Technical Note: Oxygen Optodes on Profiling Platforms: Update on Response Times, In-Air Measurements, and In-Situ Drift

Oxygen optode measurements on floats and gliders suffer from a slow time response and sensors that are or eventually run out of calibration. Based on two dual-O2 Argo floats, we show how to post-correct for the effect of the optode’s time response and give an update on optode in-situ drift stability and in-air calibration. Both floats are equipped with an unpumped Aanderaa 4330 optode and a pumped Sea-Bird SBE63 optode. Response times for the pumped SBE63 were derived after Bittig et al. (2014). Also their methods were used to correct the time response bias. Using both optodes on each float, the time re5 sponse regime of the unpumped Aanderaa optode was characterized more accurately than previously possible. Response times for the pumped SBE63 are in the range of 25− 40 s, while they are between 60− 95 s for the unpumped 4330 optode. Our parameterization can be employed to post-correct the slow optode time response on floats and gliders. After correction, both sensors agree to within 2−3 μmol kg−1 (median difference) in the strongest gradients (120 μmol kg−1 change over 8 minutes or 20 dbar) and better elsewhere. However, time response correction is only possible if measurement times are known, i.e., 10 provided by the platform as well as transmitted and stored with the data. The O2 in-air measurements show a significant in-situ optode drift of−0.4 % yr−1 and−0.2 % yr−1, respectively. Optode in-air measurements are systematically biased high during mid-day surfacings compared to dusk, dawn, and nighttime. While preference can be given to nighttime surfacings to avoid this in-air calibration bias, we suggest a parameterization of the daytime effect as function of the sun’s elevation to be able to use all data and to better constrain the result. Taking all effects into account, calibration factors have an uncertainty of 0.1 %. 15 In addition, in-air calibration factors vary by 0.1− 0.2 % when using different reanalysis models as reference. The overall accuracy that can be achieved following the proposed correction routines is < 1 μmol kg−1.

It is therefore timely and useful to revisit the foundations of both pumped and unpumped O 2 optode behaviour on autonomous, profiling platforms, to ensure optimal data post-processing and best data quality.In this technical note, we want to refine the findings of Bittig et al. (2014) on the optode time response and of Bittig and Körtzinger (2015) on in-air measurements and in-situ drift stability.
Using the faster SBE63 optode O 2 as reference, the 4330 optode time response can be derived relative to the SBE63 optode.
However, since the SBE63 optode's time response is well-defined :::: (±3 s : ) due to the pumped mode of operation, the results of Bittig et al. (2014) can be used to reconstruct a "true", time response-unbiased O 2 profile as reference.The 4330 optode time response can then be derived relative to that reference O 2 profile.Both approaches yield equivalent results and only the second one is shown below.Reconstruction of the unbiased O 2 profile is carried out as follows (see also Appendix/supplemental material): -Define the flow regime for the pumped SBE63 optode, i.e., l L = 34 µm.
-Translate the flow regime to a response time constant τ using the in-situ temperature and data from Fig. 11 in Bittig et al.
-Obtain the mean of consecutive data points of the unbiased O 2 profile from the SBE63 optode measurements using the inverse of the filter's bilinear transformation, i.e., eq.37 from Bittig et al. (2014) (eq. A3).
-Reconstruct the unbiased O 2 profile by interpolating the mean of consecutive data points of the unbiased O 2 profile to the original measurement times.
The analysis of the Aanderaa 4330 optode flow regime behaviour takes the opposite approach by filtering or "delaying" the reference O 2 profile until it matches the 4330 optode observations.For this, the forward filter (eq.-Estimate a boundary layer thickness l L for the given short interval of measurements (here: 12 consecutive data points).
-Translate l L to a response time constant τ using the in-situ temperature and data from Fig. 11 in Bittig et al. (2014) (see also supplemental material).
-Translate τ to the filter constants a and b using the measurement times t i and eq. 4 from Bittig et al. (2014) (eq. A1).
-Apply the recursive filter to the mean of consecutive data points of the unbiased O 2 profile and the previous (or inital) filtered O 2 measurement (eq. 4 and 5 in Bittig et al., 2014) (eq.A1 and A2).
-Compare the filtered O 2 profile with the Aanderaa 4330 optode measurements and iteratively refine the chosen boundary layer thickness l L for the given short interval until a best match is achieved.
Overall, the resulting boundary layer thickness l L is smaller than estimated by Bittig et al. (2014).l L decreases with increased flow at the sensor as expected (figure 1).Specifically, l L is around 210 µm at 0 dbar s −1 and 100 µm at 0.095 dbar s −1 .
For other CTD applications with the flow passing tangentially along the optode sensing foil (similar to typical glider optode The main oxycline coincides with the pycnocline and the float slows down significantly around 50 dbar, giving the O2 optodes time to adjust.After correction, both sensors agree well in the profile fine structure and the depth of the oxycline, while they diverge before.The time response effect is more pronounced in regions with a weaker density gradient. attachments), l L was close to 40 µm at 1.0 dbar s −1 (Bittig et al., 2014).Using these endmembers, we chose a piecewise linear interpolation to parameterize the flow regime / l L vs. platform speed v for floats (equation 1).For gliders, we suggest equation 2, accordingly.
This flow regime characterization can now be used to reconstruct a time response-unbiased O 2 profile from the Aanderaa 4330 optode the same way as for the SBE63 optode.It is to stress, however, that all time response corrections need to be done on a time axis ("timeseries"), not on a pressure axis ("depth profile").It is therefore important that a time stamp is logged with each optode measurement.

10
As expected, time response correction for the pumped SBE63 optode with small response times changes the O 2 data only to some extent (median absolute difference in the strongest gradient: 6 − 8 :::: 6 − 7 µmol kg −1 ; 3h) and is much more important for the unpumped 4330 optode with larger response times (13 − 17 ::::::: 13 − 15 µmol kg −1 ; 3g).After correction, both sensors agree to within 2 − 3 µmol kg −1 (median difference) in the strongest gradient region (120 µmol kg −1 change over 8 minutes or 20 dbar) and better elsewhere (3f).Especially near the surface and in the core of the oxygen minimum zone, the difference is close to zero.However, there is a residual pressure effect between the sensors of ca. 2 µmol kg −1 at 2000 dbar (3f), in line with a pressure correction uncertainty of 0.3 % per 1000 dbar each (Bittig et al., 2015a).
This demonstrates that a time response correction is both useful to remove sensor artifacts (3g, 3h) and feasible for both optode models, yielding ::::: highly : consistent results (3f).It is, however, only possible if the respective measurement times are known, i.e., provided by the manufacturer through the instrument's firmware as well as transmitted or stored in the data stream.
Interestingly, a similar analysis on the limited dataset of the third float, WMO ID 6900891, yields apparent boundary layer thicknesses l L for the 4330 optode that are 100 µm or more higher.We believe this is caused by growth of a biofilm on the sensing foil during its 5-week irregular near-surface drift, which artificially delays the time response.The SBE63 optode, being contained in the CTD's pumped path and antifouling regime, is probably better protected against such artifacts.

In-Air Measurements
The feasibility to calibrate O 2 optodes on floats by measurements in-air, proposed by Körtzinger et al. (2005), was first demonstrated by Fiedler et al. (2013).Bittig and Körtzinger (2015) showed the first long time series of in-air measurements over 15 months using a single float.They ::::: There, ::: we demonstrated a systematic effect of any water-side oxygen disequilibrium on nearsurface in-air measurements, that can and needs to be corrected.Subsequently, Johnson et al. (2015) confirmed the consistency between in-air and in-situ calibrations for a fleet of floats, while Bushinsky et al. (2016) showed the utility of several in-air data acquisitions per surfacing, a daytime dependence, and the potential for in-situ drift detection.
Bittig and Körtzinger (2015) used the then available ca. 15 months of data for float 6900890.Here, we want to expand their ::: our analysis to all data until August :::::::: December : 2016 and apply it to both dual-O 2 floats 6900889 and 6900890 (i.e., 21 and 34 :: 25 ::: and ::: 38 months, respectively).Both floats show a similar "carry-over" slope c of 0.21 and 0.23, respectively, similar to Bittig and Körtzinger (2015) and Johnson et al. (2015).Using this value, we can compensate for the surface water contamination of the in-air measurements and derive an oxygen correction factor m i (or gain factor g i as in Johnson et al., 2015) for each profile i (figure 4), where pO 2,obs is the uncalibrated optode O 2 partial pressure (or O 2 saturation or O 2 concentration) and pO 2,corr In contrast to Bittig and Körtzinger (2015), the longer float time series confirm the results from Bushinsky et al. (2016), i.e., the existence of a small but significant in-situ drift of the optode (loss in O 2 sensitivity of 0.2 − 0.4 :::::::::: 0.27 − 0.40 % per year) as well as a daytime dependence of the in-air measurements.Over the course of the day, correction factors decrease, i.e., optode in-air measurements apparently increase, by roughly 1 % between 7 h and 17 h local time, which is in the same direction but of slightly smaller magnitude than their findings (figure 5).float 6900889 (grey circles) and 6900890 (black dots).mi was calculated for each in-air measurement by compensating the surface water contamination with a uniform "carry-over" slope c for the entire deployment.There is a trend towards higher correction factors mi with time, indicating a small drift of the optodes to lower sensitivity.Daytime has a noticeable effect on in-air measurements with correction factors mi being decreased during the day by about 1 %, i.e., the optode readings being apparently increased.Both effects, the in-situ drift and daytime dependence, act in parallel and are superimposed in the data presented here.
One can postulate that the daytime bias is related to solar heating of the sensor's sensing foil and thus the solar elevation angle θ.In that case, the daytime effect ∆m i would correlate on average with the cosine of the zenith angle (eq.4).
If the bias is caused by the energy input to the float top cap, the daytime effect would be related to the direct sun irradiance, which is a function of the sun light's attenuation in the atmosphere (Meinel and Meinel, 1976, eq. 5), where AM is the air mass coefficient for the path length through the atmosphere, modelled as a spherical shell of 9 km around the Earth (radius 6371 km).The data can be adequately described by both approaches.We thus used the solar elevation angle θ and the simpler equation 4 to parameterize the daytime dependence (eq.6), where ∆t i is the time since deployment (in days) for each profile i and a is the annual drift rate in % yr −1 .If enough data are available (as in our case), the linear carry-over (see Bittig and Körtzinger, 2015, eq. 10), the daytime dependence (eq.6), and the in-situ drift (eq.7) can be fit simultaneously (with fit parameters c, m t=0 , m A , and a) to correct all effects at once and obtain an optimal fit (eq.8).
Consequently, data can be corrected by: Table 1.Simultaneous fit of "carry-over" slope c, initial O2 correction factor mt=0 relative to the deployment calibration, daytime dependence amplitude mA, and annual drift rate a (eq.8) for floats 6900889 and 6900890 using air pO2 derived from different reanalysis models.
The analysis was limited to all profiles until the most recently available ERA-Interim date, i.e., July :::: Sept. 2016.For simple calibration, however, preference should be given to dusk, nighttime, and dawn measurements (θ i ≤ θ c ), where the (yet uncharacterized) amplitude m A of the daytime dependence is irrelevant (eq.6).
For float 6900889, m A is ca. 1 :: 0.8 %, while it is around 0.6 ::: 0.5 % for float 6900890 (table 1).Both floats show a significant in-situ drift: O 2 levels decrease by −0.4 :::::: −0.40 % yr −1 and −0.2 :::::: −0.27 % yr −1 , respectively, with the 95 % confidence level being around 0.15 ::: 0.11 % yr −1 after 2 years of float data , and below 0.1 ::: and :::::: around :::: 0.06 % yr −1 after three years.The "carryover" slope c of float 6900889, operating in open ocean waters, is not as tightly constrained as for 6900890, operating closer to coastal and upwelled waters, where the range of oxygen super-and undersaturation is larger.When accounting for the surface water "carry-over" effect only, fit root mean squared errors (RMSE) are around 0.8 mbar (Bittig and Körtzinger, 2015).With all effects included, the RMSE is reduced to 0.5 mbar.The intercept m t=0 gives an indication of the uncertainty in m i after daytime and drift correction, which is on the order of 0.1 %.
ERA-Interim has a delay of 2-3 months, so the analysis was limited to all profiles until July 2016 (table 1).Differences in m t=0 are up to 0.1 − 0.2 % depending on the reanalysis model and as such on the same order of magnitude as the uncertainty of the calculation.This should be considered when assessing the absolute calibration accuracy of in-air data.Differences for c, m A , and a are within their confidence intervals.near the surface and (bottom) at depth for float 6900889 (grey) and 6900890 (black).There seems to be a conditioning effect on either one (or both) optodes during the first half year / ≈ 40 profiles, after which differences between optodes are stable.
Since the SBE63 optode on floats can not measure oxygen in-air, a direct comparison with in-air measurements is not possible.However, it can be done indirectly via the difference between the 4330 and the SBE63 optode (figure 6), both near the surface within the mixed layer as well as at depth (1900 dbar).The SBE63s have been in-situ calibrated against the 4330 optodes near the surface, so differences cluster around zero with a standard deviation of 0.2 %.At depth, however, there is a difference between both on the order of 0.5 %, which we attribute to uncertainties in the pressure correction (Bittig et al., 2015a).
Moreover, at both depths, there seems to be a conditioning effect during the first half year (≈ 40 profiles) only after which the difference between both sensors stabilizes.Since we can only discuss the difference, this can be caused by either one or both sensors.Reasons could be repeated pressure cycling, removal of surface films or contaminants from the sensing foils, or others, and the exact cause remains speculative.However, we recommend that users critically evaluate their optode data whether they see such a conditioning artifact, too.

Conclusions
A slow O 2 time response reduces fine scale resolution and causes a lag between in-situ and observed O 2 profile.It is important for data users to assess and to quantify the impact on O 2 data quality, as the effect can be substantial for the analysis (e.g., Plant et al., 2016).However, as demonstrated by Bittig et al. (2014) and here, it is feasible to correct for the time response effect on slowly-moving, buoyancy-driven platforms.In this work, we refined :::: refine : the flow regime required for such corrections for floats and gliders and summarized ::::::::: summarize the procedure for time response corrections (from Bittig et al., 2014).Reconstruction, however, is only possible if measurement times are known.They ::::: These : need to be provided by the instrument's firmware/manufacturer as well as being transmitted and stored with the original profile data.Of course, every correction introduces uncertainty, so the ideal case would be a fast (and accurate) O 2 sensor that doesn't need time response correction.
In-air measurements at every surfacing of Argo-O 2 floats prove to be an ideally-suited mean to check and correct a change (or loss) in O 2 sensitivity of optodes, the primary reason of O 2 optode drift (Bittig and Körtzinger, 2015).As such, they allow both to correct sensor drift between calibration and deployment (several % magnitude) as well as sensor drift in-situ during deployment (order of −0.5 % yr −1 ).In detail, measurements in air close to the sea surface are not purely 100 % air as one would expect under controlled, stable conditions.Most prominently ::::::::: importantly, any water-side oxygen disequilibrium has a systematic effect on near-surface in-air measurements, that can and needs to be corrected.In addition, mid-day observations are offset with respect to night time or dusk and dawn measurements, such that a failure to account for the daytime effect introduces a systematic bias.While its mechanistic origin is still unclear, the data suggests a relation to solar heating.Despite these secondary effects, any float O 2 data calibrated with regular in-air measurements is more accurate than without in-air measurements (compare Takeshita et al., 2013), but neglecting the water-side "carry-over" or daytime effect should be avoided as good practice.
Based on regular in-air measurements over 34 and 21 :: with an uncertainty of the greater of 4 µm or 10 %.
For unpumped optodes on fast profiling platforms, i.e., CTDs, the vertical velocity v -boundary layer thickness l L relation depends to a large extent on the mode and orientation of attachment (see difference between experiment 5 and 6, Bittig et al., 2014) with a range of l L (CTD) from 20 µm to 55 µm (1σ from 5 µm to 35 µm ).For slowly profiling platforms, we suggest to use eq. 1 for floats and eq. 2 for gliders.They might need adjustments depending on the actual optode attachment or mode of operation.
The boundary layer thickness l L can be converted to a response time τ using the sensing membrane thickness l M of the respective optode model, the two-layer diffusional model of Bittig et al. (2014), and temperature T .For the SBE63 optode, they used l M = 130 µm, for the Aanderaa 3830/4330 optode with standard foil 100 µm, and for the Aanderaa 4330F optode with fast response foil 50 µm.Results for the latter two were shown in their figure 11, Bittig et al. (2014).Here, the same model data of τ against l L and T can be found for all three optode sensing foils in the supplemental material.
. (2014) used a set of laboratory experiments to validate a two-layer diffusional model of optode time response.In essence, water flow determines the liquid boundary layer thickness, l L , in front of the optode's sensing foil, which is essentially independent of temperature.Temperature, however, modifies the O 2 solubility and diffusivity both in the boundary layer and sensing foil, such that the optode's response time, τ , shows a marked temperature dependence (at constant flow / l L ), being slower at low temperatures.Shipboard CTD-O 2 casts with optodes, as well as glider and float data in different ocean regions served ::::::::::::::: Bittig et al. (2014) to characterize the response time τ as well as the flow / l L regime on these profiling platforms.However, while the shipboard casts had both a fast CTD-O 2 sensor and an O 2 optode on the same platform, the floats and gliders featured only a single optode and relied on a match of the optode data to a "nearby" CTD-O 2 profile.This match introduced some uncertainty into the analysis of Bittig et al. (2014).Using the dual-O 2 floats of the present study, the response time and boundary layer regime of optodes on slowly profiling platforms can be constrained far more tightly with the same methods.

Figure 1 .
Figure 1.Boundary layer thickness lL vs. float vertical velocity for an Aanderaa optode 4330 with standard foil derived from the dual-O2 floats 6900889 and 6900890.The colour shading indicates the data density, while the dashed line gives the parameterization of equation 1

Figure 4 .
Figure 4. Time series (top) as well as daytime dependence (bottom) of the oxygen correction slope mi from in-air measurements for

Figure 5 .
Figure5.Same as figure4except that the daytime dependence was removed from the time series (eq.6, θc = 15 • ; top panel) and the in-situ drift removed from the daytime dependence (eq.7; bottom panel).Thick lines give a 6 months (time series) and 3 hours (daytime dependence) low-pass filtered trend with ±2σ as thin lines.The mi low-bias during mid-day is an important contributor to the average in-air correction (compare the time series of mi without and with compensation of the daytime dependence, figure4and 5, top panel, respectively) and needs to be considered.