Quality Control

All of the water and air xCO₂ measurements recorded during the Indian Ocean survey cruises were presented in the OTL original (preliminary) data files. Quality control (QC) flags (qflag) were used to identify "bad" (qflag = 4) measurements (later these measurements were removed from all data files), "questionable" (qflag = 3) measurements, and "good" (qflag = 2) measurements. Although there are several individual readings that can ultimately lead to a bad final value, one overall QC flag is reported for the measurement. This section describes the multilevel QC procedure performed by OTL and used to generate this flag. As described in the previous section, supporting measurements (sea surface temperature, salinity, and position) were filtered for bad values and interpolated to the times of the CO₂ measurements. Anyone interested in investigating the variability of these properties beyond its applicability to these CO₂ data is encouraged to return to the original IMET data set.

The first step in the calibration process was to normalize all of the detector CO₂ voltages to the mean detector temperature for that cruise and a pressure of one atmosphere. The first step in the QC protocols, therefore, was to remove any outliers in the detector temperature and pressure readings. Both of these measurements were very reliable with at most two to four isolated points removed on any given leg. Missing values were replaced with a linear approximation based on adjacent values.

The temperature- and pressure-normalized CO₂ voltages for each of the standards were analyzed for bad values. The collection program's criteria for determining when a CO₂ reading is stable were purposefully generous to prevent undersampling of real variability in the sample gases. Because the stability criteria were the same for sample and standard gases, the first point saved after switching to a new standard generally had not reached the equilibrium value. After visual confirmation of this phenomenon on each leg, the first point from each set of standards was filtered from the data set. Although rare, any exceptional outliers among the four remaining measurements on each standard were also visually identified and removed. The final calibration at each time was based on the mean of the remaining values.

Before the final calibrated values were calculated, a QC check of the equilibrator temperatures was performed. These data were quality controlled by examining all of the points recorded in 2-day intervals and outliers discarded based on visual inspection. Values were generally discarded when they were more than two standard deviations from a time local mean. The exact value for the cut, therefore, depended on the instrumental noise at the time. Questionable points were generally left in the data set. Bad values were replaced with a linear approximation based on adjacent values.

After calibration, the water and air data were broken into separate files. At this stage, every reading contributing to the water and air xCO₂ values has been quality controlled with the exception of the detector voltage. Unusual readings in the final data, therefore, either reflected real variability in the CO₂ concentration of the sample or bad voltage readings. Because it was not always clear which was the case and the final QC step was somewhat based on subjective ideas of how CO₂ behaves in the ocean or atmosphere, QC flags were created for each measurement. Only values that were known to be bad (qflag = 4) were removed from the final data set.

Marine air values showed little variability relative to the water measurements, which made identification of outliers easy. Values that were obvious outliers (qflag = 4) were visually identified by plotting the data from an entire leg as a function of time. High and variable values recorded when the ship was near land were only flagged when there were known detector problems since these values most likely represent real changes in atmospheric concentration. Questionable values (qflag = 3) were identified by carefully examining the data in 1- to 2-day intervals and marking isolated points that did not follow the local trend.

xCO₂ values in the surface seawater were generally much more variable than the marine air readings. A quality flag of "4" was reserved for water values that were clearly bad and for times when the seawater supply was shut down for extended periods but the automated CO₂ system continued to sample air from the equilibrator (I8S/I9S only). Measurements marked with a quality flag of "3" were either identified as data collected during brief bow pump failures or as single outliers that clearly did not fit with the surrounding data. The times of brief bow pump failure were identified by using the analyst's notes and by plotting the sea surface temperature together with the equilibrator temperature values as a function of time. The two temperatures tracked each other very well except when the bow pump shut down and the two temperature readings would decouple.

The showerhead GC underway xCO₂ system designed by Ray Weiss of SIO was running in parallel with the Princeton non-dispersive infrared (NDIR) instrument (see Appendix A) during all nine Indian Ocean cruises aboard the R/V Knorr. Both systems shared the same marine air supply and took water from the uncontaminated bow pump plumbing at essentially the same point. The sampling frequency of the two systems was very different. Approximately 25,000 water measurements and 8,000 air measurements were automatically logged by the Princeton instrument along the 10,000-km cruise track of WOCE leg I9N (from Fremantle, Australia, to Colombo, Sri Lanka). By contrast, the SIO system made approximately 2,000 water and air measurements (two samples per hour) on the same cruise. The high sampling frequency for the Princeton system (average water sample interval was 2.5 minutes) was designed to allow examination of the small-scale spatial variability in surface xCO₂ values. Changes of 10 to 20 ppm over a distance of 10 km are not uncommon in open-ocean surface waters. These gradients can be an order of magnitude greater in frontal regions or in coastal waters. Despite the different designs of the two systems (e.g., GC vs NDIR and shower vs disk equilibrator) the Princeton and SIO underway systems gave nearly identical results. Figure 5 is a plot of ΔxCO₂ (Princeton - SIO) for surface water versus time for WOCE leg I9N. To make a fair comparison, given the very different sampling rates, CO₂ values were interpolated from each data set to 24 evenly distributed times per day (the top of every hour) for the entire cruise. The range of surface water CO₂ concentrations covered in this comparison was approximately 300 to 420 ppm. The mean difference between the two systems (0.86 ± 2.7 ppm) was not statistically different from zero. The standard deviation of the difference not only reflects the potential variability introduced from the interpolations but also any real variability that may have been sampled by one system and missed by the other.