Observations of ocean temperature and salinity are essential for the initialization of seasonal prediction with a dynamical ocean–atmosphere coupled model (Vidard et al., 2007) because a potential source of seasonal predictability arises from the long memory associated with the high heat capacity of the ocean relative to the atmosphere. In particular, the sea surface temperature (SST) in tropical oceans is key, as shown in the pioneering work of Bjerknes (1969). Accurate prediction of tropical climate variability modes, such as the El Niño-Southern Oscillation (ENSO) and Indian Ocean Dipole, is crucial for successful seasonal prediction around the world (Saji et al., 1999; Shukla et al., 2000; Wang et al., 2000; Hoerling and Kumar, 2001; Behera and Yamagata, 2003; Saji and Yamagata, 2003; Kosaka et al., 2013) and, thus, for societal applications in agriculture, human health, and natural disaster prevention (e.g., Hashizume et al., 2009; Iizumi et al., 2013; Takaya et al., 2014; Yuan and Yamagata, 2015; Ikeda et al., 2017; Oettli et al., 2018).

The amount of ocean observations has increased enormously over the recent decade. SST is the most important variable to understand air–sea interactions. As of late 1981, satellite measurements have greatly improved such data coverage with respect to in situ observations alone, and they have provided fine-resolution SST data in real time (Reynolds et al., 2002, 2007). Subsurface temperature observations are primarily provided by mooring buoys, research vessels, profiling floats, and others. The mooring buoys located in the tropical oceans consist of the Tropical Atmosphere–Ocean/Triangle Trans-Ocean Buoy Network (TAO/TRITON) in the tropical Pacific (Mcphaden et al., 1998; Ando et al., 2017; Smith et al., 2019), the Research Moored Array for African-Asian-Australian Monsoon Analysis and Prediction (RAMA) in the tropical Indian Ocean (McPhaden et al., 2008; Hermes et al., 2019), and the Pilot Research Moored Array in the Tropical Atlantic (PIRATA) (Bourles et al., 2008; Foltz et al., 2019). The global Volunteer Observing Ship (VOS) program provides expendable bathythermograph (XBT) measurements along merchant shipping routes. More recently, the international Argo network of drifting profilers has greatly improved the data coverage of subsurface temperature and salinity, revolutionizing subsurface ocean sampling (e.g., Hosoda et al., 2008).

Observational datasets for use in data assimilation are thus created by collecting all measurements from the aforementioned sources and homogenizing the data in terms of format and, in some cases, vertical sampling and accuracy, before eventually bias-correcting the data. For instance, Version 4 of the Met Office Hadley Center series “EN4” gathers data from all types of ocean profiling measurements that provide temperature and (when available) salinity and provides global quality-controlled ocean temperature and salinity profiles in near real time (Good et al., 2013).

As an additional data source, ocean observational data from marine animal-borne instruments via satellites in near real time are emerging (Harcourt et al., 2019). Some marine animal-borne instruments can provide in situ subsurface temperature profile data from parts of the oceans wherein little or no other data are currently available. For example, the Argo float, a very powerful tool for quasi- real-time monitoring of high-quality subsurface temperatures and salinity fields, cannot cover areas of shallow ocean like the coastal regions and the marginal seas because the float is designed for the upper 2,000 m of the ice-free global ocean. If some of the animal-borne instruments were additionally capable of real-time subsurface temperature data beyond the data coverage already achieved by the present Global Ocean Observing System, the information could be proven for all ranges of prediction: weather, subseasonal, seasonal, multiyear, decadal, and longer-scales climate prediction. The first use of marine mammal data in operational oceanography dates back to 2004 (see Boehme et al., 2009). Previous studies also demonstrated the benefit of marine animal-borne instruments in ocean analysis, reanalysis, and nowcast/forecast systems (Roquet et al., 2013; Carse et al., 2015; Miyazawa et al., 2015, 2019; Mallett et al., 2018).

In the present study, we focused on the Arafura Sea, one of the marginal seas of the Maritime Continent. With its complex distribution of several islands and shallow oceans, the region experiences some of the warmest ocean temperatures of the world and is commonly known as the “boiler box” of the global weather and climate (Ramage, 1968; Neale and Slingo, 2003). Moreover, this region is strongly related to the northern Great Barrier Reef, wherein coral bleaching has widely been reported in marine heat waves (Hughes et al., 2017). Subsurface temperature data, however, remains largely unavailable in the area.

From June to October 2017, we successfully obtained temperature profile data from sea turtles in that area. To investigate the impacts of turtle-based temperature-depth profiles on seasonal prediction, we performed observing system experiments (e.g., Oke and Schiller, 2007; Fujii et al., 2015) with a quasi-real-time seasonal prediction system based on a coupled general circulation model (GCM), named SINTEX-F2 (Doi et al., 2017). For the original ensemble reforecast experiments, the EN4 dataset (including all types of ocean measurements except for those from animal-borne instruments) was used for the initialization. For this current study, however, we performed new ensemble reforecast experiments with additional temperature measurements from the sea turtles at the initialization and also explored the differences between the outputs of these runs and those of the original runs. We believe that our research could open a new door in improving regional seasonal prediction skills through better representation of the initial state of the subsurface ocean via integrating marine animal-borne instruments into the present Global Ocean Observing System.

We started compiling differences in the ocean mixed layer depth in August 2017, at 0–1 months lead time after initialization. The mixed layer depth of the wTurtle_forecast was shallower than that of the Ctl_forecast in the Arafura Sea, as shown in Figure 3A. The maximum difference was about 10 m in the region of 133°–138°E, 10°–5°S, which is consistent with the location of the temperature measurements from the sea turtles (Figures 1, 2). Figure 1B shows the depth–temperature profiles in this region on the August 1st 2017, i.e., the initialized date of the ocean observing system experiments. The total number of observation points from the sea turtles from the surface to around a depth of 120 m (976 points) was assimilated into the seasonal prediction system in this region on the August 1, 2017 (Figure 4). The wTurtle_forecast showed the shallower mixed layer depth and a lower temperature by about 1°C around 60–80 m depths relative to the Ctl_forecast. Thus, it is reasonable to assume that the assimilation of those observations strengthened the ocean stratification in shallow areas in which no data were available because models at a relatively coarse resolution tend to smooth out the vertical structure and stratification. As the mean mixed layer depth in the climatology was about 35 m there, the difference was beyond 20% of the mean. Thus, we can expect that this difference in the subsurface ocean may influence the seasonal climate at a few months lead time. Note that there were some differences observed outside of the Arafura Sea, e.g., around Java-Sumatra. Although the differences were statistically significant beyond the 90% confidence levels, the values were only on the order of 3-m. Therefore, we expect that these differences outside of the Arafura Sea are just noise, and do not strongly influence the seasonal predictions at a few months lead time.

Figure 3B shows the SST difference in November 2017, 3–4 months lead time after the initialization. The SST predicted by the wTurtle_forecast was higher in the Arafura Sea than that predicted by the Ctl_forecast (Figure 3B). Figure 5 shows the SST anomalies found from the observations, the Ctl_forecast, and the wTurtle_forecast. It can be seen that the wTurtle_forecast was indeed more realistic over the Arafura Sea. We calculated the spatial root mean square error (RMSE) of the difference in the SST between the observation and the predictions for the region of 130°–150°E, 15°S–5°S. While the RMSE in the Ctl_forecast is 0.58, that in the wTurtle_forecast is 0.28, only about 50% of the RMSE in the Ctl_forecast. Meanwhile, Figure 6 shows the time series of the SST anomaly prediction averaged in 130°–150°E, 15°–5°S. For August 2017, the observations show that the SST was higher than normal in this region; following this, the observations show that the positive SST anomaly persisted through the austral summer. However, the ensemble mean of the Ctl_forecast predicted that the positive SST anomaly would quickly decay and that this region would return to a normal state after October. As shown in the wTurtle_forecast, the sea turtle data assimilation was able to precondition the shallow mixed layer depth in the austral winter and sustain the positive SST anomaly in the following austral spring/summer, which, otherwise, would have soon vanished without proper constraint of the mixed layer depth, as in the prediction of the Ctl_forecast.

We have shown that additional assimilation of temperature measurements from sea turtles can potentially improve regional seasonal climate prediction around the Arafura Sea, as in the 2017 warm event case. The present study was similarly a pilot study, and the results may have been influenced by the model, assimilation, and manner of generating the ensemble members. The results were based on the model initialization within July–August 2017 as a mere case study; the animal-borne measurements by sea turtles in the study area are still unique and the available data is very limited. However, the present study can support the merits of integrating marine animal-borne instruments into the present Global Ocean Observing System from the perspective of better representing the initial state of numerical seasonal prediction. We hope that our study encourages efforts for sea turtle-borne observations development in terms of deployment, collection, and processing so that we may clearly demonstrate the robustness and importance of their impact more clearly. Note that some data from marine mammals in high-latitude regions are already included in EN4. Very recently, Patel et al. (2018) suggested that the use of sea turtle-borne telemetry devices has the potential to improve the resolution of in situ temperatures through depth data and, in turn, improve oceanographic model outputs. However, as of yet, sea turtle data have not been used for predictions. Moreover, to the best of these authors’ knowledge, this is the first time that marine animal data have been used for seasonal prediction in the tropical regions. Sometimes, the impact may be regional, as shown in this study, but highly accurate predictions of regional ocean and climate may open a new door to better societal applications for the surrounding rim countries. The turtle measurements of in situ temperatures were originally motivated by a research interest in marine biology; however, they may also contribute to operational seasonal prediction, and thus climate services, and complement the coastal ocean observation network, which is not accessible by autonomous profilers. Indeed, sea turtle data are now being collected in a delayed time mode but, in the future, may also be disseminated in near real time. Such near real-time collection and dissemination services are now in development.

Figure 7 shows the horizontal distributions of the differences in SST and mixed layer depth anomalies between the wTurtle_forecast (ensemble mean) and the Ctl_forecast (ensemble mean) from August to November 2017. The anomalously shallow mixed layer depth seems to lead the positive SST anomaly in the following month around the Arafura Sea. The signal of the SST prediction extended eastward from the Arafura Sea and then appeared in a broader region relative to the signal of the mixed layer depth at 0–1 months lead time. This feature was clearly enhanced in the warmest ensemble member of the wTurtle_forecast. How could the shallow mixed layer depth around the Arafura Sea in the austral winter persist/generate the positive SST anomaly in the following austral spring/summer? After the austral winter, this region warms up, primarily due to the surface heat flux, as a normal seasonal evolution (e.g., Halkides et al., 2011). The anomalously shallow mixed layer depth could have enhanced the sensitivity of the mixed layer temperature to seasonally modulate atmospheric warming through the shallow mixed layer depth, leading to the positive SST anomaly in the following season. That is, heat fluxes are basically the same, but with a shallower mixed layer depth they can imply a stronger warming. This mechanism could potentially work in other regions on a seasonal time scale, as shown in previous works (Doi et al., 2010; Morioka et al., 2011; Kataoka et al., 2017; Tozuka et al., 2018). Although the heat budget analysis through posterior diagnostics was performed, it was still too difficult to capture the signal relative to the noise in order to reasonably trace that mechanism. As shown in Figure 6, the ensemble spread with the wTurtle_forecast was larger than that with the Ctl_forecast. This may have been partly due to the fact that the mixed layer depth in the wTurtle_forecast was shallower than that in the Ctl_forecast and could have enhanced the response of the mixed layer temperature, i.e., proxy of SST, to the noisy atmospheric forcing in this region. We also checked the differences in the rainfall between the Ctl_forecast and the wTurtle_forecast. However, we could not find a significant impact because the rainfall prediction was very noisy. These difficulties may have been partly due to the small ensemble size of the ocean observing system experiments and the availability of the sea turtle data only in limited periods and regions. To address these points, we plan to perform large ensemble reforecast experiments for more cases, although their costs are high and beyond operational capability (Doi et al., 2019).

What is the importance of the Arafura Sea with regard to the large-scale climate? We calculated the horizontal distribution of the correlation of the SST anomaly in the Arafura Sea and the 2 m air temperature anomaly in the global regions for the October–December averages during 1983–2015 using the observational data and the reanalysis data (figures not shown). It can be seen that the warm SST anomaly in the Arafura Sea was linked with the warm 2 m air temperature anomaly in some parts of the Philippine Sea and Coral Sea. As shown in the partial correlation analysis, this relationship could be robust, even if the influences from large-scale tropical climate variations such as ENSO and the Indian Ocean Dipole were linearly removed (figures not shown). With regard to precipitation, it can be seen that the warm SST anomaly in the Arafura Sea was linked with the wetter-than-normal condition over the Maritime Continent and northern Australia (figures not shown). Although a high correlation does not always show causality, these results may suggest the potential importance of seasonal prediction of the SST anomaly in the Arafura Sea for some parts of the Philippine Sea, Coral Sea, Maritime Continent, and northern Australia. In addition, some previous works already suggested a relationship between the SST anomaly in the Arafura Sea and the larger-scale circulations. Kawamura et al. (2002) showed that the local climate over the Arafura Sea is important for the onset mechanism of the Australian summer monsoon. Indonesian rainfall variability is influenced not only by ENSO but also by local air–sea interactions, including those in the Arafura Sea region (Hendon, 2003; Kubota et al., 2011). Additionally, Zhang et al. (2016) suggested that seasonal prediction of an extratropical climate (e.g., the East Asian climate) is partly dependent upon the prediction skill for rainfall over the Maritime Continent, demonstrating the roles of remote and local forcing in the variation and prediction of the regional rainfall there. Since the complicated relationships between the local climate over the Arafura Sea and the larger-scale circulations are not yet well understood and are still difficult for climate models to capture (e.g., Schiemann et al., 2014), higher resolution and better representation of the air–sea coupling process in climate models may be needed to more clearly display the potential importance of the local climate over the Arafura Sea in the large-scale climate.

The quality of the temperature profile data captured by sea turtles was discussed and compared to the other in situ data of the Global Temperature and Salinity Profile Program (GTSPP) by Miyazawa et al. (2019). Although it was difficult to match up the sampling positions and times of the two data sources exactly, the aforementioned study checked the dependency of their mutual consistency as a function of the horizontal distance between them. A higher correlation and smaller root mean square difference between them suggested a sufficient reliability of the turtle measurements compared to the usual GTSPP data. McMahon et al. (2005) also compared the thermistor of SRDLs with that of the Argo floats. The Argo floats have much better temperature sensors than SRDLs; the accuracy of SRDL temperature data is 0.1°C, which is at least one order of magnitude lower than that of the Argo floats. Note that other manufacturers are also building loggers with various degrees of accuracy. However, some marine animal-borne instruments with SRDLs could potentially be beneficial in areas wherein little or no other subsurface ocean data are currently available. Further studies on quality assessment may be required before compiling the temperature measurements from sea turtles into the existing Global Ocean Observing System in order to assimilate these data into operational short-range prediction systems as well.

The datasets generated for this study are available on request to the corresponding author.

The fieldwork was conducted with permission from Dinas Perikanan, Pemerinta Kabupaten Tambrauw (19630701 199403 1 009). All procedures performed in this study involving animals were approved by the Animal Experimental Committee of the University of Tokyo and conducted in accordance with the Guidelines for the Care of Experimental Animals (A17-13).

TD performed the seasonal prediction system on a basis of the ocean-atmosphere coupled model, and analyzed the observation data and model prediction outputs. AS mainly contributed to the assimilation of temperature measurement from sea turtle into the seasonal prediction system. TF, HS, and KS mainly contributed to the temperature measurements from sea turtles. All authors contributed to designing the research, interpreting results, and writing the manuscript.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.