South Florida estuaries are warming faster than global oceans

From extensive evaluations, it is found that, of all satellite data products of sea surface temperature (SST), MODIS SST is the most appropriate in assessing long-term trends of water temperature in the South Florida estuaries. Long-term SST data show significant warming trends in these estuaries during both daytime (0.55 °C/decade) and nighttime (0.42 °C/decade) between 2000 and 2021. The warming rates are faster during winter (0.70 °C/decade and 0.67 °C/decade for daytime and nighttime, respectively) than during summer (0.48 °C/decade and 0.28 °C/decade for daytime and nighttime, respectively). Overall, the South Florida estuaries experienced rapid warming over the past two decades, 1.7 and 1.3 times faster than the Gulf of Mexico (0.33 °C/decade and 0.32 °C/decade for daytime and nighttime), and 6.9 and 4.2 times faster than the global oceans (0.08 °C/decade and 0.10 °C/decade for daytime and nighttime).


Introduction
Estuarine water temperature is a crucial driver of ecosystem structure and function, and can directly or indirectly affect physical, chemical, and biological processes, including photosynthesis and water stability (Livingstone andDokulil 2001, O'Reilly et al 2003) as well as fish species ranges (Sharma et al 2007) and species interactions (Hampton et al 2008). Estuaries are more sensitive to climate change as they respond more rapidly than oceans (Adrian et al 2009). Estuarine water temperature is influenced by climate and landscape factors (Sharma et al 2007), with the former including air temperature, cloudiness, and solar radiation (Shuter et al 1983, Oswald andRouse 2004) and the latter including mean depth and surface area (Wetzel 2001).
In a warming climate, a critical question is how estuarine temperature has changed over the past decades. Remotely sensed sea surface temperature (SST, • C) has often been used to study temporal changes at basin scale (e.g. Muller-Karger et al 2015) or local scale, including coastal waters, relatively large lakes, estuaries, reservoirs, and rivers (Costoya et al 2016, Virdis et al 2020, Zhang et al 2020, Torregroza-Espinosa et al 2021, Wang et al 2021. For South Florida estuaries, however, except for two studies (Boyer et al 1999, Robbins andLisle 2018) using limited in situ SST data to characterize temperature changes and one study (Carlson et al 2018) using 4 km remotely sensed SST to study long-term changes in Florida Bay (FB), it is unknown whether and how various satellite data can be used for the same purpose of assessing long-term SST changes.
The objective of this study is twofold. First, we evaluate the SST data products derived from various satellite measurements, from which the most feasible data product to study decadal changes in water temperature in several South Florida estuaries is determined. Then, using this data product, we quantify water temperature trends in the South Florida estuaries and compare/contrast these trends with those from adjacent open waters and from global oceans.  (c) Monthly MODIS SST images for each region in winter and summer, respectively. The gray buffer is the water-land boundary eroded inward by one MODIS 1 km pixel. Note that some pixels inside the estuaries have no valid SST data due to either small islands (e.g. in the FB) or MODIS data quality control. The sizes of LO, SLE, TB, CRE, and FB are 1730, 29, 1000, 56, and 2200 ecological pressure in the past decades, such as algal blooms, seagrass and coral dieoffs, and fish and lobster kills (Tomlinson et al 2004, Doering et al 2006, Peterson et al 2006, Turner et al 2006, Phlips et al 2015, Butler and Dolan 2017, Kramer et al 2018, Cannizzaro et al 2019. Numerous studies have documented the biological stress due to nutrient enrichment (e.g. Lapointe et al 2019, 2020 for FB), yet few studies focused on the thermal environment of these estuaries. To compare with temperature changes in open waters, the Gulf of Mexico (GoM) and global ocean open waters were also selected. To facilitate comparison, the GoM was divided into four quadrants, with the northern quadrant roughly coinciding with the exclusive economic zone of the United States and the eastern quadrant representing the area affected by the Loop Current (LC) and LC eddies.

Data sources and preprocessing
2.2.1. In situ data In situ surface temperature data for the study regions were obtained from the South Florida Water Management District (SFWMD), Florida Atlantic University-Indian River Lagoon Observatory Network (FAU-IRLON), National Ocean Service (NOAA NOS), Sanibel-Captiva Conservation Foundation River, Estuary, Coastal Observing Network (SCCF RECON), National Data Buoy Center (NOAA NDBC), Everglades National Park (ENP), and University of South Florida (figure 1, also see table S1 in the supplemental material). The data were considered to be the bulk temperature of the surface water. All data were collected at intervals of <1 h, which underwent extensive quality control by the data collection agency or organization.

Satellite data
Several remote sensing data products were selected to evaluate their feasibility to achieve the study goals. These included Landsat series, Advanced Very High-Resolution Radiometer (AVHRR) Pathfinder, ECOsystem Spaceborne Thermal Radiometer Experiment on the International Space Station (ECO-STRESS), and MODIS. Surface temperature inversion of data collected by the Landsat series (i.e. Landsat-4, 5, 7, and 8) from 1982 to the present is based on the open-source code provided by Ermida et al (2020), where SST was derived using the empirical relationship between the top of the atmosphere brightness temperature and SST in a single thermal infrared channel (Sun et al 2004, Freitas et al 2013, Li et al 2013, Duguay-Tetzlaff et al 2015 with water emissivity being fixed at 0.95. The AVHRR Pathfinder Version 5.3 SST dataset (PFV53) is a global twicedaily collection of 4 km SST data from 1981 to the present, generated by the SeaDAS-based system. The best quality PFV53 data were selected according to the quality level band. ECOSTRESS collects radiance with a 4-5 d revisit period, where surface temperature was derived through the Temperature Emission Separation algorithm (Gillespie et al 1998, Hulley and Hook 2010, Hulley and Freepartner 2019. The data quality and accuracy over South Florida waters have been evaluated by Shi and Hu (2021).
MODIS 1 km resolution SST data and ocean color data have been generated through a Virtual Antenna System (VAS, Hu et al 2013) using the NASA software SeaDAS (version 8). The data quality of MODIS SST is classified at 4 levels, with level 0 representing the highest quality and level >2 being invalid pixels subject to cloud contamination or other artifacts (Merchant et al 2005, Barnes et al 2011). For the winter months of December-February and summer months of June-August, SST images are visually inspected to select pixels with quality values of 0-2 while masking pixels with cloud contaminations as determined by the corresponding ocean color Red-Green-Blue true color images. This is to avoid masking valid pixels with extremely low or high SST when using quality levels of 0-1. This manual process was not applied to nighttime SST data because validation results using quality levels of 0-2 without manual masking suggested validity of the nighttime data. For other months, SST image pixels with quality values of 0 and 1 are used directly.

Satellite data evaluation
To evaluate remotely sensed SST data for the study regions, only in situ stations with a distance to land greater than 1.5 times of pixel size were selected. This is to prevent possible land-adjacency contamination of image pixels. Remotely sensed SST at each site was calculated as the average of 3 × 3 pixels centered on the site, with the time difference limited to ±1 h between the two measurements. Then, a spatial homogeneity test was used to discard pixels with large spatial variability (coefficient of variation > 0.15) in the matchup selections. The accuracy of remotely sensed SST was assessed using several statistical measures, including linear regression and coefficient of determination (R 2 ), absolute deviation (bias), and root mean square difference (RMSD).

Determine long-term trend
Long-term SST time series consists of three additive components: trend, seasonality, and noise. The trend component describes gradual changes over a time scale of >1 year (Verbesselt et al 2010). The seasonality component represents regular and periodic changes on an annual scale (Zhan et al 2014a(Zhan et al , 2014b. The noise component represents random and irregular changes caused by observation conditions, atmospheric environment, and disturbance events. This study utilizes the time series decomposition method in the statsmodels package of Python for deseasonalization to extract the trend component in the longterm MODIS SST during daytime and nighttime, respectively, which is labeled as SST t below (the subscript 't' represents 'trend'). Specifically, the decomposition uses a moving window of the same size of the seasonality (12 months in this study) to calculate the average value within the window, and then generate the trend component from the long-term averages. Then, the seasonal factors and noise component are estimated from the de-trended data.
The detection and estimation of trends in the SST time series of the trend component were performed using the modified Mann-Kendall test (Hamed and Rao 1998), which improves upon the nonparametric Mann-Kendall test to account for autocorrelation in the time series, and Sen's slope estimates (Sen 1968) through the PyMannKendall package of Python, with the results tested at the 95% confidence level.

Which SST data product to use?
All satellite-based SST data products, except for the AVHRR pathfinder data, showed agreement with in situ water temperature (see figures S1 and S2 for Landsat and AVHRR, and Shi and Hu (2021) for ECOSTRESS). Of particular importance are those from Landsat series and the recent ECOSTRESS sensors, as they provide data of high spatial resolution for estuaries. However, one drawback of these data for trend analysis is their limited temporal coverage, as shown in figure S3. Landsat 60-120 m SST has a 16 d revisit frequency, and in the 1980s and 1990s the temporal coverage was especially scarce. ECOSTRESS 70-m SST has more frequent revisits but the data collection did not start until 2018, thus not sufficient to evaluate trends.
In this regard, MODIS SST appears to be the most feasible to assess decadal changes. Not only is the accuracy acceptable (figure 2), but the two sensors (on Terra and Aqua, respectively) can provide valid SST observations in relatively large estuaries every 2-3 d after masking clouds and other artifacts. Furthermore, statistics of valid observations showed no systematic bias in either time (figure S4), space (figure S5), or long-term annual patterns (figure There is not enough valid daytime data for the SLE region due to strict data quality control. S6). Although the statistics (R 2 , RMSD, bias) are different across different water bodies, which are likely due to differences in the number of matching pairs used in the statistics and due to different stratification (MODIS instruments measure the skin temperature while in situ sensors measure the bulk temperature), they all show agreement between MODIS and in situ SST without variable biases in different SST ranges. All these characteristics make it suitable to use MODIS SST to study long-term changes.

SST trends in South Florida estuaries 3.2.1. Long-term trends
The long-term de-seasoned and de-noised SST t are shown in figure 3, with their changing rates presented in table 1. From 2000 to 2021, the South Florida estuaries warmed by an average of 1.21 • C during the day and 0.92 • C at night. All estuaries showed statistically significant warming trends, with average warming rates of 0.55 • C/decade during the day and 0.42 • C/decade at night, respectively. The LO region has warmed the most over the past 22 years by 1.76 • C and 1.21 • C during the day and at night, respectively. This is followed by the TB (1.17 • C and 0.92 • C), the CRE (1.03 • C and 0.95 • C), and the FB (0.88 • C and 0.84 • C).

Seasonal trends
The fluctuation of water temperature in South Florida estuaries is significantly higher in winter than in summer (figure 4), probably because heat loss from the water column in winter is more susceptible to varying degrees of the storm and cold-air fronts   The warming rates of water temperature in winter are much higher than in summer. The different warming rates between daytime and nighttime in summer are likely due to solar insolation. Across different water bodies, LO showed the highest warming rate while FB showed the lowest warming rate during the day and night in winter, respectively. In contrast, the regional difference in warming rates is minor during daytime and nighttime in summer.

South Florida estuaries are warming faster than open oceans
The four quadrants of the GoM have warmed by an average of 0.73 • C during daytime and 0.69 • C at night from 2000 to 2021 (figure 3 and table 1), with warming rates of 0.33 • C/decade during daytime and 0.32 • C/decade at night, respectively. Unusually cold winters of 2010 and 2011 somewhat offset the warming trends in these quadrants (Muller-Karger et al 2015), for otherwise the warming trends (without these cold winters) would be higher ( figure S7). Over the same period, the global oceans have warmed by 0.18 • C during daytime and 0.22 • C at night, with warming rates of 0.08 • C/decade and 0.10 • C/decade, respectively. Overall, South Florida estuaries are warming 1.7 and 1.3 times faster than the GoM open waters during daytime and at night, respectively, and 6.9 and 4.2 times faster than the global oceans.

Discussion
Because of the different hydrodynamic environments, it is not surprising to see that estuaries respond to climate change differently than open oceans. In this case, South Florida estuaries are found to warm at a much faster rate than either GoM open waters or global oceans. While this is the first time such a fastwarming rate is found in South Florida estuaries, similar findings have been reported for several other estuaries. For example, in situ temperature data showed faster warming rates than global oceans for the Chesapeake Bay between 1949 and 2002 (Preston 2004), for the Long Island Sound estuary between 1991 and 2013 (Staniec and Vlahos 2017), and for many of the 166 estuaries along the Australian coastline from 2007 to 2019 (Scanes et al 2020). Likewise, remotely sensed SST from MODIS and other sensors showed faster warming rates than global oceans for the Yangtze River Estuary from 1982 to 2017 (Wang et al 2021), for Tonle Sap Lake in Cambodia from 1988 to 2018 (Daly et al 2020), and for some lakes in the Tibetan Plateau from 2001 to 2012 (Zhang et al 2020).
The warming rates reported here are different from those estimated from in situ temperature data alone for the period of 1978-2008 (mean warming rate of 2.80 × 10 −3 • C yr −1 in three Florida estuaries, 3.18 × 10 −3 • C yr −1 for TB) (Robbins and Lisle 2018). It is unclear whether this is due to the difference in the reporting period (2000-2021 versus 1978-2008), but the limited in situ data may be a reason. Indeed, using AVHRR SST data of the past 35 years, Carlson et al (2018) found a warming rate of 0.03 • C yr −1 in the FB for the month of August, consistent with the findings of this study of ∼0.4 • C/decade for the same estuary. However, the 4 km AVHRR SST data are not applicable for smaller estuaries, and their accuracy is also questionable for smaller estuaries (figure S2), making the 1 km MODIS SST data an optimal choice to study temperature changes in the past two decades for South Florida estuaries.
One may question whether some of the observed long-term trends could be due to temporal or spatial aliasing, as MODIS data may be unevenly distributed in either time or space, especially across different years. For example, during the same month, the temporal or spatial distributions of valid MODIS data may be different across different years. However, a detailed view of the distributions of daily data suggests that this is unlikely (figures S4 and S5). For the long-term trend analysis, while it is true that the number of valid pixels used to calculate seasonal averages varies from year to year, there does not appear any annual variation in the number of valid data points to mimic any of the SST annual variations (figure S6), thus suggesting the validity of the estimated seasonal trends.
The results showed different warming rates from different estuaries, which may be due to multiple factors such as latitude, estuary type, average depth, and flushing time of the estuary (Scanes et al 2020). For example, although the latitude (26.95 • and 26.46 • , respectively) and mean depth (both ∼2.7 m, Phlips et al (2020), Wan et al (2013)) of the LO and CRE regions are similar, the warming rate of the former is significantly faster than the latter, possibly due to the longer residence time of the LO (3.5 years, James and Pollman (2011)) than the CRE (4-80 d, Wan et al (2013)). However, with the limited number of waterbodies considered in this study (table S2), it is rather difficult to isolate one factor while keeping all other factors the same (similar to a 'controlled' experiment). A follow-on study may expand to other estuaries around the GoM and in other mid-latitude regions to design such a 'controlled' experiment to pinpoint the reasons behind the different warming rates. Likewise, most of the waterbodies showed faster warming rates in winter than in summer (table 1). Although this is in line with the empirical rule of climate change (i.e. cold 'things' warm faster than warm 'things' , see www.climate.gov/news-features/blogs/beyond-data/ climate-change-rule-thumb-cold-things-warmingfaster-warm-things), more waterbodies are needed in a future study to understand the specific reasons.
Based on the report of the Intergovernmental Panel on Climate Change, global ocean mean SST is likely to increase by 1.6 • C by 2050 even under the best climate mitigation strategy (IPCC 2019). Under such a projected climate regime and assuming South Florida estuaries will continue to warm at rates 4-7 times faster than global oceans, by 2050 these estuaries will have SST increases by 6 • C-11 • C. While these dramatic increases may be questionable, South Florida estuaries are likely to warm faster than global oceans in the foreseeable future for the same reasons behind the observed SST trends, which has significant applications to the local ecosystems. For example, unusually high SST in the FB during the summer of 2014 caused massive coral bleaching in the Florida Keys Reef Tract (Manzello 2015). The dense turtle grass beds in FB were destroyed in 2015, and high SST was thought to be a synergistic stressor, interacting with high salinity and sulfide toxicity to cause seagrass mortality (Carlson et al 2018). Likewise, more extreme events may occur, including winter cold events, making thermal stress a threat to the ecosystem rather than just a compounding factor. Indeed, South Florida experienced cold winters from 2009 to 2011 (figure 4), leading to mass mortality of manatees (Barlas et al 2011, Hardy et al 2019 and severe coral mortality in the Florida Keys Reef Tract (Lirman et al 2011). Continuous monitoring and assessment of water temperature in South Florida estuaries will help implement adaptation strategies under a changing climate. Likewise, it is desirable to extend the study domain in the future to include other major estuaries in North America to have a systematic evaluation of coastal thermal environment.

Conclusions
Of all SST data products, MODIS 1 km SST is found to be the most appropriate for evaluating decadal temperature changes in South Florida estuaries. Statistical analysis of MODIS SST time series revealed much faster warming trends from 2000 to 2021 during both daytime and nighttime in these estuaries than in GoM open waters and global oceans. The warming rates in both daytime and nighttime are also much higher in winter than in summer for most estuaries. Whether such trends will continue under a changing climate requires continuous monitoring and assessment of water temperature in these ecologically and economically important estuaries.