Introduction

Prolonged extremely warm ocean temperatures, namely marine heatwaves (MHWs), are proven to have a great impact on the marine ecosystem, the atmosphere aloft, and human communities1,2,3,4,5. Generally, MHWs could be identified by discrete prolonged warm events with the ocean temperature (mostly, the sea surface temperature, SST) anomalies above a certain threshold6. Based on that, recent studies suggest that MHWs are becoming longer and more frequent all over the world7,8,9, while more long-lasting MHWs have been observed10,11,12. Particularly, from 1925 to 2016, the global increase in MHW frequency and duration led to 54% more MHW days13. More recently, it is also found that coastal MHW occurrence and duration increased by about 1−2 events per decade and 5−10 days per decade globally during the past 25 years, respectively14. Some studies further suggest that the globally averaged MHW days may become over 112 days per year by the end of the twenty-first century under current carbon emissions, and the most severe MHWs that occur once every hundreds of years may become annual events7.

The enhancement of MWHs has been observed globally; however, they did not increase homogeneously due to their various drivers8,13,14,15, including SST warming16,17, climate variations18,19, and ocean dynamics20,21, leading to challenges in predicting/projecting their future changes22,23. It is found that the warming of mean SST acts as the main driver of increasing MHWs16, which also has a certain impact on the SST variance and further enhances MHW occurrence18. Global and regional scale climate changes, such as the El Niño–Southern Oscillation (ENSO) and the Pacific Decadal Oscillation, greatly modulate MHW characteristics via their impacts on the surface heating (by influencing the pressure system and hence the cloud cover) and air-sea interactions (e.g., wind-current responses and mixed layer processes14,19,24). In addition to that, oceanic phenomena, like the western boundary currents15,25 and coastal upwellings26, could influence MHWs via heat transport. Moreover, these drivers may act coherently, resulting in complex responses of MHWs in some specific regions, such as Asia and the Indo-Pacific (AIP) region20,27.

On the other hand, as more and more studies focused on MHWs and their variabilities, a few studies noticed that the detected MHWs strongly depend on the selected dataset6,8,14. Particularly, a recent study was conducted using multi-product-based analysis for coastal MHWs and showed a large variety in both magnitude and trend among the results using different satellite-based SSTs14. Although some studies suggest the spatial distributions and the time series of MHWs using different SST datasets were generally similar13, the spread was actually large from a regional perspective (see Fig. S4 in ref. 13). One possible reason is that recent SST products do have similar mean states and variabilities, making them all acceptable in climate studies28. However, as shown in the studies mentioned above, it may not be true for evaluating MHWs.

Among the hotspots of MHWs, the AIP region is of particular interest, as it is dominated by complex climate and oceanic variations. With the world’s largest population and abundant marine biodiversity, including the large area of coral reefs and related marine ecosystems, the AIP region is also one of the most vulnerable regions to MHWs29,30,31. Recent studies also demonstrated that the AIP region has warmed faster than the global mean in the last four decades32,33,34,35, which induced the increase in MHW activity. Such enhancement was observed in the entire AIP region from the tropics to mid-latitudes27,36,37,38. Nevertheless, most previous studies were conducted based on only one dataset, leading to potential uncertainties in our current assessments of MHWs over this region.

Therefore, in this study, our primary goal is to estimate the general characteristics of MHWs over the AIP region and provide quantitative analyses of the potential uncertainties when using different datasets. Further interest is also given to whether the discrepancies have certain patterns or sources, which may help us reduce them. To do so, we compared MHWs obtained by three commonly used SST datasets, including the Optimum Interpolation Sea Surface Temperature (OISSTv2.139,40), the Operational Sea Surface Temperature and Ice Analysis (OSTIA41) system, and the Merged satellite and in situ data Global Daily Sea Surface Temperature (MGDSST42), which are proved to have low biases compared to in situ data28,43. Moreover, since the reference depths of each SST dataset are not identical, we focused on the occurrence and related durations, which are sufficient to represent the likelihood of the estimated MHWs. Based on our analyses, while all three datasets show some similar patterns, large uncertainties (defined by the ratios between the variance of metric differences and the variance of the baseline from OISST) were seen in their occurrence, durations, and long-term trends. These discrepancies were not related to the mean SSTs or their variances in different datasets; however, the configurations in each SST analysis may contribute partially. Globally extended works are also conducted, suggesting such discrepancies are common in all the oceans (see Supplementary Figs. 8 and 9).

Results and discussion

General Conditions of MHWs

Figure 1 shows the occurrences and mean durations of MHWs over the AIP region based on three SST datasets: OISST, OSTIA, and MGDSST. In general, during the last four decades, MHWs tend to be found near the coast and the regions dominated by warm currents, such as the Kuroshio and its extension region. Results also show that some MHW-rich areas were connected and covered large regions of the marginal seas and the Maritime Continent (MC), suggesting those MHW events may not have occurred individually but were highly organized. On the other hand, more frequent MHWs were often accompanied by shorter durations (Fig. 1e-g; also see Supplementary Fig. 1 for the relations between occurrence and durations).

Fig. 1: Metrics of MHWs based on OISST, OSTIA, and MGDSST.
figure 1

a−c Total occurrence and (e−g) the mean durations of marine heat waves (MHWs). Their zonal means are shown in (d, h) (solid lines), and the shading represents their standard deviations.

Full size image

It is found that MHWs last about 10−15 days in most regions, while only some specific regions suffered from MHWs with durations longer than 20 days, such as the eastern coast of the northern Korean Peninsula, the Oyashio region, and the western subarctic gyre region (the northeast corner of our domain). In addition, unlike the occurrence, durations of MHWs exhibited a weak but clear dependence on the latitude, which is longer in the north and shorter in the south (Fig. 1h).

Although we obtained some similar patterns of MHWs based on the three SST datasets, great differences among them could not be ignored. In the tropical regions, much fewer MHWs were detected by OISST (Fig. 1a, d), while the numbers could become almost 50% larger if we used OSTIA (see Supplementary Fig. 2). Compared to them, although MGDSST showed moderate numbers of MHWs, it showed more MHWs over the offshore regions away from coastlines, such as the central Bay of Bengal (BOB) and the areas northwest of Kalimantan Island. However, situations changed in higher latitudes. In the regions north of 20°N, OISST had the highest occurrence among the three datasets, while much fewer MHWs were found in MGDSST, especially near 40°N (Fig. 1d).

Interestingly, although the occurrence of MHWs showed large differences between OSTIA and OISST, they had similar mean durations in most regions with differences of no more than 3 days (Fig. 1e, f), except in the MC, the central North Pacific, and the Sea of Okhotsk (OS; Fig. 1d, f; also see Supplementary Fig. 2). Unlike the small differences between OISST and OSTIA, MHWs had much longer durations in MGDSST, resulting in about 5-day biases over the entire domain (Fig. 1h).

Note that we did not find clear dependence on the SST types (i.e., the reference depths) either in MHW metrics or in the errors related to observations (see Supplementary Figs. 5 and 6), and the abovementioned discrepancies will be discussed later.

Long-term variations of MHW activity

After examining the general characteristics of MHWs, we’d like to focus on the temporal variations of MWHs over the AIP region. In addition, by considering the climate regime shift in the late 1990s as mentioned in previous studies27,44, two separate periods were considered in this study. As shown in Fig. 2a,b, while both occurrence and duration of MHWs reached their highest levels in 1998, distinct differences in their trends were also found before and after 1998. To avoid unexpected under/overestimations, we excluded the year 1998 during the trend estimations. During the early period (1982−1997), MHWs remained few, and the annual durations were also short, having almost no clear trends before 1998. Meanwhile, by considering their spatial distributions, MHWs in the early period were more likely localized phenomena and less organized, while large changes were only seen over some small regions, such as some coastal areas in the East China Sea and the OS (see 1982−1997 panels in Fig. 2).

Fig. 2: Long-term trends of the annual occurrence and durations of MHWs.
figure 2

Time series of the domain-averaged annual (a) occurrence and (b) duration of MHWs (solid lines) and their linear trends (dotted lines) based on OISST (blue), OSTIA (green), and MGDSST (red) datasets in two periods: 1982−1997 and 1999−2020, together with (c−n) the spatial distributions of grid-based trends. The linear trends obtained by the Theil-Sen estimator are embedded in (a, b) (values and dashed lines) with insignificant trends marked by the Italic font, which were based on the annual values. In (c−n), only significant values (p < 0.1) are plotted.

Full size image

By contrast, MHW activity was significantly enhanced after 1998, covering large areas over both marginal seas and the open ocean. As a result, more and longer MHWs were observed over most regions in our domain. Results show that the occurrence of MHWs doubled or more during the past two decades (>0.83 per 10 yr), while their durations increased by over 30% (>18.96 d per 10 yr). Interestingly, large increasing trends were seen over almost all the tropical/subtropical and subarctic regions but were limited within a few hot spots in mid-latitudes, such as the Bohai Sea or the southern part of the Sea of Japan.

Although MHW activity obtained by the three datasets varied in a similar way, their differences were not negligible, especially in the estimated long-term trends. For example, during the early period, we did not obtain consistent results using different datasets due to the relatively weak MHW activities, even though the trends were statistically significant in OSTIA and OISST (Fig. 2a). In addition, MGDSST provided some unique patterns with strong descending signals over the tropics and central Pacific regions, which could be barely seen in other datasets (Fig. 2k, m). On the other hand, while the recent enhanced MHW activities were found in all three datasets, the estimated trends varied widely (e.g., Fig. 2b; also see panels for 1999−2020 in Fig. 2).

Uncertainties in regional evaluations

To quantify the extracted MHWs using different datasets and to find which region suffered the most, we introduced an uncertainty index based on the variance of metric differences and used the metrics from OISST as the baselines (Fig. 3; see a global version in Supplementary Fig. 9). It is found that the largest uncertainties are concentrated mainly in tropical regions, especially over the MC covering both marginal seas and open ocean areas. Meanwhile, relatively large uncertainties were also observed in the BOB, the Bohai-Yellow-East China Sea (BS-YS-ECS), and the OS.

Fig. 3: Uncertainties in the estimated metrics of MHWs, and time series of area-averaged metrics over four selected regions.
figure 3

a, c Uncertainties based on OSTIA and (b, d) MGDSST datasets, which are defined by the ratios between the variance (VAR) of metric differences and the variance of the baseline from OISST. Temporal variations of area-averaged annual occurrence (solid line) and durations (dashed lines) of MHWs over four selected regions together with the correlation coefficients (r): (e) the Maritime Continent (MC), (f) the Bay of Bengal (BOB), (g) the Bohai-Yellow-East China Sea (BS-YS-ECS), and (h) the Sea of Okhotsk (OS), which are marked by the boxes in (a).

Full size image

To further evaluate the influences of such uncertainties, Fig. 3e−h show the temporal variations of the estimated occurrence and durations of MHWs in four sub-regions. We found that the results in these regions shared similar interannual variations with correlation coefficients over 0.74; however, unlike the results in Fig. 2a, b, they showed non-monotonic biases that varied year to year, leading to the large spread in the estimated linear trends (see Supplementary Table 1).

In addition to the temporal biases, one should also be aware of the differences in their spatial distributions. Figure 4 shows the extracted MHWs in 2020 based on three datasets, and large differences in occurrence and durations among the datasets could easily be seen. It is found that MHWs may exist in one dataset but not in others, or they could be found in totally different areas in a small sub-region. For example, based on OISST, MHWs were detected near the eastern coast of BOB (Fig. 4d), but OSTIA showed they occurred mainly in the western areas (Fig. 4e). Meanwhile, only a few MHWs were detected in OISST and MGDSST within the MC; however, that was not the case in OSTIA which showed much more and longer MHWs, especially near the western coast of Sumatra Island (Fig. 4b,e). Similar to that, MHWs did not even exist near the western coast of the Korean Peninsula in OISST, but the other two datasets showed them. In addition, even for those events that were detected in all datasets, their magnitude could still have a large spread.

Fig. 4: MHWs in 2020 based on OISST, OSTIA, and MGDSST.
figure 4

Annual occurrence and durations of MHWs using (a, d) OISST, (b, e) OSTIA, and (c, f) MGDSST.

Full size image

One may consider that the regions with large discrepancies may share similar situations or environments; however, our results did not support that. First of all, these regions are located from the tropics to higher latitudes and from coastal areas to the center of open oceans with depths from ~50 m (e.g., the Bohai Sea) to thousands of meters (e.g., the tropical Indian Ocean), which are also dominated by different climate and/or weather systems. Meanwhile, their oceanic conditions are also different. Both regions with and without strong ocean currents suffered large uncertainties (e.g., the BOB and the OS), while the mixed layer depths (hence the upper ocean stratification) also varied much among these regions45, which are believed to be important in MHW formations46.

Note that the uncertainties mentioned here were designed to represent the spread among the datasets but not to indicate which dataset is better or worse. Similar conclusions could also be obtained by using OSTIA or MGDSST as the baseline (see Supplementary Fig. 3).

Influences of the mean SSTs and their variances

Previous studies suggest that MHWs strongly depend on the mean SSTs and the SST variances16,18. Therefore, one may consider the biases and related uncertainties shown above could be attributed to the differences in mean SSTs and their variance among the datasets. To confirm that, we compared three SST datasets and summarized them in Fig. 5. Results indicate that the domain-averaged SSTs and variances of SST anomalies correlated well with the duration (Fig. 2b) and the occurrence (Fig. 2a) of MHWs, respectively (also see Supplementary Fig. 4 for correlation coefficients between SSTs and MHWs). However, as shown in Fig. 5b−g, neither the differences of SSTs nor the variances could explain the distributions of uncertainties found above (Fig. 3a, b), especially for those in tropical regions. For example, although large uncertainties were found in the BOB and the MC, the differences in mean SSTs or variances were quite small. By contrast, the large differences in variance in the Sea of Japan did not lead to large uncertainties. Note that the dot-like patterns in Fig. 5c−g represent the bias-correction in three SST datasets by the Tropical Atmosphere Ocean/Triangle Trans-Ocean buoy Network (TAO/TRITON) Array, which reduced the differences among the datasets; however, our results suggest that such adjustments were limited in about 2-degree area around the buoys and did not reduce the uncertainties in MHW estimations (e.g., Fig. 3a−d).

Fig. 5: Mean SSTs and variances of daily climatology-removed SST anomalies.
figure 5

a Time series of domain averaged SSTs and variances of SST anomalies. b−g Spatial distributions of mean SSTs and the variances of SST anomalies based on OISST and (c, d, f, and g) their differences. Note that panels f-g show the differences of variances, not the variances of differences.

Full size image

Possible influences of configurations during the SST analysis

It is known that all three SST datasets ingest in situ observations from ships and buoys, while OISST also started to include the Argo floats in 201640. Therefore, it is not a surprise to see these datasets showing good agreement with each other and having low errors when compared to most in situ data28,43,47,48 (also see Supplementary Figs. 5, 6, and Supplementary Table 2 for a simple comparison using in situ SST dataset49). However, their inputs and related procedures could possibly influence their performance in MHW assessments.

Among the three datasets, OISST only uses the Infra-Red (IR) sensor, while the others combine both IR and microwave (MW) sensors. Although the IR sensors have a better spatial resolution, they also suffer from the influences of cloud cover and weather conditions50. With the help of in situ observations, it may not be a large problem; however, as shown in Fig. 1, OISST did exhibit the lowest MHW occurrence in the MC where vigorous convection develops51. Similar issues may also influence the uncertainties in coastal regions. When the IR image is not available, the MW sensors cannot obtain high-quality signals near land52, leading to relatively large errors and hence the discrepancies among the datasets near the coastline47,53. On the other hand, although we regridded OSTIA onto a 0.25° mesh, its original high-resolution data could still have some advantages in coastal regions43 and may result in the relatively higher MHW occurrence there (Fig. 1).

In addition to that, previous studies suggest that the interpolation procedures used for generating satellite SSTs could induce attenuation or spurious small-scale features54, when a few high-resolution inputs are available. This problem could further induce part of uncertainties by producing spurious MHWs. Meanwhile, MGDSST may suffer more from the smoothing issue, because of its cut-off procedures for high-frequency signals55,56, which might be the source of its 5-day-longer biases.

Another influence could come from the differences in sea ice treatments. OISST set the SST to −1.8 °C for 100% sea ice concentration (IC), and it also uses a linear ice-to-SST conversion algorithm when IC is larger than 50%39 (35% in version 2.1 from 201640). Same as OISST, MGDSST also set the SST to −1.8 °C when IC reaches 100%, but no other modification was applied42. Unlike the other two datasets, OSTIA uses a NEMOVAR data assimilation scheme57, which determines the background SST with a ‘decay timescale’ varying from 5 days to 30 days depending on IC41. Therefore, the SSTs over the frequently freezing regions (e.g., the OS) could have great differences among the datasets, leading to large uncertainties in MHW assessments.

Note that one may consider the multi-product ensemble SST analysis could have better performance in representing SSTs and hence MHWs. However, previous studies suggest that the ensemble median SST does not have to show the lowest errors28,43. In addition, our simple three-dataset mean SST also shows similar results and exhibits lower MHW occurrence in the AIP region (see Supplementary Fig. 7), although three datasets are not enough to obtain a robust ensemble mean. Finally, we would like to reiterate that we only tried to assess the uncertainties in MHWs obtained by different datasets but not to provide a quality judgment of them, which should be conducted based on independent in situ data and more proper analyses.

Conclusion

In this study, we analyzed the general characteristics (occurrence and duration) of MHWs over Asia and the Indo-Pacific (AIP) region based on three commonly used SST datasets. In summary, during the last four decades, MHWs tend to occur in the regions near the coastlines and/or with strong warm currents. The occurrence and durations of MHWs were highly correlated, and more frequent MHWs often led to their shorter durations. By contrast, results showed a weak but clear latitude dependence in the mean duration of MHWs, but none was found in their occurrences. In addition, MHWs were rare and inactive before 1998 but were significantly enhanced after 1998 with more and longer MHWs in almost the entire AIP region.

On the other hand, although all SST datasets showed some similar patterns, the large regional differences could not be ignored. Table 1 summarizes the area-averaged MHW metrics (occurrence, duration, and uncertainties) over the AIP region and four subregions mentioned above. Overall, OSTIA showed a 5−10% larger MHW occurrence in most AIP regions, while MGDSST exhibited the lowest, especially in high latitudes. One exception was found in the Maritime Continent, where OISST showed much fewer MHWs. In addition to the occurrence, a large diversity of the locations of MHWs was also found among the datasets. As for MHW duration, a 5-day-longer bias was seen in MGDSST, while the differences between the other two datasets were relatively small. These differences further lead to large uncertainties in MHW metrics and their long-term trends in both coastal and open ocean areas, especially in the Maritime Continent and the East Asia marginal seas; yet no clear dependence was seen in their spatial distributions. Based on an uncertainty index defined by the ratios of variances, discrepancies among the datasets could exceed 0.6 or more. Moreover, although we found that the domain-averaged metrics of MHWs generally followed the mean SSTs and the variances of climatology-removed SST anomalies, neither of them could explain the uncertainties and their distributions. On the other hand, these discrepancies may partially be contributed by sensors, interpolation/smoothing procedures, and sea ice treatments in different SST datasets, which greatly influence the small-scall and high-frequency features that are important in MHW assessments. Overall, our study suggests that the use of multiple datasets is crucial for evaluating extreme events.

Table 1 Mean metrics of MHWs in 1982−2020 over Asia and the Indo-Pacific (AIP) and four selected regions: the Maritime Continent (MC), the Bay of Bengal (BOB), the Bohai-Yellow-East China Sea (BS-YS-ECS), and the Sea of Okhotsk (OS) together with their uncertainties (in brackets), which are defined by the ratios between the variance of metric differences and the variance of the baseline from OISST (see Fig. 3)
Full size table

The discrepancies and spread of datasets are often considered and well-treated in model-based studies58; however, they were barely mentioned in the studies using observational datasets, especially the widely-used SSTs. Although previous studies demonstrated the mean state and variations of present SST products are consistent and good for climate studies28, this may not be true when investigating extreme events. Nowadays, more frequent and longer MHWs are observed not only in Asia and the Indo-Pacific region but also around the global oceans due to the rapidly warming climate59,60. Our extended work shows that a large spread in MHW assessments could also be found globally (see Supplementary Figs. 8 and 9), which could further induce large uncertainties in the evaluations and decision-making processes for their impacts on the ecosystem and socio-economy. Although some studies suggest using an ensemble mean of multiple datasets13, such a process may also lead to the loss of extreme values that are critical in MHW assessments. In addition to that, some recent studies suggest detecting MHWs as spatiotemporally connected pixel-based events61,62, which could help facilitate the driving mechanisms of MHWs and suppress unexpected spurious events. It could be one possible solution for mitigating some of the discrepancies, but it requires further analyses. More studies also need to be done on this issue, such as more detailed evaluations of SST datasets by independent in-situ observations, investigations on optimal MHW detection algorithms or definitions for different datasets, and the improvement of satellite retrieval algorithms52.

Method

Data

The daily SST datasets were utilized to diagnose the physical processes of individual MHW events over Asia and the Indo-Pacific regions (80° ~ 170°E, 10°S ~ 60°N). Concerning the different sources and algorithms among different datasets28, three commonly used global SST analysis datasets that cover the study period of 1982−2020 are selected for comparison, including the National Oceanic and Atmospheric Administration (NOAA) 0.25° daily OISSTv2.139,40, the U. K. Met Office 0.05° OSTIA41, and the 0.25° daily MGDSST42 from the Japan Meteorological Agency (JMA). In particular, for ease of intercomparison, the 0.05° OSTIA SST was regridded to 0.25° based on the first-order conservative remapping by using the Climate Data Operators software63. In addition, the NOAA In situ SST Quality Monitor (iQuam) dataset49 is also obtained for error estimations in our extra analyses in the Supplementary Information.

All three satellite SST datasets include in situ observations but with different combinations of satellite sensors. OISST uses the Advanced Very High Resolution Radiometer (AVHRR) only during our study period. MGDSST combines both IR and MW sensors, including the AVHRR, the Visible/Infra-red Imager Radiometer Suite (VIIRS), the Spinning Enhanced Visible Infra-Red Imager (SEVIRI), the WindSat, the Advanced Microwave Scanning Radiometer 2 (AMSR-2), and the AMSR-Earth Observing System (AMSR-E). OSTIA uses both IR and MW sensors, including AVHRR, VIIRS, the Sea and Land Surface Temperature Radiometer (SLSTR), and AMSR-2. Despite different observation inputs, these datasets also use different algorithms and thresholds for the areas with ice cover. For ease of comparison, a summary of general information for the three datasets is given in Table 2.

Table 2 Descriptive summary of the SST datasets used in this study, where the abbreviations of Infra-Red (IR) and microwave (MW) sensors are listed here: the Advanced Very High Resolution Radiometer (AVHRR), the Visible/Infra-red Imager Radiometer Suite (VIIRS), the Spinning Enhanced Visible Infra-Red Imager (SEVIRI), the Advanced Microwave Scanning Radiometer 2 (AMSR-2), the AMSR-Earth Observing System (AMSR-E), and the Sea and Land Surface Temperature Radiometer (SLSTR)
Full size table

Detection of MHWs

Following previous studies, we defined MHWs as the discrete prolonged anomalously warm events based on the 90th percentile threshold of the climatology-removed SST anomalies from 1982 to 20206. The same period was used for calculating the daily climatology, which is simply defined as the mean SST on a given day of a year computed from the total 39 years. After that, an MHW event is identified when the abnormal warm period lasts over five consequent days or more at each grid point. The same processes were applied to all three datasets using their own thresholds. Note that although no moving window was applied to the 39-year daily climatology, our estimates on MHWs are consistent with previous studies13,15 and the discrepancies among the three datasets are still clear when using a 31-day moving threshold and 2-gap-day-allowed method6 (see Supplementary Fig. 1e−g).

After the detection of MHWs, two metrics (occurrence and duration) were estimated for an MHW event at each grid point. In addition, to investigate the long-term variations of MHWs, the monthly metrics were also calculated. Specifically, we calculated the ‘monthly occurrence’ based on the fractions of their total durations for the long-lasting MHWs that continued over several months9,64. For example, the occurrence of a 5-day MHW event from July 30 to August 3 will be counted as 0.4 in July (i.e., two days of its 5-day duration) and 0.6 in August, respectively. Consequently, for simplicity, the monthly duration is obtained by the sum of all MHWs in one month (i.e., MHW days in previous studies7), and the annual metrics were the sum of all months in a year. Accordingly, the long-term trends were calculated based on the Theil-Sen estimator using the Python package PyMannKendall65.