sfd-correction. However, other variables such 318 as wind speed or incoming short wave radiation are spatially and temporally much more 319 variable, for example related to single wind gusts or cloud movement. Differences in the 320 measured fluxes at two EC towers caused by those spatial-temporally highly variable weather 321 variables cannot be captured well with the sfd-correction term due to this "random character". 322 However, a weather filter can account for this because it compares the differences in weather 323 variables at each single time step. Therefore a filter for similar weather conditions was 324 applied in addition to the sfd-correction following Hill et al. (2012) and Richardson et al. 325 (2006) to only include half hourly NEE data, if the weather conditions at the second EC tower 326 are similar to those at the permanent EC1 tower location in Rollesbroich. Following the 327 definition in Richardson et al. (2006), similar weather conditions were defined by a 328 temperature difference < 3°C; wind speed difference < 1 m/s and difference in PPFD < 75 329 μmol m -2 s -1 . The weather-filter was applied before the (classical) uncertainty estimation and 330 the sfd-correction. As shown e.g. in Tsubo and Walker (2005), the incoming short wave 331 radiation (or solar irradiance SI) and the photosynthetically active radiation (PAR) are 332 linearly correlated. Accordingly SI and PPFD measured at the EC1 station in Rollesbroich 333 were also linearly correlated. Because direct PPFD measurements were not available for all 334 measurement periods, we derived a linear regression equation on the basis of all SI and PPFD 335 data for the permanent EC tower station (EC1). Using this equation, missing PPFD values 336 were estimated if only SI but no PPFD data were available at a certain time step. 337

Footprint analysis 338
The footprint analysis was applied to quantify the percentage footprint overlap of the two EC-339 stations during the measurement periods. This information was not used to filter the data but 340 to allow for a better understanding of the mean uncertainty estimates for the different 341 scenarios. Using the analytical model of Kormann and Meixner (2001)  observations. Generally, σ was considerably lower than σ ℎ . The total raw-data based 367 random error σ [μmol m -2 s -1 ] was calculated by adding σ and σ ℎ "in quadrature" 368 (σ = √ σ ℎ 2 + σ 2 ) according to Aubinet Fig.3 show the linear regressions of the random error σ(δ) (also referred to as 384 "standard error" or "uncertainty") as function of the NEE magnitude according to the 385 classical two-tower approach for the different EC tower distances without weather-filter 386 ( Fig.2) and with weather-filter (Fig.3). The dashed linear regression lines denote that the 387 linear correlation between σ(δ) and NEE is weak (p > 0.1), which is in particular true for the As illustrated in Tab.2, the mean NEE uncertainty estimate based on the classical two-tower 407 approach increased as a function of EC tower distance. However, without applying the 408 weather-filter, the mean uncertainty σ(δ) was nearly identical for the two largest distances 409 (20.5 km and 34 km), although e.g. the land cover and management in Merzenhausen (EC3 410 tower at 34 km separation) were different from the Rollesbroich site. As a result of the 411 weather-filtering, the mean uncertainty was less overestimated for the distances 173m and 412 20.5 km. However, for the 95 m and 34 km distance, the overestimation of the uncertainty 413 estimate increased by the weather-filtering (Tab.2). This implies that for the classical two-414 tower approach (without sfd-correction) weather-filtering did not clearly reduce the 415 overestimation of the uncertainty for largest EC tower distances (20.5 km and 34 km) where 416 weather-filtering is expected to be particularly relevant. 417 Comparing the mean uncertainty estimates of the classical two-tower approach with the 418 reference random error estimates σ cov , indicates that both with and without weather filter the 419 uncertainties were overestimated (Tab.2), for all EC tower differences. This could be 420 expected for the large distances, because basic assumptions for the application of the classical 421 two-tower approach are violated for these large distances. But results illustrate that even for 422 short EC tower distances NEE uncertainty estimated with the classical two-tower approach is 423 larger than the raw-data based estimates (Tab.2). 424

Extended two-tower approach 425
The scatter plots in Fig.4 illustrate the effect the sfd-correction (Eq.2) had on the difference 426 of the NEE data simultaneously measured at both EC towers (NEE -EC1-and NEE -EC2-). The 427 sfd-correction reduced the bias and scattering, because systematic differences of the 428 measured fluxes, e.g. induced by different environmental conditions, were removed. As 429 expected, the effect of the sfd-correction was considerably higher for the larger EC tower 430 distances because environmental conditions are also expected to differ more if the distance of 431 two locations is larger. For the 8 m EC tower distance for instance, the effect of the sfd-432 correction is very minor because footprints are often nearly overlapping. However, for the EC 433 tower distances >= 173 m, the bias and scattering of NEE -EC1-and NEE -EC2-was considerably 434 reduced by the sfd-correction. 435 A comparison of Fig.2 and Fig.5  Additional application of the weather-filter (Fig.6) on the sfd-corrected NEE corr data reduced 447 the mean uncertainty estimate σ(δ) corr by 23.3% and 2.9% for the 20.5 km and the 34 km EC 448 tower distance and reduced Δσ cov by 57.7% and 7.7%. The effect of the weather-filter on the 449 uncertainty estimates of the shorter EC tower distances was very minor (Tab.2). The 450 uncertainty estimates σ(δ) corr,f determined with the extended two-tower approach agree best 451 with the independent reference values σ cov for the EC tower distances 95m and 173 m, 452 suggesting that those distances were most suitable for the application of the extended two-453 tower approach. 454

Discussion 455
The results show that the two-tower based uncertainty estimates (both classical and extended 456 two-tower approach) were smallest for the 8 m distance. This can be explained with the 457 results of the footprint analysis: While the average percentage footprint overlap is 13% 458 (normalized 19%) for the 95 m EC tower distance and only 4% (7%) for the 173m EC tower 459 distance, it is 68% (80%) for the 8 m EC tower distance. The stronger overlap of the 8 m 460 distance footprint areas is associated with a more frequent sampling of the same eddies. As a 461 consequence, part of the random error was not captured with the two-tower approach. If EC 462 towers are located very close to each other (< 10 m) and the footprint overlap approaches 463 100%, only instrumental errors and stochasticity related to sampling of small eddies will be 464 captured with the two-tower based uncertainty estimate. Because the EC measurements are 465 statistically not independent if the footprints are overlapping, the classical EC tower method 466 is not expected to give reliable uncertainty estimates for very short EC tower distances 467 (Hollinger et al., 2004;Hollinger and Richardson, 2005). However, without applying the sfd-468 correction, the mean uncertainty estimate σ(δ) was higher than the raw-data based reference 469 value σ cov which includes both the instrumental noise σ and the stochastic error σ ℎ . 470 The raw-data based σ itself was only 0.04 μmol m -2 s -1 of 0.64 μmol m -2 s -1 for the 471 dataset of the 8 m EC tower distance. The mean uncertainty value derived with the extended 472 two-tower approach σ(δ) corr,f for the same dataset was lower than σ(δ) but still considerably 473 higher than σ , suggesting that even at 8 m EC tower distance instrumentation errors were 474 only a minor part of the two-tower based uncertainty estimate. Taylor, 1994; Orchard and Cook, 1983) as well as the role of grassland management (e.g. 487 Allard et al., 2007). Results indicate that an overestimation of the two-tower based 488 uncertainty caused by different land surface properties in the footprint area of both EC towers 489 can be successfully filtered out by the extended approach. It should be noted that a shorter 490 moving average interval of the sfd-correction term (e.g. 6 hours instead of the applied 12 491 hours window; Tab.A2), results in slightly lower uncertainty estimates compared to the 492 reference. This can be explained by a possible "over-correction" of the NEE data related to a 493 too short moving average interval for calculating the sfd-correction term. It needs to be 494 emphasized that the estimated mean NEE values of the moving average intervals are 495 associated with uncertainty. As mentioned, the moving average interval should be long 496 enough to exclude random differences of the simultaneously measured fluxes but short 497 enough to limit the impact of non-stationary conditions. However, the 12hr running mean 498 NEE1 and NEE2 values ( 12 ) as well as the respective means of NEE1 and NEE2 499 ( 2 _12 ) used to calculate NEE corr (Eq.2) are uncertain because they still contain the 500 random error part which cannot be corrected or filtered out. This uncertainty in the mean is 501 expected to be higher for a shorter averaging interval such as 6 hours. Therefore, completely 502 correcting the difference in mean NEE slightly overcorrects systematic differences in NEE. In 503 general results were not very sensitive to different moving average sizes of the sfd-correction 504 term and data coverage percentages defined for this interval (Tab.A3). 505 It is expected that systematic differences in measured NEE caused by spatially variable land 506 surface properties are stronger during the night than during the day since they affect 507 respiration more directly than photosynthesis (see e.g. Oren et al., 2006). Moreover, during 508 night-time and/or winter (positive NEE), some conditions associated with lower EC data 509 quality such as low turbulence, strong stability, and liquid water in the gas analyzer path 510 prevail more often than in summer and/or daytime (negative NEE). The less severe cases of 511 such conditions are not always completely eliminated by the quality control. In time series of 512 eddy-covariance fluxes this typically shows up as implausible fluctuations of the flux during 513 calm nights. This is reflected by plots of NEE flux magnitude versus uncertainty (Fig.2-3; 514 Fig.5-6) showing higher uncertainties for positive compared to negative NEE data which 515 agrees with previous findings (e.g. Richardson et al., 2006). 516 At very large EC tower distances (20.5 km, 34 km) footprints were not overlapping and the 517 environmental conditions were considerably different; in particular for the EC tower setup 518 Rollesbroich/Merzenhausen with different land use (grassland/crop) and climate conditions. 519 For those distances, the relative difference Δσ cov between σ cov and σ(δ) (classical two-tower 520 approach) was much larger than Δ σ cov between σ cov and σ(δ) corr,f (extended two-tower 521 approach). Δσ cov was reduced by 85.7% for the 20.5km distance and 79.3% for the 34km if 522 both sfd-correction and weather filter were used. However, after applying the sfd-correction 523 and the weather-filtering, the mean uncertainty estimate was still higher than the raw-data 524 based reference value (Tab.2), suggesting that for these large EC tower distances the sfd-525 correction and the weather-filter do not fully capture systematic flux differences and 526 uncertainty is still overestimated by the extended two-tower approach. This can have 527 different reasons. We assume the major reason is that the weather-filter is supposed to 528 capture all measured flux differences that can be attributed to different weather conditions at 529 both EC towers which cannot be captured with the sfd-correction. Applying stricter 530 thresholds could increase the efficiency of the weather filter but in our case the reduced 531 dataset was too small to allow further analysis. In general, the weather-filter did not improve 532 the uncertainty estimates as much as the sfd-correction. However, this does not imply that 533 differences in weather conditions are negligible when applying the extended two-tower 534 approach for larger EC tower distances. In fact the systematic part of measured EC flux 535 differences between both towers caused by (steady, systematic) among-site differences in 536 weather conditions were already partly captured with the sfd-correction. In contrast, such 537 systematic differences were difficult to capture with the weather-filter because much lower 538 thresholds would have been required. 539 The absolute corrected and weather-filtered uncertainty value σ(δ) corr,f [μmol m -2 s -1 ] was 540 slightly lower for the 34 km EC tower distance than for the 20.5 km EC tower distance 541 (Tab.2). The raw-data based reference σ cov [μmol m -2 s -1 ] however was also smaller for the 34 542 km dataset than for the 20.5 km dataset which can be related to the different lengths and 543 timing (i.e., different seasons) of the measurement periods for each of the five EC tower 544 distances: The roving station was moved from one distance to another within the entire 545 measurement period of ~ 27 months. During this entire time period of data collection, the 546 length and timing of the single measurement periods varied for the five EC tower separation 547 distances (Tab.1). This is not optimal because the random error is directly related to the flux 548 magnitude and the flux magnitude itself is directly related to the timing of the measurements. 549 Because in spring and summer flux magnitudes are higher, the random error is generally 550 higher as well (Richardson et al., 2006). To reduce this effect, we captured spring/summer as 551 well as autumn/winter months in each measurement period. However, the timing of the 552 measurements and the amount of data available were not the same for the five EC datasets. In 553 particular the permanent EC tower in Merzenhausen was measuring considerably longer (> 2 554 years) than the roving station did for the other four EC tower distances. Therefore, 555 differences of the mean uncertainty estimates for the five measurement periods were partly 556 independent of the EC tower distance. This effect gets obvious when looking at the mean 557 uncertainties σ cov estimated with the reference method, which should be independent of the 558 distance but were also found to be different for each dataset of the five EC tower distances. 559 Against this background, statements about how EC tower distances affect the two-tower 560 based uncertainty estimate need to be treated with caution. 561 The NEE uncertainty σ(δ) corr,f estimated for the grassland site Rollesbroich agree well with 562 the NEE uncertainty values for grassland sites by Richardson et al. (2006), and also the 563 regression coefficients (Fig. 2-3; Fig.5-6, Tab. A1) do not show large differences. This can be 564 expected since Richardson et al. (2006) applied their method for a very well-suited tower pair 565 with low systematic differences, such that the classical approach and our extended approach 566 should approximately converge. However, identical results are unlikely because even for two 567 very similar neighboring sites some systematic differences occur. In addition, the random 568 error is expected to vary between sites (see e.g. Mauder et al., 2013) which is in part related 569 to instrumentation. 570

Conclusions 571
When estimating the uncertainty of eddy covariance net CO 2 flux (NEE) measurements with 572 a two-tower based approach it is important to consider that the basic assumptions of identical 573 environmental conditions (including weather conditions and land surface properties) on the 574 one hand and non-overlapping footprints on the other hand are contradicting and impossible 575 to fulfill. If the two EC towers are located in a distance large enough to ensure non 576 overlapping footprints, different environmental conditions at both EC towers can cause 577 systematic differences of the simultaneously measured fluxes that should not be included in 578 the uncertainty estimate. This study for the grassland site Rollesbroich in Germany showed 579 that the extended two-tower approach which includes a correction for systematic flux 580 differences (sfd-correction) can be used to derive more reliable (less overestimated) 581 uncertainty estimates compared to the classical two-tower approach. An advantage of this 582 extended two-tower approach is its simplicity and the fact that there is no need to quantify the 583 differences in environmental conditions (which is usually not possible due to a lack of data). 584 Comparing the uncertainty estimates for five different EC tower distances showed that the 585 mean uncertainty estimated with our extended two-tower approach for the 95 m and 173 m 586 distances were nearly identical to the random error estimated with the raw-data based 587 reference method. This suggests that these distances were most appropriate for the 588 application of the extended two-tower approach in this study. Also for the largest EC tower 589 distances (20.5 km, 34 km) the sfd-correction significantly improved the correlations of the 590 flux magnitude and the random error and significantly reduced the difference to the 591 independent, raw data based reference value. We therefore conclude that if no second EC 592 tower is available at a closer distance (but available further away), a rough, probably 593 overestimated NEE uncertainty estimate can be acquired with the extended two-tower 594 approach although environmental conditions at the two sites are not identical. 595 A statement about the transferability of our experiment to other sites and EC tower distances 596 requires further experiments. However, we assume transferability is given if both EC towers 597 are located at sites of the same vegetation type (e.g. C3-grasses, C4-crops, deciduous forest, 598 coniferous forest, etc.). Flux differences caused by a different phenology can be very hard to 599 separate from the random error estimate, even though they are expected to be mainly 600 systematic and could therefore be partly captured with the sfd-correction. Moreover, the EC 601 raw data should be processed in the same way (as done here) and the measurement devices 602 should be identical and installed at about the same measurement height. Important is also that 603 the instruments are calibrated thoroughly and consistently. Because this was true for the three 604 EC towers included in this study, we conclude that systematic flux differences that are 605 corrected for with the sfd-correction arise mainly from different environmental conditions 606 whereas calibration errors are assumed to have a very minor effect. Different weather 607 conditions at both EC tower sites are a main drawback for applications of the two-tower 608 approach. While systematic differences of the weather conditions are expected to be captured 609 by the sfd-correction, less systematic weather fluctuations e.g. related to cloud movement, are 610 difficult to be filtered of the two-tower based uncertainty estimate. Applying very strict 611 thresholds can lead to a too small dataset, especially if the measurement periods are short. If 612 EC raw data is available, we recommend to use an uncertainty estimation scheme like the one 613 presented in Mauder et al. (2013). 614