Measurement Biases in Ocean Temperature Profiles from Marine Mammal Dataloggers
1. Introduction
Regional, spatial, and temporal coverage of temperature and salinity of the global ocean is very diverse and still shows regions of inadequate resolution. Within the last decades, the ship-based measurements have been augmented by data from unmanned platforms. Data from autonomous profilers (Argo floats) have significantly improved global data coverage and added important information in regions undersampled because they were not visited by ships. More recently, an important contribution particularly in high-latitude and polar regions at almost all seasons comes from instruments attached to marine mammals. For improving global and regional heat and freshwater budgets, they need to be incorporated into global hydrographic datasets.
Dataloggers attached to marine mammals represent an instrumentation type which provided a significant amount of ocean temperature profiles since the 2000s, particularly where the sampling by Argo floats is limited or absent. In the World Ocean Database 2018 (WOD18) under the category autonomous pinniped bathythermograph (APB), marine mammal data contribute to ∼61% of temperature profiles south of 50°S for 2005–18, ∼42% of temperature profiles north of 40°N. Two main logger models have contributed to APB: conductivity–temperature–depth Satellite Relay Data Loggers (CTD-SRDL) manufactured by Sea Mammal Research Unit (SMRU) Instrumentation and time–depth recorders (TDRs) manufactured by Wildlife Computers. In this study, we examine the quality of the temperature profiles by comparing them with the more accurate CTD and Argo data. The overlap between the CTD-SRDL temperature data in WOD18 and in the database distributed by the Marine Mammal Exploring the Ocean Pole to Pole (MEOP) consortium was found to be incomplete, so the most recent version of the MEOP-CTD database is used along with the data from the World Ocean Database missing in the MEOP collection. TDR data, mostly found in the North Pacific sector, represent the majority of APB profiles (55%) because they are provided at higher temporal frequency. Since both recorders contribute a significant part of all temperature profiles available in the World Ocean Database, proper estimation and elimination of possible instrumental biases are necessary for improving the global ocean heat content time series. Using the collocated temperature profiles from the ship-based CTD and Argo floats, we calculated corrections for instrumental biases in both recorder types.
2. Oceanographic data from dataloggers attached to marine mammals
a. General description
Since the late 1990s, scientists started to equip marine mammals with oceanographic dataloggers. The pioneering work of the autonomous pinniped environmental sampler (APES) demonstrated the feasibility of the approach (Boehlert et al. 2001). The first datalogger developed during the 1990s could measure temperature and pressure only (therefore the original name “pinniped bathythermographs”), and the instrumentation and handling are also described in detail by Boehlert et al. (2001). A fast technological progress led to the rapid development of compact CTDs tested on fur seals and whales (Hooker and Boyed 2003; Lydersen et al. 2002). A major breakthrough was made in 2003 when the CTD-SRDL was developed by the SMRU (St Andrews, United Kingdom), incorporating a high precision CTD unit manufactured by Valeport Ltd. (Boehme et al. 2009; Fedak 2004). The first large-scale deployments of CTD-SRDL started in 2004, as part of the program Southern Elephant Seals as Oceanographic Samplers (SEaOS), that led a few years later to a large number of datalogger deployments from different national teams coordinated within the consortium MEOP (Biuw et al. 2007; Charrassin et al. 2008; Treasure et al. 2017). In parallel, temperature-only TDR were developed by Wildlife Computer and used extensively as part of the project Tagging Of Pacific Predators (TOPP) (Block et al. 2011) in the North Pacific Ocean.
Apart from the importance for genuine biological studies, the temperature profiles from the dataloggers have a special value for oceanographic applications as they are numerous and cover the areas which are less frequently sampled by other hydrographic instruments (Charrassin et al. 2008). For instance, in the Southern Ocean, the CTD-SRDL data effectively fill the gap in areas, which are not sampled both by classical ship-based hydrography and by the Argo floats (Roquet et al. 2013). The maximum depth of the profiles depends on the respective mammal, with some species able to reach the maximum depth of almost 2000 m. The temperature and salinity measurements are recorded during their dive, with the typical position accuracy of ±5 km. Most of the CTD-SRDL profiles are postprocessed, including editing, correction, and calibration. The correction includes a method to reduce thermal cell effects and density inversion, as well as adjustments of temperature and salinity data (Mensah et al. 2018; Siegelman et al. 2019). Temperature offsets can be estimated in the case where seals foraged in freezing cold waters, using the known freezing temperature as a reference (Siegelman et al. 2019). The accuracy of the postprocessed measurements for the loggers manufactured after 2008 was estimated to be ±0.03°C in temperature and ±0.05 or better in salinity (Roquet et al. 2014). Ship-based CTD comparisons indicate that pressure measurements are within the range provided by the manufacturer (<5-dbar deviation at 1000 m; Roquet et al. 2011). The MEOP-CTD database provides the largest collection of quality-controlled CTD-SRDL profiling data (Treasure et al. 2017; McMahon et al. 2021). According to Boehlert et al. (2001), TDR loggers manufactured by Wildlife Computers measured temperature and depth with a temperature resolution of 0.1°C and accuracy of 0.5°C. Together, different kinds of dataloggers attached to marine mammals contributed about 2.5 million temperature profiles to the World Ocean Database.
b. World Ocean Database and MEOP collections of marine mammal data
Oceanographic data from marine mammal dataloggers reside in two big collections. The WOD18 is the largest archive of hydrographic data, including temperature profiles of various instrumentation types (Boyer et al. 2018). Marine mammal data can be found within the WOD under the instrumentation type APB, starting in 1998 with APES data along the North American Pacific coast. The second large archive of marine mammal data is the MEOP collection which includes data collected between 2004 and 2019 by the MEOP consortium.
Note that the two databases partly overlap, as WOD18 incorporates an older version of the MEOP-CTD database. So we will use predominantly the MEOP-CTD database (version 2020-01-02) for CTD-SRDL data and WOD18 database for TDR data. Yearly number of temperature profiles is shown in Fig. 1a.
(a) Yearly number of TOPP temperature profiles; (b) the total number of TOPP profiles in 1° × 2° latitude/longitude boxes; (c),(d) as in (a) and (b), but for the MEOP CTD-SRDL profiles (the red line indicates the percentage of MEOP profiles also available in the WOD database); (e),(f) as in (a) and (b), but for non-MEOP CTD-SRDL temperature profiles within the WOD database. Color lines show average positions of the following fronts in the Southern Ocean: red, Subantarctic Front; magenta, Antarctic Polar Front; dark blue, Southern Boundary.
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
Comparing coordinates and the dates of the profile occupation, we find that 78.3% of MEOP profiles are also found in the WOD archive. The percentage varies between 70% and 100% between 2004 and 2017 dropping to zero for the year 2019 (Fig. 1c). WOD profiles which are duplicates to the profiles in the MEOP collection were eliminated in the current analysis in favor of the MEOP profiles. The retained WOD profiles are distributed among the TDR profiles in the northwest Pacific Ocean (Figs. 1a,b) and CTD-SRDL profiles in both hemispheres (Figs. 1c–f). Due to the high sampling rate, the TDR profiles provided a dense spatial coverage in the North Pacific Ocean between 2004 and 2010. The CTD-SRDL temperature profiles provided a good spatial coverage in the entire Southern Ocean south of the Subantarctic Front, in the North Pacific Ocean north of about 35°N, and in the part of the Nordic seas in the Atlantic Ocean. Profiles in the northeast Pacific are much more numerous compared to the Southern Hemisphere, due to the fact that the TDR data are provided at high daily frequency (∼30 profiles day−1). Unlike TDR, the SRDL tags report profiles only from the deepest dive within each 6-h period. SRDL profiles are most abundant in the Indian sector of the Southern Ocean around Kerguelen Islands and in the shelf and continental slope regions between 100°–30°W and 30°–170°E. The spatial pattern correlates with the locations of the marine mammal colonies.
Figures 1e and 1f indicate that there are quite a few CTD-SRDL profiles missing in the MEOP collection. Two factors may explain this: 1) some profiles available only in WOD18 may have been discarded from the MEOP collection during the postprocessing; 2) some data reached WOD18 through the Global Telecommunication System (GTS) but were not included in the MEOP version used for this study. The large drop in the MEOP profile percentage in 2018 corresponding to the largest number of CTD-SRDL profiles unique for the WOD database is consistent with the latter point.
c. SRDL sensors
Temperature profiles from the MEOP collection were obtained by means of several sensor models (tags). Most of them have been manufactured by Valeport Ltd. and include VP1, VP2, and VP3 models. VP1 was the first model with low accuracy, VP2 was an intermediate model which had some technological issues, and VP3 corresponds to the standard model. CTF is a new model incorporating a fluorescence sensor next to the CTD cell. The tag model is unknown for 15.39% of all MEOP profiles. Most of these profiles correspond to the Canadian data in the Labrador Sea area; 78.21% of all profiles were obtained by means of the VP3 sensor, whereas the older VP1 and VP2 sensors contribute 3.99% and 1.59%, respectively. The CTF sensor represents the smallest group of profiles with 0.83%. The total number of MEOP profiles used for the current analysis amounts to 588 358.
d. Distribution of profiles between different mammal species
There are at least 19 species of mammals to which the tags have been attached. The species live in different geographical areas and are characterized by a different maximum dive depth (between several 100 m and almost 2 km). The dive velocity may differ between the species. Suzuki et al. (2014) estimated the diving speed for captive Steller sea lions to be around 1.5–2.0 m s−1. Elephant seals ascend at a similar >1 m s−1 vertical speed (Mensah et al. 2018). Taking into account the diving speed might be important in case of a time lag in datalogger response, although Mensah et al. (2018) did not find a substantial sensitivity to diving speed when they estimated thermal lag correction coefficients.
3. Reference data
For the calculation of the systematic offsets in marine mammal data, reference data are required. Argo and CTD temperature profiles are used as reference bias-free data in this study. The typical accuracy of the CTD temperature data is about 0.003°C. The accuracy of the Argo float data is lower than that of the ship-based CTD, estimated to be about 0.01°C. Whereas the ship-based CTD temperature profiles are typically available along a coarse set of hydrographic sections, Argo float profiles are distributed much more evenly over the World Ocean and provide a much denser net of profiles. In addition, the ship-based temperature profiles are less numerous in polar regions. These two factors result in 96.6% of all reference profiles being those from the Argo floats.
4. Data quality control
All data used for the study were quality controlled using the procedure developed in Gouretski (2018). According to the comprehensive intercomparison study by Good et al. (2022) of several quality control procedures, this procedure demonstrates a good performance effectively rejecting the data outliers and retaining good data.
The quality control procedure was applied both to the marine mammal data and to the collocated reference Argo and CTD profiles (see profile distribution maps in Figs. 1c,d) between 2004 and 2019. Outlier percentage for different instrumentation types are shown in Fig. 2. A suite of quality checks results in a low total percentage of outliers for WOD CTD (0.98%), WOD profiling float (WOD PFL; 0.47%), MEOP (2.99%), and APB (3.77%) profiles. Both reference data types demonstrate the lowest percentage of outliers. The original WOD APB and the WOD PFL data indicate a higher percentage of crude errors corresponding to the unrealistically high and low temperature values. In contrast, the MEOP collection is characterized by a lower percentage of crude outliers indicating the impact of the preliminary quality control. The WOD APB data also indicate a higher outlier percentage between 2004 and 2008 due to the less accurate TDR profiles. Both the MEOP and the WOD APB data indicate a higher outlier percentage for the years 2006 and 2011–12, probably related to the same quality issues as both collections overlap for SRDL recorders. The presence of outliers in PFL data is partly explained through the fact that a part of Argo profiles within the WOD is represented by the real-time data which have not undergone quality control at Argo data acquisition centers.
Temperature outlier percentage for (a)–(d) MEOP, (e)–(h) APB, (i)–(l) CTD, and (m)–(p) PFL profiles. Percent of outliers in (a),(e),(i),(m) year–depth bins; (b),(f),(j),(n) temperature–depth bins; (c),(g),(k),(o) year–temperature bins; and (d),(h),(l),(p) 1° × 1° latitude/longitude boxes.
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
5. Calculation of temperature bias
For the estimation of possible temperature bias in marine mammal data, we use the so-called collocation method, which was previously applied to diagnose biases in expendable bathythermograph data (Gouretski and Reseghetti 2010; Cheng et al. 2016), mechanical bathythermograph data (Gouretski and Cheng 2020), and historical Nansen bottle data (Gouretski et al. 2022). In this method, biased temperature profiles are compared with the reference (unbiased) profiles within a certain spatial and temporal collocation bubble. The choice of collocation bubble size (e.g., the maximum distance to the reference profile and the maximum time difference between profile occupations) aims both to keep the bubble as small as possible and to guarantee the sufficient amount of collocated profile pairs to obtain stable statistics.
a. Method of bias calculations
b. Selection criteria for collocated pairs
Based on the collocation analysis for XBT, mechanical bathythermograph (MBT), and Nansen bottle profiles (Gouretski and Reseghetti 2010; Gouretski and Cheng 2020; Gouretski et al. 2022), the initial search of collocated reference data for each MEOP profile is done within the radius of 150 km and for the time difference between the two profiles less than 45 days. We note that the Southern Ocean is characterized by the pronounced zonation, with strong fronts dividing the waters with different characteristics (Deacon 1937; Orsi et al. 1995). From the north to the south these fronts are the Subantarctic Front, the Antarctic Polar Front, the Southern Antarctic Circumpolar Current (ACC) Front, and the Southern Boundary. The frontal divisions can significantly exaggerate the apparent temperature bias if MEOP and reference profiles are located on different sides of the frontal division. Therefore, in case of multiple reference profiles available for one and the same MEOP profile, only one profile most similar to the MEOP profile is selected. As the characteristic of the profile similarity, we use the depth-averaged absolute temperature difference between the MEOP and the reference profile (we call it hereafter the profile similarity index) with the similarity index histogram shown in Fig. 3. Finally, the profile similarity index for the selected pair is compared with the threshold value to retain or reject the selected collocated pair. Using the most similar profile pairs reduces the noise due to the meso- and small-scale variability and the respective error of the temperature bias estimate. The threshold value of 1.5° for the similarity index corresponds to the rejection of 5% of profile pairs.
Profile similarity index histogram and the threshold value (green line).
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
To further reduce the noise due to imperfect collocations of MEOP and reference profiles additional selection thresholds are imposed:
-
Taking into account strong topographic control of the currents (and fronts) in the Southern Ocean, only collocated pairs with the bottom depth difference ΔH not exceeding the following threshold are retained: ΔHmax = Hmeop × 0.5, where ΔHmax is the maximum permitted bottom depth difference and Hmeop is the bottom depth at MEOP location.
-
A considerable fraction of MEOP profiles is located on the continental shelf or slope, whereas reference Argo floats are located mostly in deeper regions. To avoid artificial bias due to this difference in the sampling pattern, only MEOP profiles with bottom depth exceeding 1000 m are retained for the analysis.
Spatial distribution of the finally selected collocated profile pairs for different sensor types is shown in Fig. 4. VP3 sensor represents the most abundant group of collocated profiles. Almost all of the unknown sensor profiles are found in the Northern Hemisphere, and the CTF sensor profiles have the more limited geographical coverage compared to other sensors.
Number of MEOP profiles in 1° × 1° boxes collocated with Argo/CTD profiles for distinct sensors: (a) VP1, (b) VP2, (c) VP3, (d) CTF, (e) unknown sensor, and (f) all sensors.
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
The percentage of collocated pairs from the total number of MEOP profiles between 2004 and 2019 varies in the range 7%–53% (Fig. 5), with the average value of 26% for the entire period and is better than 20% for the years after 2011 which is explained by the growing number of the reference Argo float data.
Yearly percentage of MEOP profiles having collocations with reference profiles.
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
As noted above, the difference in spatial patterns between MEOP and collocated reference profiles can potentially have undesirable effects on the bias estimate. If MEOP profiles are located predominantly to the south/north of the reference profiles, the artificial cold/warm MEOP bias could arise. As demonstrated in Fig. 6, MEOP profiles more frequently (52.6% of all pairs) have their location to the south of reference profiles. To reduce possible effects of this sampling feature on bias estimates, the resulting bias is calculated as the average of two separate estimates: the one is based on pairs where MEOP profile is located to the south of the reference profile and the other is based on the pairs with MEOP profiles located to the north of the reference profile. The predominance of the cases where MEOP profiles are located to the south of Argo profiles (more frequently embedded into the current cores related to front divisions) may be linked to the specific feeding behavior of the marine mammals.
Histograms of the latitudinal difference between MEOP and reference profiles south of 30°S. Total number of profiles with negative and positive latitudinal difference is shown on the left and right side, respectively.
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
c. Sensor-type attribution for profiles with unknown sensor type
As illustrated by Fig. 4, marine mammal temperature profiles are distributed between several models of sensors, with sensor model attribution being not possible for a certain percentage of MEOP profiles (unknown sensor type). Based on the yearly number of profiles for distinct sensors (Fig. 7), we attributed unknown sensor profiles to the VP1 type for the year 2004, to the VP2 type for the years 2005–06, and to the VP3 type for the years 2007–19. Since the temperature sensor CTF has identical characteristics as VP3, the CTF profiles were blended with profiles from the VP3 type. The same attribution was made for the CTD-SRDL profiles from the WOD archive which are not present in the MEOP collection.
Yearly number of profiles from different sensors: (a),(f) VP3; (b),(g) CTF; (c),(h) VP2; (d),(i) VP1; and (e),(j) unknown sensor. (left) Northern Hemisphere; (right) Southern Hemisphere. Sensor-type attribution for the unknown sensor group is indicated by the respective color.
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
6. Results: Diagnosed temperature bias
a. Biases for distinct mammals
MEOP temperature profiles which have collocations with reference data are distributed among 19 mammal species each bearing the SRDL recorder. A total of 678 animals (e.g., distinct tags) contributed with temperature profiles to the collocated dataset. According to the histogram of the profiles per mammal (Figs. 8c,f,i), the median number of profiles from each tag is around 150 for VP3/CTF profiles. Because of the noise from the mesoscale variability, this amount of profiles is not sufficient for the estimation of a temperature offset vertical profile for each individual tag. Therefore, we calculated depth-averaged offsets for distinct tags within three layers (Figs. 8a,d,g): 0–300 m (in this layer, offset changes with depth are most pronounced), 300–1000 m (here, the bias remains rather constant), and for the entire water column between 20 and 1000 m.
(a),(d),(g) Layer-averaged temperature biases for distinct mammals vs time (sensor group VP3/CTF). Number of temperature profiles is shown by color. (b),(e),(h) Temperature bias histograms with the number of mammals which with negative (NB<0) and positive (NB>0) biases, respectively. (c),(f),(i) Histograms of number of profiles per mammal. (a)–(c) Layer 20–1000 m; (d)–(f) layer 20–300 m; (g)–(i) layer 300–1000 m; and (j)–(r) as in (a)–(i), but for TDR profiles. Green lines in (a), (d), (g), (j), (m), and (p) correspond to 1% rejection limits for the lowest and highest biases.
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
The offsets in each layer are predominantly negative, which is also supported by the respective histograms (Figs. 8b,e,h). The deep layer is characterized by the mean negative bias of −0.033°C. The mean bias within the upper 20–300 m is −0.013°C. This is due to the complicated vertical structure of the bias profile in this layer, which is described in the next section. After exclusion of one percent of tags with lowest and highest biases, we find that the biases for distinct tags vary between −0.35° and 0.25°C (Fig. 8a), suggesting a stable performance over the time period under consideration.
In contrast to the VP3/CTF sensor group, TDR temperature profiles are characterized by the predominantly positive biases (Figs. 8j,m,p), with the overall mean bias of 0.097°C for the layer 300–1000 m. The median number of temperature profiles per mammal is about 1500, reflecting the much higher sampling rate compared to the SRDL sensors.
b. Overall temperature biases for SRDL and TDR profiles
Figure 9a summarizes the results of the overall temperature bias calculations at depth levels with bias estimates based on all collocated pairs regardless of the year of observation. Here, we use the blended MEOP/WOD set of temperature profiles after the attribution of sensor types (see section 5c) for the profiles with unknown sensors. The horizontal bars at each level show the respective standard error, where the number of distinct tags is taken as the number of degrees of freedom. For each 5-m level, the overall mean vertical temperature gradient is also shown (Fig. 9b). The gradient is calculated over 10-m layers using the temperature values 5 m above and below the analyzed level, with depth values positive downward. The number of collocated pairs decreases with depth, with profiles reaching at least 800-m depth comprising only 11% of all profiles having collocations with reference data.
(a) Overall temperature bias for different sensors with standard error bars; (b) respective mean vertical temperature gradient; (c) number of collocated pairs at depth levels.
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
The overwhelming majority of collocated profiles from the marine mammals is linked to just two sensor types: VP3 and CTF sensors for SRDL profiles and profiles from the TDR recorders used during the TOPP experiment (Fig. 9c). Respectively, the uncertainty in bias is smaller for these sensor types compared to the sensors VP1 and VP2.
The VP1 sensor overall offset is negative over the whole water column, whereas the VP2 model exhibits the change from the positive bias above 350-m levels to the negative bias below that level. The newer and improved models VP3 and CTF are characterized by a negative overall bias at all levels except for the layer 110–210 m with a rather stable bias profile below about 350 m.
TDR recorders are characterized by the overall temperature bias which is positive over the entire depth range. The bias increases from 0.02°C between the surface and 250-m depth, varying between 0.07° and 0.11°C below this level. The possible causes for the diagnosed temperature biases are discussed in section 7.
c. Time variations of temperature bias
To assess the time variability, monthly vertical bias profiles for the years between 2004 and 2019 were calculated using all collocated pairs within the 3-yr time window centered at the respective month (Figs. 10a–c). The biases are shown only for the levels with at least 1000 collocation pairs. Also shown are the depth/time sections of the vertical temperature gradient (Figs. 10e–h), water temperature (Figs. 10i–l), and the number of collocated pairs (Figs. 10m–p).
(a)–(d) Monthly temperature bias for VP1, VP2, VP3/CTF, and TDR sensors; (e)–(h) as in (a)–(d), but for the vertical temperature gradient; (i)–(l) as in (a)–(d), but for water temperature; and (m)–(p) as in (a)–(d), but for the number of collocated pairs. Calculations are done using data within the 3-yr time window centered at the respective year and month.
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
The smallest amount of collocated pairs is available for the VP1 sensor, which is characterized by a negative bias over the whole depth range (Fig. 10a). For the VP2 sensor, temperature offsets are predominantly positive for the years 2005–06, turning to negative values for the years 2007–08 (Fig. 10e). The time evolution of the temperature bias for the VP3/CTF sensor group can be traced for the years after 2006 (Fig. 10i), with the persistent positive bias found in the layer between 50 and 300 m. This layer is located within the envelope of positive vertical temperature gradient values (see Figs. 9b,f,j).
In spite of the fact that on average an order of magnitude less collocated pairs are available for each calendar year, the vertical structure of the yearly biases is qualitatively similar to the overall bias profile, indicating a rather constant negative temperature bias below about 250–300 m (e.g., a pure thermal bias) and a characteristic layer of positive biases between 100 and 300 m. The mean water temperature calculated over the respective collocation pairs (Figs. 10c,g,k) varies between 0° and 4°C indicating no correlation with the derived bias. The number of available collocated pairs is shown in Figs. 10m–p and differs significantly among the sensor types.
Figure 11a provides a more detailed view of the bias change between 2007 and 2019 for the VP3 sensor. Within the layer 300–700 m, the bias ranges between −0.045°C (for the year 2007) and 0.005°C for the year 2013. The largest variations are confined within the layer 100–200 m, which also corresponds to the layer with the strongest change of the vertical temperature gradient (Fig. 11b). The yearly vertical profiles of the TDR temperature bias (Fig. 11c) exhibit a rather similar vertical structure within the layer below about 150-m level, where the changes of the vertical temperature gradient become small (Fig. 11d). In this layer, the bias varies within the range 0.06°–0.12°C, with the exclusion for the year 2010 which is characterized by the lowest bias values.
Yearly vertical profiles of the (a) temperature offset and (b) vertical temperature gradient for the VP3/CTF sensor group; (c),(d) as in (a) and (b), but for the TDR sensor. Calculations are done using data within the 3-yr time window centered at the respective year.
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
d. Spatial pattern of temperature bias
To investigate the spatial bias patterns for the SRDL recorders (VP3/CTF sensor group), the mean yearly temperature offsets within the layer 350–700 m were calculated for 4° × 8° latitude/longitude geographical boxes (Fig. 12). The scope of the spatial sampling exhibited significant changes over the 13-yr period with the number of 4° × 8° boxes sampled varying between 176 in 2008 and 71 in 2017. There are also changes in the geographical pattern. For instance, in the northwest Atlantic Ocean, there are no collocated pairs in the years 2012–17. In the Southern Ocean, collocated pairs are absent in the Pacific sector between 2015 and 2019, with the Atlantic sector having a good coverage only for the years 2008–10. In spite of the changing sampling pattern, all maps (except for the year 2018) show the prevalence of boxes with the mean negative offset. For the entire time period between 2007 and 2019, 65% of boxes exhibit mean negative bias. The maps do not indicate persistent patterns of negative or positive biases in certain geographical areas.
Mean yearly temperature offsets for the VP3/CTD sensor group within 4° × 8° latitude/longitude boxes and for the layer 350–700 m, for the years (a) 2007; (b) 2008, (c) 2009, (d) 2010, (e) 2011, (f) 2012, (g) 2013, (h) 2014, (i) 2015, (j) 2016, (k) 2017, (l) 2018, (m) 2019, and (n) average 2007–19. In the middle of each map, the number of boxes with negative (blue) and positive (red) bias is shown.
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
The significant sampling pattern change over the analyzed time period (Figs. 12 and 13) may be one of the possible causes for the apparent bias change between the years. As noted by one of the reviewers, other factors like changes in sensor selection and calibration, change of the position of the temperature probe on the tag, and the kind of thermal contact with the tag can also contribute to the apparent bias variation over time.
Mean yearly temperature offsets for the TDR sensor within 2° × 4° latitude/longitude boxes (years are shown in the upper-left corner of each map) and for the layer 200–600 m: (a) 2004; (b) 2005, (c) 2006, (d) 2007, (e) 2008, (f) 2009, (g) 2010, and (h) 2004–20. The number of boxes with negative (blue) and positive (red) bias is shown on Alaska.
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
Similar calculations of the box-averaged biases were conducted for the TDR sensor for the period 2004–10 (Fig. 13). The smaller 2° × 4° boxes are used because of the denser sampling. The box-averaged biases are calculated for the layer 200–600 m where the vertical bias profile is less variable (see Figs. 11c,d).
Similar to MEOP, the spatial sampling during the TOPP project exhibits variation with the number of sampled 2° × 4° boxes ranging between 46 in 2004 (Fig. 13a) and 106 in 2007 (Fig. 13d). For all years, box-averaged biases are predominantly positive in agreement with the overall mean bias profile (Fig. 9a). The percentage of boxes with positive offsets varies between 65% in 2009 and 84% in 2006. We conclude that the yearly maps of box-averaged biases are in agreement with the overall bias structure described in section 6b, with VP3/CTF and TDR tags exhibiting, respectively, negative and positive bias.
e. Consistency of the reference database
For the calculation of the systematic offsets in marine mammal data, reference data are required. Argo and CTD temperature profiles are used as reference bias-free data in this study. The typical accuracy of the CTD temperature data is about 0.003°C. The accuracy of the Argo float data is lower than that of the ship-based CTD, estimated to be about 0.01°C. Unlike the ship-based CTD, Argo sensor cannot be calibrated in the laboratory.
To assess the degree of consistency of the Argo/CTD blended reference dataset, we selected Argo/CTD collocated pairs falling within 1° × 2° latitude/longitude squares where SRDL profiles collocated with Argo or/and CTD pairs are also available (Figs. 14a–c). Based on these Argo/CTD collocated pairs, Argo temperature offset versus CTD data was calculated using the same method as for SRDL profiles (Fig. 14d). Similar to the vertical profile of the CTD-SRDL bias, the Argo-CTD temperature offset exhibits larger variability in the upper part of the water column implying impact of the strong temperature gradient above 150 m (Figs. 14d,e), with the absolute offset values less than 0.008°C below the depth of about 350 m. The offset spread in the deeper part of the water column can be interpreted as the instrument-related uncertainty due to the possible bias in Argo data. It should be noted that the number of collocations with CTD profiles both for Argo and CTD-SRDL data is at least an order of magnitude less compared to the number of SRDL profiles collocated with Argo or/and CTD profiles (Fig. 14f). However, below the 350-m level, SRDL offsets relative to Argo profiles and relative to CTD profiles agree within about 0.01°C. Based on the above results, we estimate the possible uncertainty in the diagnosed CTD-SRDL temperature bias to be the order of 0.01°C.
Number of collocated pairs in 1° × 2° latitude/longitude boxes: (a) VP3 profiles collocated with CTD and/or ARGO, (b) VP3 profiles collocated with CTD profiles, (c) Argo profiles collocated with CTD profiles, (d) the corresponding overall temperature offsets, (e) mean vertical temperature gradient, and (f) number of collocated pairs at depth levels.
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
7. Possible causes for diagnosed biases
To explain the cause for the diagnosed temperature bias, we consider the derived biases and the accompanying vertical temperature structure for the two most numerous groups of the recorders: the CTD-SRDL recorders equipped with the VP3/CTF sensor (Figs. 15a–d) and the TDR recorders used during the TOPP experiment (Figs. 15e–h). Apart from the pure thermal biases for the temperature sensor, two other factors can theoretically cause temperature bias in case of a nonzero vertical temperature gradient: a time lag in temperature sensor response and a systematic error in pressure sensor. Since the temperature profile is obtained during the ascent phase of the mammal dive, the nonzero sensor time lag would contribute with a negative/positive addition to the bias estimate in case of a negative/positive temperature gradient, respectively. Similarly, pressure overestimation/underestimation by the pressure sensor would result in a warm/cold temperature bias in the case of a negative vertical temperature gradient (with the opposite association in the case of a positive vertical temperature gradient). As noted in section 2a, we used the postprocessed MEOP data, with the data processing method accounting for thermal cell effects, density inversions, and adjustments of temperature and salinity data. The issue on the possible sensor time lag requires further consideration and is discussed in section 7a.
(a) Overall mean temperature bias (VP3/CTF sensor group) for all collocated pairs (green), pairs from the warm year half (red) and pairs from the cold year half (blue); (b) as in (a), but with the thermal bias (−0.025°C) subtracted and the depth levels shifted 2 m downward; (c) as in (a), but for the mean vertical temperature gradient; (d) as in (a), but for the mean temperature; (e) as in (a), but for the number of collocated pairs; (f)–(j) results for the TDR sensors with (g) showing temperature offset after reducing original depths by the factor of 0.940.
Citation: Journal of Atmospheric and Oceanic Technology 41, 7; 10.1175/JTECH-D-23-0081.1
Figures 15a–e show derived temperature biases (VP3/CTF sensors) and the respective thermal structure for three groups of collocated profiles: 1) all pairs, 2) pairs from the warm year half (December–May), and 3) pairs from the cold year half (June–November). To mitigate the effect of the different geographical sampling for the warm and the cold year halves, the selection of profiles was restricted only to those 2° longitude × 1° latitude boxes where profiles from both year halves were available. Figures 15c–d indicate a complicated thermal structure which is characteristic for the Southern Ocean (Gordon and Huber 1990), where the majority of collocated pairs for the group VP3/CTF comes from. The two-layer thermal structure is typical for the winter, when the cold Winter Water (formed during the winter convection process with temperatures close to the freezing point due to the ice building) overlies the warm Circumpolar Deep Water. In summer, the thermal structure is characterized by three distinct layers: a warmer near surface layer, the underlying remnant layer of the cold Winter Water, and a warmer layer of the Circumpolar Deep Water below (Fig. 15d). These seasonal temperature changes are confined to the upper 200-m layer and lead to the profound changes in the vertical temperature gradient (Fig. 15c). These changes can be linked to the vertical structure of the diagnosed temperature bias within the upper 200 m (Fig. 15a).
The seasonal changes of the thermal structure of the North Pacific area (Figs. 15h,i) covered by the TDR profiles during the TOPP experiment are mostly confined to the upper 100-m layer with the vertical temperature gradient remaining negative throughout the year (Fig. 15h). In contrast to MEOP data, there is a considerable difference in the number of TOPP temperature profiles available for the warm and the cold year halves (cf. Fig. 15e and Fig. 15j).
a. Temperature and depth corrections for CTD-SRDL profiles
Between 200- and 800-m depth, the diagnosed bias (Fig. 15a) based on all collocated pairs ranges between −0.005° and −0.034°C, with both the mean temperature and the mean temperature gradient exhibiting just negligible seasonal changes (Figs. 15c,d). For this layer, the profile suggests a rather constant mean bias value of −0.027°C. The largest bias variability over depth is restricted to the upper 200-m layer. The maximum positive bias value of about 0.015°C near the 190-m level corresponds to the temperature gradient maximum of about +0.005°C m−1. Within the upper 100-m layer, the bias is confined to the range from −0.28° to −0.20°C, close to the mean value for the deeper part of the water column. Bias profiles for the warm and cold year halves follow closely the profile based on all pairs up to the level of 120 m (Fig. 15a). Within the upper 100 m, biases for warm and cold year halves exhibit a difference of 0.05°C.
We investigated the impact of the thermal lag possibly remaining in the postprocessed data. Figure 15b shows bias profiles after applying the following corrections: 1) accounting for the negative thermal bias of −0.027°C and 2) accounting for the time lag of 1 s, corresponding to the vertical downward displacement of depth levels by 2 m if the vertical upward velocity of 2 m s−1 is assumed. Taking account for the time lag reduces the range of bias variability throughout the water column for all three groups of collocated pairs. For the all-pairs case, the implementation of the two corrections results in the absolute bias reduction everywhere except for the layer 115–205 m, where the average absolute bias for the corrected profiles is larger than the absolute bias for the original profiles by about 0.015°C. The standard deviation of the overall bias for all pairs within the layer 0–200 and 200–800 m is 0.015° and 0.005°C, respectively. The downward shift of the observed levels leads to the reduction of the bias standard deviation to 0.011°C in the upper and to 0.003°C in the deeper layer. We note that the downward level displacement imitating a possible logger time lag does not fully remove the characteristic structure of the diagnosed bias profiles within the upper 200-m layer.
As discussed in section 6c, our results indicate a certain time variability of the temperature bias for SRDL recorders (Figs. 10a–c and 11a). However, considering significant changes in geographical sampling pattern from year to year (Fig. 12), the small magnitude of the diagnosed bias, and the absence of the metadata on possible technical changes during the tag manufacturing process, we suggest applying a constant thermal correction which is equal to the overall temperature bias within the layer 200–800 m. This layer corresponds to the most homogeneous part of the water column with typical vertical gradients less than 0.001°C m−1 (Fig. 11b). The bias corrections for different sensors of SRDL recorders are given in Table 1, with all sensors exhibiting a negative bias: −0.027° ± 0.044°C for VP1, −0.007° ± 0.023°C for VP2, and −0.027° ± 0.007°C for VP3/CTF sensor groups. Further, to account for the remaining datalogger time lag, the downward displacement of the observed level depths by 2 m is recommended.
Thermal corrections for three types of SRDL sensors with standard error bars (corrections to be subtracted from the original temperature).
b. Temperature and depth (pressure) corrections for TDR profiles
Similar bias calculations were conducted for TDR temperature profiles (Figs. 15f–j). In the part of the North Pacific Ocean covered by the TDR observations during the TOPP project, the vertical temperature gradient is strongest within the 20–50-m layer (Fig. 15h) corresponding to the position of the seasonal thermocline. The mean gradient remains negative throughout the entire water column. Most significant seasonal changes in temperature and temperature gradient occur within the upper 100-m layer (Figs. 15h,i). In contrast to the CTD-SRDL recorders, the TDR recorders are characterized by the positive depth (pressure) bias, which is translated into the positive temperature bias under the conditions of a generally negative vertical temperature gradient characteristic for this region of the North Pacific Ocean. The maximum reduction of the overall temperature bias is achieved for the constant depth correction factor of 0.940 (Fig. 15g). The need for this correction arises partly because the TDR’s reported pressure is in “meters of freshwater” which overestimate pressure (dbar) by a factor of 0.980 665 (M. Rutishauser 2023, personal communication). It means that the actual pressure correction factor should be 0.958. This is still larger than the 1% accuracy claimed by the manufacturer, which we do not know how to explain. Application of the depth correction effectively reduces original bias throughout the entire water column (Fig. 15g). The diagnosed bias for the warm year half was found to be systematically lower compared to the bias for all collocated pairs and for the cold half year pairs. The difference might arise due to a considerable seasonal difference in the number of collocated pairs, with roughly 2/3 less data available for the warm year half. Further, upward shift of the original sample levels for the warm year half does not lead to the bias reduction within the upper 50-m layer. We explain this by the insufficient vertical resolution of the TOPP temperature profiles.
8. Conclusions
This study represents a first attempt to estimate instrumental biases in temperature profiles obtained by means of dataloggers attached to marine mammals. Two big groups of data have been considered: the data from SRDL recorders (e.g., small CTDs measuring both temperature and salinity) and the data from TDR temperature recorders. These two instrumentation types provided a significant part of temperature profiles for the upper 1000 m within the global hydrographic archive between 2004 and 2019. To estimate the instrumental bias, temperature profiles for these two groups of recorders are compared with quasi-collocated reference temperature profiles obtained by means of ship-based CTDs and Argo floats, with Argo profiles overwhelming in number of ship-based profiles. Estimation of the temperature offset between collocated CTD and Argo profiles in the same geographical regions where marine mammal profiles are available suggests a small negative bias of −0.008°C for Argo data, which can be interpreted as one source of uncertainty for biases derived for CTD-SRDL and TDR recorders.
The derived temperature biases for SRDL recorders depend on the sensor type and exhibit variations with depth and over time. The bias estimates are more reliable for the most numerous VP3/CTF sensor. The average bias derived for this sensor group is negative except for the layer 50–300 m (the thickness varies over time) where the bias is predominantly positive. Such vertical bias structure can be linked to the thermal structure of the water column. Taking account for possible time lag of SRDL loggers (not completely accounted for through the data postprocessing algorithm) reduces the bias standard deviation throughout the water column but does not eliminate the residual bias.
Based on the results of this study, we cannot fully explain the vertical structure of the CTD-SRDL temperature offset within the upper 200-m layer, with the predominantly positive bias values within the layer 100–200 m. This vertical bias structure may be linked to a thermal layering of the water column in the Southern Ocean and can be attributed to the combination of such factors as sensor time lags, different vertical velocities of Argo floats and mammals, and different degrees of water entrainment during the upward movement. Further research is needed to explain the complicated upper layer bias structure for SRDL data but remains beyond the scope of this study. Below the depth of 200 m, bias variations with depth are small and the bias remains negative for all years between 2007 and 2019 typically within the range from −0.01° to −0.04°C, e.g., within the uncertainty declared by the manufacturer. Temperature offset between Argo- and ship-based CTD data also show increased bias variation over depth within the upper layer suggesting link to the thermal structure.
The less accurate TDR recorders exhibit a different depth–time bias pattern, showing the predominantly positive bias (0.08°–0.14°C) below 100 m (i.e., well within the manufacturer declared accuracy of 0.5°C). The North Pacific region is characterized by stronger vertical temperature gradients compared to the Southern Ocean. Unlike in the case of the CTD-SRDL recorders, there is a clear dependence of the total temperature bias on the vertical temperature gradient indicating pressure (depth) overestimation.
Based on the results of the current study, we suggest corrections for instrumental temperature biases in CTD-SRDL and TDR recorders. For CTD-SRDL recorders, the thermal sensor–specific corrections are suggested (Table 1) along with a downward shift of observed levels by 2 m. For TDR recorders, we recommend reducing the reported sample depth by the factor of 0.940. In spite of the relatively small bias magnitude (especially for the CTD-SRDL recorders), taking these instrumental biases into account is necessary for the more accurate estimation of the regional ocean heat content change.
Acknowledgments.
This study is supported by Strategic Priority Research Program of the Chinese Academy of Sciences (XDB42040402), Youth Innovation Promotion Association, CAS (2020-077), and the National Natural Science Foundation of China (42076202 and 42122046). We are thankful to the staff of the National Center for Environmental Information for preparing the invaluable collection of the hydrographic data which served as a basis for this study. We thank Peter Koltermann for his useful suggestions and for the careful proof of the manuscript. We are grateful to the International MEOP Consortium and the national programs which made the marine mammal data freely available. We greatly appreciate detailed comments and suggestions from two anonymous reviewers that helped to improve the manuscript.
Data availability statement.
CTD, Argo, and APB temperature profile data analyzed during the current study reside in the World Ocean Database (Boyer et al. 2018) and are available from the following public domain resource of National Center for Environmental Information (https://www.ncei.noaa.gov/access/world-ocean-database-select/dbsearch.html). The MEOP-CTD database is freely available on the MEOP portal (http://www.meop.net).
REFERENCES
Biuw, M., and Coauthors, 2007: Variations in behavior and condition of a Southern Ocean top 0 predator in relation to in situ oceanographic conditions. Proc. Natl. Acad. Sci. USA, 104, 13 705–13 710, https://doi.org/10.1073/pnas.0701121104.
Block, B. A., and Coauthors, 2011: Tracking apex marine predator movements in a dynamic ocean. Nature, 475, 86–90, https://doi.org/10.1038/nature10082.
Boehlert, G. W., D. P. Costa, D. E. Crocker, P. Green, T. O’Brien, S. Levitus, and B. J. Le Boeuf, 2001: Autonomous pinniped environmental samplers: Using instrumented animals as oceanographic data collectors. J. Atmos. Oceanic Technol., 18, 1882–1893, https://doi.org/10.1175/1520-0426(2001)018<1882:APESUI>2.0.CO;2.
Boehme, L., P. Lovell, M. Biuw, F. Roquet, J. Nicholson, S. E. Thorpe, M. P. Meredith, and M. Fedak, 2009: Technical Note: Animal-borne CTD-Satellite Relay Data Loggers for real-time oceanographic data collection. Ocean Sci., 5, 685–695, https://doi.org/10.5194/os-5-685-2009.
Boyer, T. P., and Coauthors, 2018: World Ocean Database 2018. NOAA Atlas NESDIS 87, 207 pp.
Charrassin, J.-B., and Coauthors, 2008: Southern Ocean frontal structure and sea-ice formation rates revealed by elephant seals. Proc. Natl. Acad. Sci. USA, 105, 11 634–11 639, https://doi.org/10.1073/pnas.0800790105.
Cheng, L., and Coauthors, 2016: XBT Science: Assessment of instrumental biases and errors. Bull. Amer. Meteor. Soc., 97, 924–933, https://doi.org/10.1175/BAMS-D-15-00031.1.
Deacon, G., 1937: The Hydrology of the Southern Ocean. Cambridge University Press, 124 pp.
Fedak, M., 2004: Marine animals as platforms for oceanographic sampling: A “win/win” situation for biology and operational oceanography. Mem. Natl. Inst. Polar Res., 58, 133–147.
Good, S., B. Mills, T. Boyer, F. Bringas, G. Castelão, R. Cowley, G. Goni, V. Gouretski, and C. M. Domingues, 2022: Benchmarking of automatic quality control checks for ocean temperature profiles and recommendations for optimal sets. Front. Mar. Sci., 9, 1075510, https://doi.org/10.3389/fmars.2022.1075510.
Gordon, A. L., and B. A. Huber, 1990: Southern Ocean winter mixed layer. J. Geophys. Res., 95, 11 655–11 672, https://doi.org/10.1029/JC095iC07p11655.
Gouretski, V., 2018: World ocean circulation experiment – Argo global hydrographic climatology. Ocean Sci., 14, 1127–1146, https://doi.org/10.5194/os-14-1127-2018.
Gouretski, V., and F. Reseghetti, 2010: On depth and temperature biases in bathythermograph data: Development of a new correction scheme based on analysis of a global ocean database. Deep-Sea Res. I, 57, 812–833, https://doi.org/10.1016/j.dsr.2010.03.011.
Gouretski, V., and L. Cheng, 2020: Correction for systematic errors in the global dataset of temperature profiles from mechanical bathythermographs. J. Atmos. Oceanic Technol., 37, 841–855, https://doi.org/10.1175/JTECH-D-19-0205.1.
Gouretski, V., L. Cheng, and T. Boyer, 2022: On the consistency of the bottle and CTD profile data. J. Atmos. Oceanic Technol., 39, 1869–1887, https://doi.org/10.1175/JTECH-D-22-0004.1.
Hooker, S. K., and I. L. Boyed, 2003: Salinity sensors on seals: Use of marine predators to carry CTD data loggers. Deep-Sea Res. I, 50, 927–939, https://doi.org/10.1016/S0967-0637(03)00055-4.
Lydersen, C., O. A. Nøst, P. Lovell, B. J. McConnell, T. Gammelsrød, C. Hunter, M. A. Fedak, and K. M. Kovacs, 2002: Salinity and temperature structure of a freezing Arctic fjord – monitored by white whales (Delphinapterus leucas). Geophys. Res. Lett., 29, 2119, https://doi.org/10.1029/2002GL015462.
McMahon, K. W., W. G. Ambrose Jr., M. J. Reynolds, B. J. Johnson, A. Whiting, and L. M. Clough, 2021: Arctic lagoon and nearshore food webs: Relative contributions of terrestrial organic matter, phytoplankton, and phytobenthos vary with consumer foraging dynamics. Estuarine Coastal Shelf Sci., 257, 107388, https://doi.org/10.1016/j.ecss.2021.107388.
Mensah, V., F. Roquet, L. Siegelman, B. Picard, E. Pauthenet, and C. Guinet, 2018: A correction for the thermal mass induced-errors of CTD tags mounted on marine mammals. J. Atmos. Oceanic Technol., 35, 1237–1252, https://doi.org/10.1175/JTECH-D-17-0141.1.
Orsi, A. H., T. Whitworth III, and W. D. Nowlin Jr., 1995: On the meridional extent and fronts of the Antarctic Circumpolar Current. Deep-Sea Res. I, 42, 641–673, https://doi.org/10.1016/0967-0637(95)00021-W.
Roquet, F., J.-B. Charassin, S. Marchand, L. Boehme, M. A. Fedak, and G. Reverdin, 2011: Delayed-mode calibration of hydrographic data obtained from animal-borne satellite relay data loggers. J. Atmos. Oceanic Technol., 28, 787–801, https://doi.org/10.1175/2010JTECHO801.1.
Roquet, F., and Coauthors, 2013: Estimates of the Southern Ocean general circulation improved by animal-borne instruments. Geophys. Res. Lett., 40, 6176–6180, https://doi.org/10.1002/2013GL058304.
Roquet, F., and Coauthors, 2014: A Southern Indian Ocean database of hydrographic profiles obtained with instrumented elephant seals. Sci. Data, 1, 140028, https://doi.org/10.1038/sdata.2014.28.
Siegelman, L., F. Roquet, V. Mensah, P. Rivière, E. Pauthenet, B. Picard, and C. Guinet, 2019: Correction and accuracy of high- and low-resolution CTD data from animal-borne instruments. J. Atmos. Oceanic Technol., 36, 745–760, https://doi.org/10.1175/JTECH-D-18-0170.1.
Suzuki, I., S. Katsusumi, A. Fahlman, Y. Naito, N. Miyazaki, and A. W. Trites, 2014: Drag, but not buoyancy, affects swim speed in captive Stellar sea lions. Biol. Open, 3, 379–386, https://doi.org/10.1242/bio.20146130.
Treasure, A. M., and Coauthors, 2017: Marine mammals exploring the oceans pole to pole: A review of the MEOP consortium. Oceanography, 30 (2), 132–138, https://doi.org/10.5670/oceanog.2017.234.
