The most common indicators of low-quality data are either low backscatter amplitudes or low signal-to-noise ratios, which express the strength of the signal relative to a background noise level (SNR=20log⁡10(Asignal/Anoise)). Critical cut-off values for discarding data are instrument- and environment-specific. For the Nortek Vector used in the benchmark datasets, the manufacturer recommends values are 6 dB above the noise floor (around 50 dB) and 15 for amplitude and SNR, respectively . In contrast, an obstruction of an acoustic beam may manifest as high values in both backscatter and SNR. This obstruction may be identified by a large difference in amplitude between the adjacent acoustic beams, especially if another beam is obstructed.

ADV systems use pulse-to-pulse coherent technology in which a pair of pulses separated by a known time lag are emitted. The similarity between the measured echo of the two pulses is assessed and reported as a percentage value, which provides a further indication of data quality . It is worth noting that coherence and amplitude are functions of different parameters with varying sensitives, and thus provide two separate measures of quality control. Low correlation values can be used to identify bad data; but the converse is not necessarily true, that is, a large value doesn't indicate good quality data in all cases. A canonical cut-off value of 70 % has been shown to generally reduce variance within the dataset; however, values of 50 % are also commonly used as a cut-off.

Our quality control data format provides users with the ability to specify the threshold and rationale in the NetCDF data file, which should also be described in the methods of their scientific publications. We recommend that these additional user-defined flags also consider other data quality concerns, e.g., the pressure sensor indicating that the instrument was out of the water or change in pitch, roll, or heading indicating that the instrument has moved. Once bad data has been flagged, other QC measures such as phase unwrapping and velocity despiking can be applied to improve data quality.

For pulse-coherent instruments, the Doppler phase shift can only be determined from -π to π. For values outside this range, when the along-beam velocity exceeds the ambiguity velocity as set by the time-lag between pulses, “phase wrapping” can occur. This effect manifests as a sudden and unrealistic change in velocities, which is usually also accompanied by a change in sign. These “phase-wrapped” velocities can sometimes be corrected for in beam coordinates by subtracting or adding twice the ambiguity velocity. Several algorithms are available to transform phase-wrapped to actual velocities with reasonable success . However, it is nonetheless still preferable to set the nominal maximum velocity to a sufficiently large value when programming the instrument for deployment to reduce data processing issues . This maximum velocity setting requires prior knowledge about the expected velocities at a given site, which can be gained through previous field studies, use of operational tidal or hydrological predictions, or even specialised numerical modelling for the site. The velocities from our different benchmarks provide some examples for how velocities may vary across environments (Table ). For example, weak flows under sea ice (Under-ice MAVS) to very strong flows in constricted channels (Tidal MAVS). However, the velocities can still be much higher than expected from modelling and previous studies. A prime example is the Tidal shelf ADV benchmark, which was programmed with an expected maximum horizontal velocity of 0.8 m s⁻¹ which was expected to be sufficient given the environment (0.4 m above the seabed in 250 m of water), but velocities over 1 m s⁻¹ were observed, resulting in phase wrapping that needed to be addressed.

Short-lived transient spikes may contaminate velocity measurements, often only a few samples in duration but with different amplitude to the neighbouring “correct” signal. It is critical to remove as many of these spikes as possible, as they can dramatically alter the velocity spectra, which are required for fitting the inertial subrange model. Spikes in ADV measurements may manifest because of aliasing of the Doppler signal, which happens when pulses are contaminated through reflection off complex objects and boundaries .

To despike ADV velocity measurements, we recommend the median filter-based method described by . This method was originally developed for atmospheric measurements and was also recommended by review for despiking high-frequency atmospheric measurements of carbon dioxide used to quantify vertical turbulent fluxes. They found that the median filter was more robust than other filters such as the commonly applied phase-space thresholding method developed by . In particular, the median filter method (i) is better at handling missing points, (ii) copes with the presence of low-frequency coherent turbulent structures, and (iii) is less biased by spikes than other methods .

The method requires a window length, over which to calculate the median, and a threshold for spike identification. The window length must be longer than the duration of spikes, which can span consecutive samples but also must be sufficiently short to compute a reasonable local median see Table 1 of. For the spike-identification threshold, we recommend a method that calculates a histogram of velocity differences between the original minus the smoothed velocities . In this technique, the local minimums on either side from the center define the positive and negative thresholds, and velocity differences exceeding these thresholds are deemed to be spikes (Fig. ).

Each step flags velocity samples that do not meet the quality-control criteria. Each criterion uses binary (1/0) values that are then transformed into a “bitwise” flag. The maximum flag value depends on the number of criteria Nc used to assess data quality. The overall value of the boolean flag increases as the number of failed quality-control metrics increases. The maximum possible value for the flag will be less than 2Nc, which translates to 255 when eight quality-control metrics are being evaluated (Nc=8).

To flag the raw velocities, we use an 8 bit boolean number calculated from: 8VEL_FLAG=∑j∈INc2j-1, where j is the flag number amongst those applicable to each velocity sample. For example, I={2,5} if the second and fifth quality-control flags apply to the velocity samples, translating to a boolean value of 18. These boolean flags allow tracking the state of multiple metrics simultaneously while reporting only one number. A value of 0 means the velocity sample was not flagged and thus is likely high-quality. These boolean flags were designed with the Climate and Forecast (CF) metadata conventions (https://cfconventions.org/cf-conventions/cf-conventions.html#flags, last access: 11 March 2026) in mind.

Flags are identifiable in the NetCDF file by the FLAG suffix appended to the variable name, i.e., XYZ_VEL_FLAG for raw velocities collected in the instrument's XYZ coordinates, or EPSI_FLAG for ε^ estimates (Tables and in Appendix A). We also included additional metadata in the variables' attributes. We added “flag thresholds” and “flag thresholds meanings” to CF recommended “flag_meanings” and “flag_masks” attributes (see Table ). The “flag thresholds” provide the thresholds associated with a given flag, while the “flag thresholds meanings” briefly explain how the thresholds are applied. These extra attributes allow future users to revisit the flags, particularly the thresholds used, and apply more stringent (or less stringent) thresholds at their discretion. We also recommend that the these quality-control flags be briefly described in the methods of publication, along with the chosen thresholds for each flag.

We considered six methods to fit the model Ψj(k) to the spectral observations Φj(k), and subsequently estimate ε. Four methods fit the model spectra directly to the spectral observations, while the two remaining methods fitted log-transformed spectral observations by minimizing the sum of squared residuals or minimizing the sums of the absolute residuals logLSQ and logLAD methods in. Of the four methods applied directly against the spectral observations, two methods minimized the sum of squared residuals using a gradient or gradient-free minimization algorithms power and LSQ methods in. The third method in this category minimized the absolute residuals using the same gradient-free minimization as the LSQ method LAD method in. The fourth method tested relied on the maximum likelihood estimator and presumed statistical distribution of the spectra MLE method in. The full details of the methods and their assessment against synthetic spectra are described in . Spectra were synthesized by specifying ε in Eq. () and adding uncertainty to the spectra via two different statistical distributions . The sensitivity of the fitting methods was also evaluated against the uncertainty (smoothness) of the spectral observations, the number of spectral samples Ns, and the decadal range: 12δ=log⁡10kN/k1 between the first k1 and the last kN fitted wavenumber .

From that analysis, we recommend using fitting methods applied to log-transformed spectral observations. These methods returned more accurate results than those applied to untransformed spectra Figs. 3 and 4 of. These methods were less sensitive to outliers that might be introduced in the spectra by unremoved spikes in the raw timeseries. Specifically, we recommend minimizing the least-absolute deviation (residuals) between the fitted model and the observations logLAD method in. This method requires no assumption about the statistical distribution of the observations, unlike least-square regression, which expects normally distributed data. Minimizing the absolute residuals is considered less mathematically stable than least-square regression . However, since only the intercept β0 that best fits the log-transformed spectral observations Ψ^(k) is required; the technique amounts to: 13β0=medianln⁡(Ψ^i)-β1ln⁡(ki). The spectral slope β1 is set to the expected value of -5/3 for the inertial subrange (Eq. ), and i denotes each observation in the spectra. Linear least squares would take the mean of Eq. () rather than the median. Both least-square regression and least-absolute deviation performed well against the synthetic spectra . However, the estimated ε^ from least-absolute deviation were less biased than other methods, especially for spectra with low degrees of freedom, i.e., high degrees of uncertainty because of limited spectral averaging from using short segments τε.

From the fitted intercept β^0, we can estimate ε^ using: 14ε^=exp⁡(β0)ajCk3/2, where ajCk are the constants already defined above for the inertial subrange model (Eq. ). Using the logLAD fitting method and FFT-lengths that are 1/4 of the segment length (d≈14), the estimated ε^ is expected to be within 43 % of the actual ε if ten samples are fitted over one decade Figs. 3 and 4 of. This error reduces to 15 % when the number of samples increases to 100, which tends to occur when the fit falls at wavenumbers nearing the high wavenumber limit of the inertial subrange.

The synthetic spectra were also used to evaluate the ability of the fitting technique to estimate the spectral slope β^1 . The spectral slope estimates help flag spectra that do not exhibit a clear inertial subrange because of poor data quality, low energy, and anisotropy, or simply because the sampling protocol or spectral averaging cannot resolve the inertial subrange. The latter occurs when sampling too slowly or using fft-lengths that are too short to resolve the entire inertial subrange (Sect. ). As with the estimation of ε^, the methods applied to the log-transformed data were better at recovering the spectral slope β^1 than those used to untransformed data. The logLAD method was less sensitive to outliers than the least-square regression . The logLAD method involves minimizing the sums of absolute residuals rather than the sums of squared residuals see Eqs. 7 and 4, respectively of.

A primary challenge in obtaining ε^ is identifying the spectral wavenumbers that most likely belong to the inertial subrange, as these wavenumbers depend on ε^ – the sought quantity. Additionally, the location in wavenumber space also depends on the mean speed past the sensor U‾ when converting timeseries into spatial spectra. With appropriate choices for data sampling, segmenting (Sect. ), and spectral averaging (Sect. ), this step amounts to avoiding the low wavenumbers dominated by anisotropy and the high wavenumbers dominated by instrument noise. Other wavenumbers to avoid are those impacted by vortex shedding off instrument frames (e.g., Under-ice MAVS in Fig. ), or surface waves.

We evaluated four strategies for identifying the inertial subrange within the spectra by adding white noise to the same synthetic spectra used to test fitting methods . The two first strategies require computing the absolute deviation between each of the fitted log-transformed spectral observations and its log-transformed model spectrum given by Eq. (): 15Δlog⁡Ψi=|log⁡Ψ^i-log⁡Ψi| and then taking the mean or median of these deviations across different subsets of wavenumbers (Fig. d, e respectively). The third strategy estimated the mean absolute deviation of Ψ^i/Ψi Eq. 24 of. Our fourth strategy estimated ε^ and the spectral slopes β^1 over short wavenumber subsets of the spectra (see Fig. c). The wavenumbers with the estimated slope closest to the expected β1=-5/3 for the inertial subrange are then selected to calculate the spectrum's final ε^. The other three methods situate the inertial subrange over the wavenumbers that yield the minimum value for the calculated statistics.

The slope method was recommended because it performed slightly better than the other three against synthetic spectra of for which we added white noise floor to mimic the impact of instrument noise on observed spectra. The slope method tended to situate the inertial subrange at lower wavenumbers than the other three methods even when the synthetic spectra were further flattened by adding white noise at low wavenumbers. This additional flattening crudely replicated the narrowing of the inertial subrange when the energy containing range begins to impinge on the inertial subrange (e.g., k^≲1 cpm for Tidal Shelf ADV example in Fig. ). We note, however, that all four methods generally situated the wavenumbers within the inertial subrange of the synthetic and real spectra (e.g., Fig. c–e). Overall, we recommended the slope method because it performs slightly better than the other methods, while the estimated slope β1 is required later for flagging of ε^ at level 4.

In practice, the user must select the size of each subset, in addition to the wavenumber overlap for each subset within the spectra. For our benchmark datasets, we used an overlap equivalent to 1/20 of a decade, but other users may prefer shifting the window by one spectral sample at a time. We recommend a minimum decadal wavenumber range of δ=0.8 and that each fitted subset includes at least ten samples, especially when δ<1 . Users may also specify a maximum kmax and minimum wavenumbers kmin that can be fitted upon inspection of the spectral observations (e.g., Fig. ). For example, kmax≈2π/Lb where Lb is the distance to the nearest boundary, while kmin≈π/ℓ depends on the dimension ℓ of the sampling volume of the instrument. For each subset, the estimated ε^i is used to calculate LK, and verify whether the fitted wavenumbers are within the inertial subrange. The user may provide some allowance, but we recommend ensuring that the median fitted kmed is within the inertial subrange, i.e., kmed<0.1/LK. Spectra for which none of the fitted subsets satisfy this requirement were flagged (see Sect. ).

In some instances measurement noise may adversely impact the estimated ε^, and render the spectra unusable for estimating turbulence quantities. The drowning of the inertial subrange by noise is particularly common in low-energy environments such as the ocean interior or in lakes. To deal with this issue, some authors have presumed the shape of the noise spectra to either remove it from the spectral observations e.g.,, or add the presumed noise shape to the model used for fitting the unaltered spectral observations e.g.,. The former is challenging because measurement noise levels change with the flow conditions and must be estimated from the spectra itself see. The noise floor can be particularly hard to estimate in high energy flows being drowned by the signal. The latter method has the same challenges but prevents us from using spectral fitting methods applicable to log-transformed spectra. Rather than use these strategies, we recommend comparing the estimated ε^ to the minimum εm when the corresponding theoretical spectral energy levels in the inertial subrange (Eq. ) exceeds the noise floor Φn : 16εm=ΦnajCk3/2k5/2. This recommendation is based on the relatively low positive bias introduced when leaving the noise “as-is” in the spectra (Fig. ). The noise floor Φn typically appears as white noise (flat) in velocity spectra and can be estimated by averaging the energy levels of the highest wavenumbers where β^1≈0. Equation () depends on the fitted wavenumbers k and can be generalized further as a function of the smallest turbulent length scales LK defined in Eq. () : 17k=αLK=αν3/ε1/4. The inertial subrange corresponds to α≤0.1. Substituting this relationship for k into Eq. () yields the following minimum εm: 18εm=α20/3ν5ΦnajCk4. This equation demonstrates that noise is most detrimental when fitting wavenumbers approaching the theoretical high wavenumber bound of the inertial subrange (α=kLK=0.1). The minimum resolvable εm can be viewed as the measurement detection limit. The limit increases with noise levels, albeit the increase is amplified when when fitting large wavenumbers i.e., k close to the beginning of the viscous subrange (Fig. ). For example, the inertial subrange sits above the noise floor for ε>10-7 W kg⁻¹ at wavenumbers ten times smaller than the highest within the inertial subrange (α=0.01) provided the noise levels Φn are less than 10-6 m² s⁻² (rad m⁻¹)⁻¹ (Fig. a). However, if fitting the highest wavenumbers within the inertial subrange (α=0.1), the spectra sit above the noise floor only if Φn≲2×10-8 m² s⁻² (rad m⁻¹)⁻¹ (Fig. b).

By leaving the noise floor “as-is” in the spectral observations, a positive bias occurs when estimating ε^. The over-estimation is in addition to other errors associated with spectral fitting techniques described above. When ε^≈εm, the estimated ε^ over-predicts the prescribed value ε by a factor of 3 (circles in Fig. ). When ε^≈100εm, the estimated ε^ over-predicts ε by a factor of 1.5 (squares in Fig. ). These positive biases are relatively small compared to the range of ε encountered in environmental flows, especially considering how quickly the bias lessens when including lower wavenumbers (α≤0.1) in the fit. Hence, the biases shown in Fig. are conservative when α is determined from the largest fitted wavenumber kN.

To flag overly noisy spectra with Eq. (), the user must estimate the spectral noise floor Φn from the observations. For our benchmarks, Φn was calculated as the average spectral energy levels over the last δ=0.1 decadal range or the last 30 spectral samples, whichever leads to a larger number of averaged samples. The resulting Φn is stored as SPEC_NOISE_UVW in the same units as the spectra from which the average was determined at Level 3 UVW_VEL_SPEC. The spectral noise estimate Φn is used with the maximum fitted wavenumber kN stored in K_BNDS and the estimated dissipation ε^ stored in Level 4 to estimate εm using Eq. (). The estimated εm is stored as MIN_EPSI_NOISE at Level 4 so that this value can be used to generate quality control flags for ε^ in Sect. .

Nearby instrument frames may obstruct the flow or shed vortexes, contaminating the velocity observations. This contamination is often recognizable as a narrow band peak in the velocity spectra when the frame obstruct the flow upstream of the sampling following (Fig. ). In this situation, the contamination frequency fc can be determined from the Strouhal number Sr: 19fc=0.21U‾D for high Reynolds flow Sr≈0.21,. The frequency of the disturbances increases with decreasing diameter D of the obstruction or an increase in the velocities past the frame U‾. These disturbances are typically associated with a specific flow direction relative to the instrument's frame of reference. Hence, if the instrument is fixed in space, the flow direction can be used to identify when vortex shedding is potentially contaminating the velocity spectra, Alternatively, near boundaries, the estimated ε^ can be compared to the predicted values for the log-law of the wall and the flow direction see Fig. 4 of. This contamination may occur at wavenumbers higher than those within the inertial subrange, as apparent in the first 25 min of the Tidal Shelf ADV benchmark. The flow disturbances will sometimes be significant enough to increase energy levels within the inertial subrange. Thus, it is preferable to avoid the problem by placing the sampling volume of the instruments far from the obstruction. A common rule of thumb is leaving a space larger than ten times the diameter of the obstruction between the frame and the sampling volume. When the sampling volume is obstructed, ε^ will be over-estimated downstream of the disturbance see Fig. 4 of.

Another contamination source specific to the MAVS velocity sensor is the rings that house the acoustic transducers. These rings shed vortexes, contaminating the velocity measurements. This contamination affects most directions transverse to the instrument's main shaft. For a horizontally mounted instrument, such as our Tidal MAVS benchmark, the longitudinal direction was the least affected by the rings' vortex shedding . In contrast, the transverse direction was the most affected. The other Under-ice MAVS benchmark was mounted vertically on a rod lowered beneath the ice. The vertical velocity direction was the least affected, followed by the longitudinal, although the vertical component still shows evidence of a high-frequency flow disturbance in its spectra (Fig. ). Our quality-control metrics below include optional flagging for ε^ estimates affected by flow disturbances.

Our first metric presents the results from testing the stationarity of turbulent velocities. Both spectra computations and the inertial subrange model rely on the assumption of stationarity. The test is applied to each velocity component separately. For this purpose, we used the nonparametric test by , which involves calculating two statistics (standard deviations and average) over shorter subsets of each velocity timeseries segment (of duration τε). We then compare the number of runs (crossings) of each statistic about its average to the expected values in Table A.6 of . For this evaluation, the user must choose the significance level and the number of subsets to subdivide each segment. We used 20 subsets and a 95 % level, resulting in an acceptable number of runs between 6 and 15. We flagged ε^ as non-stationary if the number of runs for the standard deviation and means for the velocity segment are outside the expected range.

The second quality metric focuses on Taylor's frozen turbulence hypothesis. We recommend flagging ε^ estimates associated with low advection velocities U‾ relative to the root mean square of the turbulent velocity fluctuations along the direction of mean advection urms. This condition translates to flagging ε^ associated with urms/U‾ exceeding a threshold TH: 22urmsU‾>TH. We suggest TH=0.33 based on the work of who showed that the error for ε^ for this value is about 10 %. Larger errors are expected when urmsU‾ increases see Fig. 12 of. Thus, TH could be increased but it should always remain smaller than 1. The chosen threshold should be specified in the metadata of EPSI_FLAGS accompanying the ε^ estimates as detailed in Table .

The third flag considers whether the fitted spectra are drowned by noise. This criteria involves calculating the minimum resolvable εm using Eqs. () and () from the dissipation ε^ estimates, the maximum fitted wavenumber kmax, and the estimate of the spectral noise floor Φn (see Sect. ). The estimated ε^ is compared to the minimum resolvable εm. Dissipation estimates are flagged as being drowned by noise when the ratio between these two quantities is less than the user-defined threshold Tn: 23ε^εm<Tn. The threshold Tn must always be larger than one and specified in the ε^ flag's metadata (see Table ). We suggest Tn=3 owing to the positive bias shown when the highest wavenumbers of the inertial subrange sit near spectral observations as illustrated in Fig. ).

Our fourth flag involves estimating the spectral slope β^1 from the observations to identify spectral slopes that deviate too much from the expected -5/3 value (thus rendering suspect the associated estimate of ε^). These deviations may occur because of excessive noise, anisotropy, or other contamination (e.g., vortex shedding). An estimated slope that deviates from -5/3 can also occur when the slope is actually -5/3 due to statistical variability of the computed spectra d, the number of spectral points Ns used for the fitting, and the decadal range δ fitted Figs. 7 and 8 of . Their results were obtained by fitting the inertial subrange to 3200 synthetic spectra (100 spectra per degrees of freedom tested) with uncertainty generated using the χd2-distribution . Figure recasts their results in a different form to show that the estimated 99.7 % bounds can be represented by: 24β^1>β1±Aδ2dNs with empirical estimates of A ranging between 7 to 17. The maximum A from the numerical experiments was about 17 for Ns=100 and δ=1 (Fig. ). This A denotes the maximum deviation obtained from the expected β1=5/3 when fitting idealized velocity spectra.

We chose this maximum A=17 to flag ε^ with spectral slope estimates that fall outside the bounds given by Eq. (). Reducing A would render the threshold more stringent, flagging an increased number of ε^ estimates. Users should always specify their choice of A in the metadata of the Level 4 NetCDF quality-control ε^ flags (EPSI_FLAGS in Table ). This information in combination with the variables for estimating the acceptable bounds (δ, Ns, and d in Eq. ) should be also stored as variables at Level 4 to enable other users to re-flag ε^ if desired (see Table ).

Our fifth flag involves identifying segments with significant data loss during the quality-control of raw velocities, which render the spectra unreliable for estimating ε^ (Sect. ). Limited testing involving the random removal of velocity samples from our benchmarks showed that spectral shapes deviate considerably from the original when more than 10 % of samples are removed. As the percentage of data loss increases, the interpolated time series yield spectra with increased energy levels at low k and decreased energy at high k. We suspect acceptable data loss depends on data quality (i.e., noise levels) and underlying turbulence captured in the original time series. We thus recommend users specify their threshold TP for the maximum percentage of data loss in the segment after quality-controlling the raw velocities. For our benchmark datasets, we use 10 % as the minimum percentage of good samples in a segment.

The sixth metric is for flagging ε^ estimates associated with low energy turbulence, resulting in anisotropic spectra. This flag is not for excluding anisotropic spectra, when one velocity component is more impacted by the flow properties than the others. This impact manifests itself as flatter spectral slopes at low wavenumbers preferentially in one direction compared to the others (e.g., Ψ^w vs. Ψ^u for k<8 rad m⁻¹, Fig. ). This feature does not render the spectra unusable for estimating ε until the turbulence becomes so weak that the scales of the energy containing subrange start approaching those at the high end of the inertial subrange. The spectral slopes may start to deviate from the -5/3, but oftentimes a -5/3 slope still exists but its energy levels drop below those expected for isotropic turbulence (Eq. ). Spectra are considered too anisotropic when the ratio L/LK is small, which results in underestimating ε^.

This flag is optional as it typically requires an estimate of the largest turbulent overturns L (see Sect. ). The size of the large overturns depends on the mean flow characteristics and so necessitate measuring the background stratification (LO, Eq. ) or the background shear (LS, Eq. ) unless the distance from the boundary is a suitable alternative for L (Eq. ). To flag this issue, we compare the ratio L/LK to a user-defined threshold TA: 25LLK<TA. The threshold TA depends on the chosen measure for L and the velocity component seefor an extensive review. The longitudinal direction tends to have a broader inertial subrange than the vertical and transverse directions. We recommend a similar threshold TA=100 to , noting that higher thresholds might be necessary when the transverse or vertical components are used to estimate ε^. The user should specify the definition of their largest L and the threshold TA for flagging the data in the EPSI_FLAG metadata. For our benchmarks, this length-scale was set L=κLb and stored as at level 4.

The seventh flag identifies spectra when most fitted wavenumbers sit outside the inertial subrange. This situation arises when the median fitted wavenumber kmed during the search of the inertial subrange (see Sect. ) are high and always lies within the viscous subrange: 26kmed≳0.1LK . This situation typically arises when the speed past the sensor is very slow, or the spectra are very noisy, such that the algorithm places the inertial subrange at very high wavenumbers.

Alternatively, the fitted wavenumbers may be too small and thus outside the inertial subrange. This situation will arise if the inertial subrange is unresolved because the sampling frequency is too slow or the largest overturns' are negligible. For the benchmarks, these wavenumbers are those smaller than those dictated by the distance to the nearest boundary (see Fig. ).

The last flag is reserved for missing ε^ estimates or any other user-defined flag. For example, the user may flag data loss onboard the instrument, ε^ outliers in the time series, or perhaps unrealistically different ε^ between velocity components. Occasionally, all components will yield ε^ estimates, passing all quality control criteria, but significant differences still exist between components. This situation may occur, for instance, because of vortex shedding from nearby flow obstacles (Sect. ). We used this user-defined flag to denote velocity directions from the MAVS benchmark datasets that were impacted by vortex shedding.

In the NetCDF file, a final estimate for ε^ is stored as a 1d timeseries EPSI_FINAL. This parameter is effectively the “best” ε^ issued from the data processing and flagging with the EPSI_FLAGS. There can often be large differences in dissipation estimates among the three velocity components caused by differing impacts of noise, anisotropy at low wavenumbers, and vortex shedding on the spectral observations. Thus, we select one of the velocity component to produce the final ε^ estimates, and document the choice in the EPSI_FINAL metadata.