In drift analysis of salinity sensors, one major problem is the difficulty in delineating sensor drift from water mass changes. In the present study, a new method is proposed for finding sensor drift that is free from water mass changes. The efficiency of this new method in finding out possible drift in the Argo salinity is demonstrated in the Sea of Japan (SOJ) by using the “near-linear” subsurface salinity structure of the SOJ. The new method is based on the time–space decorrelation scale. The salinity difference between two neighboring observations within the time–space decorrelation scale (SALD) is used to find out possible drift. Neighboring observations within the time–space decorrelation scale are referred to as matchups. The SALD derived from matchups between Argo floats and shipboard CTD observations from the SOJ shows linear drift. Although all four selected floats (5 yr completed) from the SOJ show linear drift (<0.001 PSS yr−1), the drift alone is not so significant to affect the objective of the Argo program in understanding climate variability. In the SOJ, SALD identified salinity error other than drift in good quality data that are flagged by the Argo delayed-mode quality control (ADMQC) method. Therefore, SALD could be used as an effective additional tool in the Argo data quality control. To examine the applicability of SALD in open ocean regions, in addition to confined basins such as SOJ, SALD was applied successfully to detect salinity error in Argo data from the subtropical North Pacific (SNP).
Salinity plays a major role in ocean circulation to affect regional and global climates. Therefore, quality observations of salinity are important for understanding ocean circulation and climate variability. Argo is the first global ocean network of profiling floats providing real-time observation of surface and subsurface salinity, along with temperature (Davis et al. 2001; Larson et al. 2008). The Argo program and its data management system began with regional arrays in 1999, scaled up to global deployments by 2004, and achieved its target of 3000 active instruments in 2007 (Roemmich et al. 2009). Observations from this large number of floats have been contributing significantly toward our continuous effort in understanding climate variability.
During the last decade, Argo has generated a large volume of data, and more will be generated in the future. Data, in general, are susceptible to errors. In the case of Argo, the susceptibility is higher. Because 90% of the Argo floats have electrode-type conductivity cell (see Sea-Bird online at http:/www.seabird.com/), the “prolonged and unattended” presence in the ocean make them susceptible to fouling–biofilm formation inside the cell, which alters the conductivity and thereby the salinity measurements. Reports based on postdeployment calibrations of conductivity cells from six recovered floats (Table 1) have shown no significant drift (Oka 2005; Janzen et al. 2008). Although these preliminary findings on salinity drift are encouraging for the Argo program and end users, it is important to note that the calibration was not in situ and none of the recovered floats in Table 1 have completed the expected life span of 4 yr. Moreover, two of them ran aground and one remained in the land for more than 1 yr.
Various objective methods are employed for correcting gross errors in large volumes of data, both real and delayed mode (Wong et al. 2003; Owens and Wong 2009; Gaillard et al. 2009). However, objective methods may not be efficient enough to identify the possible long-term sensor drift in the conductivity cell. For example, Wijffels and Tchen (2007) intercompared objective methods used by various data assembly centers (DACs) and reported large inconsistencies in methods used and results obtained in the delayed-mode quality control. Reasons for such inconsistencies could be due to (i) natural variability (even at deeper depths), (ii) the lack of good quality background data, and (iii) difficulty in delineating sensor drift from long-term water mass change.
Considering all of the above facts, in the present study, long-term sensor drift in the conductivity cells of floats that completed 5 yr are examined using the “near-linear” subsurface salinity structure in the Sea of Japan (SOJ). In drift analysis, one major problem is the difficulty in delineating sensor drift from water mass changes. In this study by defining a new parameter—the salinity difference between two neighboring observations within the time–space decorrelation scale (SALD)—we have examined the drift. SALD’s definition and its merit in identifying salinity drift compared to existing objective methods are provided in section 2. The near-linear aspect of the salinity structure in the SOJ compared to other regions of the global ocean is explained in section 3. Time–space decorrelation scales of salinity in the SOJ are presented in section 4. The selection of quality floats for drift analysis and the observed drift are explained in section 5. Section 6 provides applicability of SALD in open ocean regions, in addition to the confined basin as SOJ. Discussion is provided in section 7, which is followed by a summary and conclusions in section 8.
2. SALD: A derived parameter for drift analysis
Salinity trends in the global oceans (Bryden et al. 2003; Read and Gould 1992) could be misrepresented as salinity drift (Wijffels and Tchen 2007). Therefore, any attempt to identify the salinity drift in Argo data should take care of long-term trends in salinity. In the present study, this is taken care by considering the new derived parameter SALD. Neighboring observations within the time–space decorrelation scale are referred to as matchups.
The decorrelation scale is defined as the time or space distance over which neighboring data points are correlated at some specified confidence level. Because the natural variability within the decorrelation scale is insignificant by definition, if SALD is found to be significant in value and variability, it could be attributed to sensor error. If the observed error shows trends, then the same could be attributed to sensor drift or water mass changes. Because SALD is derived from matchups (neighboring observations within the time–space decor relation scale), it is independent of water mass changes. Thus, compared to the existing objective methods used in the Argo program (Wong et al. 2003; Owens and Wong 2009; Gaillard et al. 2009), SALD has the merit of delineating drift from long-term water mass changes.
3. Salinity structure in the SOJ
The SOJ is primarily a confined basin with an average depth of 1500 m, and it is connected with the North Pacific through four straits with sill depths that are shallower than 140 m. The SOJ possesses its own deep-water formation in wintertime, the so-called deep convection, and comprises various water masses, such as intermediate water, central water, and bottom water, just like in the open oceans (Kim et al. 1996). The SOJ exhibits a long-term salinity trend of the order of +0.0006 PSS yr−1 at a depth of 500 m (Kwon et al. 2004). The near-linear subsurface salinity structure of the SOJ compared to that of other regions of the world oceans is demonstrated in Fig. 1. In the SOJ, at depths greater than 500 m, the vertical salinity distribution is almost linear, which is very distinct compared to other basins. This special feature of the salinity structure facilitates SOJ as a suitable basin for calibration of Argo salinity. The variability in salinity for the waters at a depth of 1000 m is of the order of 0.001 PSS, which is one order of magnitude smaller than that of the open ocean (Kim et al. 2004). Thus, SOJ is a suitable basin for in situ validation of Argo salinity to an accuracy of 0.001 PSS. Being a marginal sea, floats are confined to the SOJ and the confinement facilitates repeated annual calibration with available shipboard CTD observations.
Since 2001, 243 floats have been deployed in the SOJ, contributing toward the highest float/profile density in the global ocean (Argo 2010). The confined nature of the basin restricts the floats to drift within the basin, facilitating more shipboard CTD matchups for validation. Some of these floats used in this study have completed 5 yr of data collection with good quality results.
4. Time–space decorrelation scale in the SOJ
Using periodic shipboard measurements of 40 yr off the Korean coast, Park and Kim (2007) reported a decorrelation length scale of 20–77 km at depths greater than 500 m by analyzing the correlation coefficients among stations. In the present study, because the Argo data are from the open sea and the validations are at higher depths (about 1000 m), the decorrelation length scales are expected to be different from that reported by Park and Kim (2007).
For identifying the appropriate time–space decorrelation scale in the SOJ, the high-quality shipboard CTD (Seabird 41) archive available in the Japan Oceanographic Data Centre (JODC) for the period 2001–10 is analyzed. For finding time–space decorrelation scales, we have used only shipboard CTD stations (nearly 10 000 matchups are used). SALD is derived at a potential temperature (θ) surface of 0.4°C (~1000 m). Figure 2 represents the time–space decorrelation scale derived from matchups using shipboard CTD observations from JODC. While Fig. 2a shows the decorrelation scale of time (day difference between matchups), Fig. 2b represents space (distance between matchups) decorrelation. As depicted in Figs. 2a,b, for a time scale of 6 days and space scale of 100 km, the regression coefficients are close to zero and are not statistically significant at 99% confidence level.
5. Selection of quality floats for drift analysis and the observed drift
Shipboard CTD data from JODC are used to check the quality of float salinity. The data have been collected on board Research Vessel (R/V) Seifu Maru, operated by the Japan Meteorological Agency (JMA). Delayed-mode float data for the SOJ has been downloaded (see http://www.usgodae.org/). From inventory information of ship stations and float profiles, shipboard and float CTD matchups have been selected based on the time–space decorrelation scale (6 days and 100 km), as illustrated in Figs. 2a,b. Although there are about 200 floats in the SOJ, only 9 floats could identify matchups with shipboard CTD within the decorrelation scale. The mean and standard deviation of SALD derived from the matchups of nine floats are given in Table 2. Mean SALD for floats 5900193 and 5900195 are greater than 0.02 PSS and therefore are not selected for the drift analysis. Although the remaining seven floats show SALD less than 0.006 PSS, and only four floats (2900438, 2900440, 2900441, and 2900525) are selected for the drift analysis because the number of matchups for theses floats are statistically significant. The expected life of an Argo float is 4 yr. However, the selected four active floats have collected data for 5 yr. Thus, these four floats provide a good opportunity to examine Argo salinity drift beyond the expected life span of Argo floats. The trajectory of the four selected floats with profile positions and the shipboard CTD matchup positions (blue filled circles) are provided in Fig. 3. Although three floats (2900438, 2900440, and 2900441) are deployed in 2004 (the float 2900525 is deployed in 2005), matchups (shipboard and float CTD) for all four floats are available only from 2005. JMA used the Sea Bird Electronic 911 plus CTD for data collection from 2005, and the conductivity sensor is routinely calibrated (calibration information are available at http://www.data.kishou.go.jp/). The shipboard CTD data used for the drift analysis are of high quality.
The annual mean and standard deviation of SALD for each float are presented in Fig. 3e. Float 2900438 has shipboard CTD matchups spanning from 2005 through 2010 without any gap. While floats 2900438 and 2900400 have positive biases and positive drifts, float 2900525 shows negative bias and negative drift. Float 2900441 shows negative bias, whereas drift is positive. Table 3 shows annual mean SALD and standard deviation. Maximum bias of 0.00489 is observed for the float 2900440 and the minimum bias of −0.00074 is observed for the float 2900441. Although the magnitudes of the biases are different, linear drift shows a near-uniform distribution, with an average annual linear drift of 0.0009 yr−1. This annual linear drift does not consist of long-term water mass changes because SALD is free from long-term water mass changes. From the observed biases and drift, it is evident that once the biases are removed for each float, the observed drift of the order of <0.001 PSS yr−1 becomes insignificant.
Long-term drift and biases in the Argo salinity discussed above have been derived using float-to-shipboard CTD matchups. SALDs derived from float-to-float matchups using only Argo observations in the SOJ identify error other than long-term drift. This is illustrated in Fig. 4. In this case also, the same time–space decorrelation scale of six days and 100 km are used. SALD values derived from Argo matchups are flagged as erroneous data if the former is larger than the maximum SALD derived from the shipboard-to-shipboard CTD matchups (Fig. 2).
From Fig. 4, except for floats 2900437 and 5900195, all other floats used in matchups of Fig. 4 SALDs are well within the 0.01 PSS (gray dots) and confirm the good quality flag (1) assigned by Argo delayed-mode quality control (ADMQC). However, it is obvious that float 2900437 (blue dots in Fig. 4) has large errors in the initial profiles; this is also true for float 5900195 (red dots in Fig. 4). Calibration with shipboard CTD has also confirmed large error in float 5900195 (Table 2). Thadathil et al. (2004) have reported such errors in the initial profiles from the Indian Ocean. Because the matchups of 2900437 and 5900195 are from good quality Argo data flagged by the ADMQC method, the observed errors in these floats, as seen in Fig. 4, substantiate the efficiency of SALD in Argo data quality control.
6. Applicability of SALD in the subtropical North Pacific
In regions where natural variability is of the same order of possible salinity drift, SALD could be used for drift as well as error analysis. If natural variability is greater than the possible drift, SALD could be used only for finding errors unrelated to drift. This possibility is examined in the subtropical North Pacific (SNP) by considering the area marked as the blue square in Fig. 1. As discussed in the case of SOJ, SALD has been derived for SNP using shipboard CTD data from JODC. SALD from SNP is derived at the θ surface of 2.5°C (the approximate depth limit of the Argo data). Figure 5 represents the time–space decorrelation scale derived from matchups using shipboard CTD observations from JODC. While Fig. 5a shows the decorrelation scale of time (the day difference between matchups), Fig. 5b represents space (the distance between matchups) decorrelation. As depicted in Figs. 5a,b, for a time scale of 5 days and a space scale of 80 km, the regression coefficients are close to zero and are not statistically significant at the 99% confidence level.
Using the above time–space decorrelation (5 days and 80 km) SALD values are derived from Argo matchups using Argo delayed-mode data from SNP (http://www.usgodae.org/). In the case of SOJ, delayed-mode data available in the site have a good quality flag (1). However, in the case of SNP, the delayed-mode data available in the site comprise good (flag 1) and bad (flags other than 1) quality data. SALD values derived from Argo matchups are flagged as erroneous if the former is larger than the maximum SALD derived from the shipboard matchups (Fig. 5).
Gray dots in Fig. 6a show common erroneous Argo data identified by both SALD and ADMQC method. Blue (float 2900238) and red (float 29035) dots show erroneous Argo data identified by SALD and flagged by ADMQC for good quality. For further substantiating the observed error in floats 2900238 and 29035, shipboard CTD observation within the time–space decorrelation is available only at one station (shown in Fig. 7) for float 29035. However, the plot of Argo salinity at 2.5°C (Fig. 6b) for matchups involving these floats (with other floats from the same region) further corroborate the observed error in floats 29035 and 2900238. In the case of float 2900238, the error exists only in initial profiles. As shown in Fig. 7 (filled brown box), the SALD derived from matchups of the first profiles of floats 29035 and 29034 with shipboard CTD observation (JNSR_200074) clearly identifies the observed error in profiles of float 29035 by SALD. Because there is no shipboard CTD observations for substantiating the observed error in the initial profiles of float 2900238 by SALD, comparative analysis of float-to-float matchups involving three floats from the same region (2800222, 2900238, and 2900240) are shown in Fig. 7. The large mean SALD values for matchups involving 2900238 (shown in filled blue box of Fig. 7), compared to a matchup not involving 2900238, show existing errors in the initial profiles of float 2900238 as identified by the SALD method. Table 4 shows a comparison of the SALD method and ADMQC in identifying erroneous salinity from the SNP. Both the ADMQC and SALD methods identified six common floats (serial numbers 1–6 in Table 4) that show salinity errors. While for three floats (2900142, 2900492, and 2900504) SALD could not identify the existing errors, ADMQC could not identify errors existing in floats 29035 and 2900238, which were detected by SALD.
All four selected floats for drift analysis show drift and biases, both negative and positive. Negative drift can be expected by fouling–biofilm formation inside the cell and the resulting change in the conductivity. However, in the present study, except for float 2900525 three floats (2900438, 2900440, and 2900441) show positive drift. Therefore, the cause(s) of the observed drift is not always the same. This is consistent with previous observations from shorter records (Oka 2005; Janzen et al. 2008) and in a different basin (Gaillard et al. 2009). The results obtained in the present study from longer records show very convincingly that the mean linear drift (of the order of 0.0009 yr−1) is very minimal, and a cumulative drift for 5 yr (0.0045 PSS) will not be a major concern. However, drift and biases together may cross the set threshold of 0.01 PSS by the Argo program. Therefore, it is important to find out possible biases of each float and correct it in the delayed-mode dataset.
Although delayed-mode data have been used in this study, comparison with shipboard CTD has revealed that significant error remained undetected in delayed-mode data. This shows the need for improving the existing objective methods used by DACs. SALD could be used as one of the standard estimates in the Argo quality control scheme before subjecting to other objective methods. The efficiency of SALD in identifying drift is demonstrated in Fig. 3e. As illustrated in Fig. 4 (SOJ) and Fig. 6a (SNP), the SALD derived from float-to-float matchups also could be used to identify erroneous profiles. Because SALD depends on salinity structure at deeper θ surfaces, errors identified by this method could be depth independent. For example, the observed salinity drift in Fig. 3e is depth independent. Therefore, SALD may not be an efficient tool in identifying random errors at depths above the reference θ. Also, in areas where strong salinity fronts exist, the SALD method may have difficulties in detecting errors. From the above discussion, it is implied that SALD could be used as an effective additional tool for Argo data quality control tool with the existing ADMQC method.
8. Summary and conclusions
Long-term sensor drift in the conductivity cells of float that completed 5 yr are examined using the near-linear subsurface salinity structure in the Sea of Japan (SOJ) and the new method of SALD. SOJ provides a better facility to validate Argo salinity to an accuracy of 0.001 PSS. It has also the highest float/profile density in the global ocean. The SALD derived from matchups between the four floats (completing 5–6 yr in the SOJ) and shipboard CTD observations from the SOJ within the defined time–space decorrelation scale shows linear drift. Although all four floats show linear drift (<0.001 PSS yr−1), it is encouraging to note that those are insignificant to affect the objective of the Argo program in understanding the climate variability. Nevertheless, if added to biases, these drifts could be of concern.
Applicability of SALD in open ocean regions (which are different from the confined basin of SOJ) was also examined by applying the SALD method in SNP. SALD derived from Argo matchups from the SNP could identify most of the erroneous floats flagged by the ADMQC method. In addition, SALD could also identify salinity error in some floats, which are flagged for good quality by ADMQC. As illustrated in this study, SALD is a simple and powerful tool to identify drift, biases, and erroneous profiles. It has the added advantage of delineating water mass changes from drift. Different oceanic regions have different decorrelation scales of time and space. With the availability of good quality background data, it is possible to define the decorrelation scales and then, considering the less variable region of the water mass (within the Argo depth range), SALD could be used to check the data quality. Because SALD makes use of subsurface salinity structure, it may not capture existing random errors in data from upper layers.
The Argo data were collected and made freely available by the International Argo Project (http://www.usgodae.org/). Argo is a pilot program of the Global Ocean Observing System. Shipboard CTD data used in this study were downloaded from the Japan Oceanographic Data Centre (http://www.JODC.go.jp). We thank anonymous reviewers for constructive suggestions that helped to improve the quality of the paper.
National Institute of Oceanography Contribution Number 5037.