Quantifying Uncertainties in CERES/MODIS Downwelling Radiation Fluxes in the Global Tropical Oceans

The Clouds and the Earth's Radiant Energy System program, which uses the Moderate Resolution Imaging Spectroradiometer (CM), has been updated with the launch of new satellites and the availability of newly upgraded radiation data. The spatial and temporal variability of daily averaged synoptic 1-degree CM version 3 (CMv3) (old) and version 4 (CMv4) (new) downwelling shortwave (Q S ) and longwave radiation (Q L ) data in the global tropical oceans spanning 30°S–30°N from 2000 to 2017 is investigated. Daily in situ data from the Global Tropical Moored Buoy Array were used to validate the CM data from 2000 to 2015. When compared to CMv3, both Q S and Q L in CMv4 show significant improvements in bias, root-mean-square error, and standard deviations. Furthermore, a long-term trend analysis shows that Q S has been increasing by 1 W m − 2 per year in the Southern Hemisphere. In contrast, the Northern Hemisphere has a − 0.7 W m − 2 annual decreasing trend. Q S and Q L exhibit similar spatial trend patterns. However, in the Indian Ocean, Indo-Pacific warm pool region, and Southern Hemisphere, Q L spatial patterns in CMv3 and CMv4 differ with an opposite trend (0.5 W m − 2 ). These annual trends in Q S and Q L could cause the sea surface temperature to change by − 0.2 to 0.3 °C per year in the tropical oceans. These results stress the importance of accurate radiative flux data, and CMv4 can be an alternative to reanalysis or other model-simulated data.


Introduction
The performance of climate and weather forecasting models depends on the availability of high-resolution accurate surface radiation datasets and their assimilation into climate and weather forecasting models [1].The atmosphere and ocean interact through mass, momentum, and heat fluxes at the ocean-atmosphere interface.Heat fluxes modulate intraseasonal oscillations in tropical oceans [2][3][4].Shortwave radiation and latent heat flux are the principal contributors to heat flux variation in the tropics [5].Tropical oceans receive the largest solar irradiance in the form of shortwave radiation.The excess heat over the tropical ocean balances through turbulent mixing, longwave radiation, and transport to the higher latitudes through ocean circulation for more extended periods [6].On weekly timescales, tropical and extratropical cyclones make up the most prominent energy transfer from the ocean to the atmosphere through the release of latent heat [7].Heat fluxes also influence large-scale tropical coupled processes such as the El Niño Southern Oscillation (ENSO), Indian Ocean Dipole (IOD), and Atlantic Meridional Mode [8][9][10][11][12][13].Ocean general circulation models (OGCMs) use the near-surface atmospheric state to parameterize the heat and momentum fluxes [14].Similarly, uncertainties in turbulent and radiative heat fluxes could affect the heat balance in the boundary conditions prescribed in the OGCMs [15,16].These limitations in the forcing fields are partially responsible for the OGCMs making unrealistic simulations of intraseasonal, seasonal, interannual, and climate variability [17].Hence, accurate near-surface atmospheric fields are essential for realistic simulations in a forced ocean and climate model.
The sum of the net radiative flux and turbulent flux adds up to the net air-sea flux at the sea surface [18][19][20].Longwave and shortwave radiation are the 2 components of radiative flux.The downwelling fluxes of shortwave (Q S ) and longwave radiation (Q L ), together with other meteorological variables and the initial ocean state, are used to force the OGCMs.A significant extent of the ocean model simulation accuracy depends on the accuracy of these downwelling radiative fluxes.At present, the ocean modeling community uses Common Reference Ocean-Ice Experiments (CORE-II) and Japanese 55-year atmospheric reanalysis (JRA)-do datasets as the prime source to run the global ocean and sea-ice model [14,[21][22][23].The National Centre for Environmental Prediction (NCEP-R2) released corrected data as CORE-II [24], and the improved JRA reanalysis is called JRA-do.However, CORE-II uses radiation data obtained from the International Satellite Cloud Climatology Project [25], available only until 2009.Hence, CORE-II forcing fields are not available beyond 2009.Therefore, climate forecasting system reanalysis (CFSR, CFSv2) [26,27] replaced the NCEP-R2 data.Although other radiation datasets have taken over the International Satellite Cloud Climatology Project since 2009, there are still gaps in this area.
Kumar et al. [28] used satellite data (Moderate Resolution Imaging Spectroradiometer [MODIS]) as the reference to evaluate the radiative fluxes from the Ocean Moored Network for the Northern Indian Ocean (OMNI).Rahaman and Ravichandran [29] evaluated the near-surface air temperature and humidity from CORE-II, Objectively Analyzed Air-Sea Fluxes, and Tropical Flux to find a better dataset to use in the model forcing.An evaluation of the surface radiation fluxes with buoy observations in the Pacific Ocean during 2000 to 2012 concluded that satellite data match well with the in situ observations, followed by reanalysis and model data [30][31][32].Trolliet et al. [33] compared the irradiance data from MERRA-2 and ERA5 and 3 other satellite-derived datasets: HelioClim-3v5, Surface Solar Radiation Data Set -Heliosat (SARAH-2), and Copernicus Atmosphere Monitoring Service with 5 buoys in the Atlantic for the period 2012 to 2013.At the ocean-atmosphere interface in the Atlantic, heat budgets derived from satellites and blended products were compared with in situ observations during 2003 to 2005 [20].While the performances are similar between the 3 satellite-derived datasets, existing reanalysis data have significant biases, errors, and poor correlation values compared with independent in situ observations.In addition, the Earth's radiation balance from satellite observations has a more significant bias over the ocean than their better agreement over land [34].These differences are due to frequent changes in satellite observing systems, the degradation of sensors, the restricted spectral intervals and viewing geometry of sensors, and changes in the quality of atmospheric inputs that drive the inference schemes.

Global tropical moored buoy array
Moored buoy observing systems in all 3 tropical oceans comprise the Global Tropical Moored Buoy Array (GTMBA).This program is a multinational effort to obtain surface meteorological and subsurface oceanic near-real-time data for research and applications.It has 3 components, namely, the Tropical Atmosphere Ocean/Triangle Trans-Ocean Buoy Network (TAO/TRITON) in the tropical Pacific [35], the Pilot Research Moored Array in the Tropical Atlantic (PIRATA) in the tropical Atlantic [36], and the Research Moored Array for African-Asian-Australian Monsoon Analysis and Prediction (RAMA) in the tropical Indian Ocean [37,38].The GTMBA program in the Indian Ocean (RAMA) is relatively new, as this program was first started in the Pacific and later followed by the Atlantic.The GTMBA data undergo rigorous 3-stage quality control (daily, weekly, and monthly) procedures to ensure high accuracy standards before being delivered to users [39][40][41][42][43].
The uncertainty in Q S in the GTMBA data is 2% due to drift criteria, and the reported monthly mean accumulation biases due to high-dust accumulation can reach −200 W m −2 in the Atlantic.Similarly, record-length mean biases in the Q S from the moorings in the Atlantic (PIRATA) can reach −10 W m −2 , potentially leading to significant negative Q S biases ( [15,44] and the references detailed the GTMBA project).Among the 27 mooring sites in the Indian Ocean, daily averaged (0000 UTC to 2300 UTC) Q L data are available only at 0°N, 80.5°E; 15°N, 90°E and 8°S, 67°E (3 locations), while daily averaged Q S data (0000 UTC to 2300 UTC) are available at 19 sites during the study period.Among the 7 stations delivering Q L in the Atlantic Ocean, only 4 have consistent data, and Q S is available from 17 of 21 locations during the study period.The Pacific Ocean has more moorings with better data availability than other oceans.We use the Q L data delivered from 11 moorings and Q S data obtained from 32 of 34 sites in the Pacific Ocean.We use all available mooring locations for Q S and Q L in the tropical oceans during 2000 to 2015.As a result, more Q S observations with a wider distribution are available than fewer and more sparsely distributed Q L observations in the tropical oceans.However, many buoy locations have data gaps, and few stations exist where only Q S or Q L are available.Therefore, the data availability period of all mooring sites is not the same.Data gaps and the spuriousness of mooring data are considered when validating satellite data with GTMBA data.Figure S1 shows the present status of GTMBA and the location of buoys.

Clouds and the Earth's Radiant Energy System (CERES/MODIS)
The Clouds and the Earth's Radiant Energy System (CERES) project started in 1997 was conceived as a successor to the Earth Radiation Budget Experiment to compile a data record for the investigation of interannual variations in climate [45][46][47][48][49][50].This program also provides an alternative for the radiative flux components available from 2000 to the present [46,48,50].In total, 7 instruments have been launched, with the latest instrument (FM6) launched in 2017 on the National Oceanic and Atmospheric Administration's (NOAA) Joint Polar Satellite System 1.The release of the instrument and Earth Radiation Budget Experimentlike data from FM6 occurred in June 2018.The instruments and platforms used to collect these data include imaging radiometers on the Geostationary Satellites platform; CERES Flight Model 1 (FM1), CERES FM2, CERES Scanner, and MODIS on Terra; and CERES FM3, CERES FM4, and MODIS on Aqua.
The CM project produces a long-term, integrated global climate data record for detecting decadal changes in the Earth's radiation budget from the surface to the top of the atmosphere.Thus, the CM program supports climate model evaluation and improvement through model-observation intercomparisons.CM is the only project to produce global climate data records of Earth's radiation budget using polar-orbiting and geostationary satellites accounting for variations in radiation at hourly, daily, and monthly timescales and at spatial scales ranging from 20 km to 1°.The CM program focuses on measuring outgoing longwave radiation radiances to an accuracy of 1% and reflected solar radiances to 2%.CM estimates of incident solar radiation agree better with surface measurements at monthly rather than Downloaded from https://spj.science.orgon February 01, 2023 at daily timescales.These estimates can also capture the seasonal variation in incident solar radiation very well [51].Barkstrom [52] and Smith et al. [53] presented complete technical details on the status of CM.
Past studies have evaluated radiative fluxes from different satellites with observations over the land and ocean [20,[30][31][32]54,55].Rutan et al. [56] compared CM surface radiation flux data at 85 globally distributed land (37) and ocean buoy (48) surface observations as well as several other publicly available products on global surface radiation flux data.The downward fluxes from synoptic 1-degree (SYN1deg) have a monthly bias (standard deviation) of 3.0 W m −2 (5.7%) for Q S and −4.0 W m −2 (2.9%) for Q L compared to surface observations.Inclusion of the diurnal cycle of cloud changes minimized the standard deviation between surface Q S flux calculations and observations at the 3-hourly timescale.Kato et al. [48] estimated the bias (root-mean-square error [RMSE]) between computed and observed monthly mean irradiances calculated with 10 years of CM data as 4.7 (13.3)W m −2 for Q S and −2.5 (7.1) W m −2 for Q L over global oceans.Nevertheless, all these studies either have a single downwelling radiation parameter or focus on a narrow region for a brief period.Venugopal et al. [16] and Thandlam and Rahaman [15] performed a similar analysis in the global tropical oceans, with a subset of CERES/MODIS data, hereafter CM, version 3 datasets during 2000 to 2009.In the present work, we evaluate both the components of downwelling radiation (Q S and Q L ) from CM version 3 (CMv3) and CM version 4 (CMv4) for a longer period (2000 to 2017).To our knowledge, no study has focused on validating near-surface Q S and Q L from updated CMv4 data in tropical oceans with in situ observations for such a long period.
Hence, this study aims to undertake a comparison of the Q S and Q L from the regularly updated downwelling radiation from satellite data over tropical oceans with in situ observations to assess whether CMv4 data can be complementary to reanalysis data to force OGCMs and can be used to evaluate climate models.Evaluating both versions with independent in situ observations over the global tropical oceans could give a glimpse of their performance and help to choose them in developing hybrid/ blended forcing data, including other atmospheric and ocean variables.We also study the spatial variability of CM (Q S and Q L ) available during 2000 to 2017.This paper is organized as follows.Materials and Methods describes the various datasets used in the study.Results and Discussion includes the evaluation of satellite data with in situ observations.This section also includes spatial variability of downwelling radiation (Q S and Q L ) from CMv3 and CMv4 in the global tropical oceans at different time scales.We also quantified the annual changes in sea surface temperature (SST) associated with annual trend in Q S and Q L from CMv4.We also illustrate the details and results of the validation process with reference data from the GTMBA [37,38].The main findings and highlights of the study are offered in Conclusions and Summary.

Materials and Methods
We use GTMBA mooring radiation data [37,38] to evaluate Q S and Q L fluxes from CM available in real time during 2000 to 2015.We compare these products with GTMBA collocated and concurrent datasets with nearest grid-point data.Trolliet et al. [33] show that comparing point location with gridded data is valid over the tropical Atlantic Ocean since no strong systematic gradient in irradiance was present over a short distance in the tropical oceans.The mean, standard deviation, and RMSE are computed using Eqs. 1 to 3 from Ref. [15].
Multiple data collection systems provide data from the GTMBA.The older "Next Generation ATLAS" systems have been replaced in TAO by the National Data Buoy Center TAO Refresh system and are in the process of being replaced by the Pacific Marine Environmental Laboratory with the recently developed "T-Flex" systems in PIRATA and RAMA.A pyranometer is used to measure the downwelling radiation fluxes with a resolution of 0.4 W m −2 , a range of 0 to 1600 W m −2 , and an accuracy of ±2% [44,57].The downwelling radiation data are sampled and stored at 1-min intervals with a sampling rate of 1 Hz [58].The hourly mean and standard deviation of data are transmitted in real time through the Global Telecommunications System (GTS).We used all-weather daily averaged (0000 UTC to 2300 UTC) GTMBA data obtained from the GTMBA data delivery platform to compensate for the daily averaged CM data.Although the real-time data have high frequency with cloud variability on a buoy compared to time-averaged values, the unavailability of high-frequency data over a few time steps in each day could lead to bias in the data.
Table 1 shows the details of the GTMBA data used in the present study.Similarly, Fig. 1 shows the frequency distribution of GTMBA data in the tropical oceans during 2000 to 2015.Except in the higher range (300 to 350 W m −2 ), the number of shortwave observations in the Indian Ocean and Atlantic Ocean shows a large difference with those in the Pacific Ocean (Fig. 1A).On the other hand, the Atlantic receives no longwave observation beyond 450 W m −2 (Fig. 1B).In the present study, we use observed geostationary enhanced, temporally interpolated surface radiative fluxes for all-sky conditions CMv3 and CMv4 (CM version 4) [59] [16,60,61].We use the daily averaged CM data to evaluate daily in situ observations from the GTMBA.To glimpse the spatial patterns in the data, Fig. S2A shows the 3-h Q S data from CMv3 on 2017 February 28, and Fig. S2B shows the global daily mean of computed Q S from the 3-h data.Rutan et al. [56] detailed the product computation, methods, and validation of CM.Finally, we use EN4-Hadley interpolated monthly mean temperature profiles with 1° spatial resolution to compute mixed-layer depth (MLD) climatology in the global tropical oceans.This is used further to estimate the annual change in SST due to the annual trend in Q S and Q L from CMv4.These datasets are obtained from Met Office Hadley Centre observations datasets during 1993 to 2017 [62].These temperature profiles are at 42 depth levels from 5 m to 5,250 m below the ocean surface.

Results and Discussion
The Indian Ocean Figure 2 shows the frequency distribution of downwelling shortwave and longwave radiation data in the Indian Ocean.
In this study, we use all the available concurrent and collocated ) in tropical oceans using daily data during 2000 to 2015.
Downloaded from https://spj.science.orgon February 01, 2023 data from the RAMA buoy locations (19).This number turns out to be 42,412 and is the same for all (RAMA, CMv3, and CMv4).Most of the Q S values lie in the 200 to 300 W m −2 range.Satellite-derived products underestimate the lower (<150 W m −2 ) and higher range values (>300 W m −2 ).Slightly better agreement in the CMv4 Q S product is seen concerning observations in the lower (<150 W m −2 ) and higher (>300 W m −2 ) ranges compared to CMv3.The mismatch in the satellitederived products with the lower and higher limits of Q S observations in the Indian Ocean could be due to the retrieval error or may have arisen due to the large intraseasonal, seasonal, and interannual variability over this region.Both satellite-derived products overestimated the Q S in the 150 to 250 W m −2 range compared to the observations (Fig. 2A).The maximum number of observations (~14,000) falls in the 250 to 300 W m −2 range, and both satellite-derived products are near this observed value.On the other hand, most observations and satellite Q L data lie in the range of 350 to 450 W m −2 .Q L data retrieved from the CM show a good fit with observations in all ranges (Fig. 2B).Thus, both versions of CM are coherent in sensing the downwelling longwave data in the Indian Ocean.Figures 3 and 4 show the temporal variability of Q S and Q L , respectively, over selected locations with continuous data availability.Each subpanel also gives the mean and standard deviation values of the observation and satellite products.Both CMv3 (red) and CMv4 (blue) show similar temporal variability with observations (gray).While Q S from observations and both versions of CM data in the equatorial region show the absence of prominent seasonal variability (Fig. 3B to D), the seasonal changes are significant in the northern Bay of Bengal (Fig. 3A) and the thermocline ridge region over the southwestern Indian Ocean (Fig. 3E).The observed Q S over the northern Bay of Bengal (Fig. 3A) peaks in the spring due to clear sky conditions with the highest solar insolation [63] and shows low values in the summer and winter months to cloudy conditions and precipitation during the summer and winter monsoons.below 150 W m −2 over different stations could be due to the weakly resolved radiation transfer algorithms during the presence of deep convective clouds and aerosol optical depth [64].The satellite-derived Q S products can capture the observed temporal variations over all the RAMA buoy locations (Fig. 3A to E).The statistical values show that the CMv4 product performs slightly better than CMv3 in terms of the mean and standard deviation compared with the observed values (Table 2).Hence, the CMv4 product could be a better option when studying Indian Ocean radiative flux variability studies and validating other products.Contrary to Q S , Q L shows large values over the northern Bay of Bengal during the summer and winter months due to cloudy conditions and low values during spring due to relatively clear sky conditions (Fig. 4A).Although the CMv4 Q L mean (411.5 W m −2 ) is close to the observed mean, the CMv3 data show a better standard deviation (26.3 W m −2 ) in the northern Bay of Bengal.On the other hand, Q L from CMv4 shows better agreement with observations both in standard deviation and mean (Fig. 4A).Following the Q S , Q L also shows no seasonal or intraseasonal variability over the equator, and values lie above 400 W m −2 ; thus, this region receives large Q L irrespective of the season.The location of this station is in the eastern equatorial Indian Ocean, and these high values could correspond to persistent convective clouds due to the presence of the intertropical convergence zone (ITCZ) throughout the year [65].However, there is no significant improvement in the correlation (r) between observations and Q L from CMv4 (r = 0.89) compared to CMv3 (r = 0.87) in the Indian Ocean (Fig. S4A  and B; Table 2).This could be due to Q L being underestimated by CMv4 and CMv3 in the higher range (>450 W m −2 ), as shown in Fig. S4A and B, respectively.
Table 2 shows the statistics of both Q S and Q L from all stations in the Indian Ocean compared with observations during 2000 to 2015.Both CMv3 and CMv4 overestimate the Q S with a negative bias of −1.43 W m −2 in CMv4 and −4.5 W m −2 in CMv3.It is worth mentioning that this daily mean bias is much lower than the earlier reported monthly biases compared with similar buoy observations [48,56].The CMv4 RMSE (27.34 W m −2 ) is significantly lower than that of CMv3 (41.33 W m −2 ).However, both satellite products cannot capture the variability in standard deviation, yet CMv4 is closer to the observed values than CMv3.The Q S in CMv4 shows a better correlation (r = 0.92) than that in CMv3 (r = 0.80) (Fig. S3A and B).Q L underestimates the observations with a positive bias of 6.95 W m −2 and 3.15 W m −2 in CMv3 and CMv4, respectively.Other statistical values also show improvement in CMv4 compared to CMv3 (Table 2).

Tropical Atlantic Ocean
Although the tropical Atlantic is the smallest in the tropical oceans, the intraseasonal and interannual variability of downwelling radiation over this region is key to global climate processes.Among other atmosphere-ocean phenomena, the ITCZ variability twice a year defines the evolution of synoptic-scale oscillations in the Atlantic.The fluctuations in downwelling radiation received due to variability in cloud cover control the intensity of the Atlantic zonal mode and meridional modes in the tropical Atlantic [10].Additionally, these synoptic oscillations induce changes in the global monsoon precipitation patterns and the poleward movement of heat and momentum [66,67].The frequency distribution of Q S and Q L in the Atlantic is shown in Fig. 5. Unlike the data-scarce Indian Ocean, the number of observations is greater in the Atlantic Ocean.Satellite-derived  5A).The discrepancies in the satellite estimations of shortwave radiation could be due to the ineffectiveness of radiative transfer models in estimating Q S during cloudiness and precipitation.
Nevertheless, Q L in CMv4 shows better agreement than that in CMv3 (Fig. 5B).Most of the observed values lie in the 350-450 W m −2 range, and both satellite products could capture the observed distribution in this range, with CMv4 being slightly better than CMv3.However, downwelling fluxes are primarily a function of local cloud properties, and the discrepancies between satellite products and in situ measurements in the lower and higher ranges might also be due to errors in retrieving cloud optical thickness and surface albedo in the satellite data [68].
Furthermore, the temporal variability of Q S from PIRATA compared with CMv3 and CMv4 is shown in Fig. 6.The prominent intraseasonal oscillations in Q S dominate the seasonal and interannual oscillations over the equator (Fig. 6b).However, the seasonal cycle is more dominant off the equator than the intraseasonal oscillation (Fig. 6A, C, and D).The northward movement of the ITCZ in boreal spring increases the cloud cover, leading to low Q S over the equator.CMv4 shows a better standard deviation (37.1 W m −2 ) with observations (44 W m −2 ), and both versions of CM capture these oscillations.Seasonal and interannual oscillations of Q S are dominant north and south of the equator (Fig. 6A, C, and D).However, Q S in the North Atlantic shows low values in the winter and peaks in the boreal summer and vice versa in the southern Atlantic.Thus, the ITCZ movement plays a key role in modulating the intensity of Q S in the Atlantic and the SST, continental monsoon forcing, and air-sea interactions [69].CMv4 shows an improved standard deviation over CMv3 compared with observations over all locations.This enhancement in CMv4 is also reflected in the correlation values (r = 0.89), which is low in CMv3 (r = 0.79) (Fig. S3C and D).
Similarly, Q L from CMv4 is also controlled by the location of the ITCZ and shows better agreement with standard deviation values nearing the observations over all locations.However, correlation values (r = 0.91) show no improvement in CMv4 compared to CMv3 (r = 0.91, Fig. S4C and D).Nevertheless, both products display a tight fit with the observations with less scattering in the higher range, as noticed in the Indian Ocean.While Q L in the equator and North Atlantic has seasonal and interannual variability (Fig. 7A and B), south of the equator shows no such signal (Fig. 7C).Substantial intraseasonal variability dominates Q L at 10°S instead.Q L starts to peak in boreal spring at the equator, shows low summer values and is opposite in the north.At the same time, Q L in the South Atlantic shows no such strong signal.Table 3 shows the detailed statistics of satellite data compared with observations over all locations in the Atlantic.Similar to the Indian Ocean, Q S in both CMv4 (−7.4 W m −2 ) and CMv3 (−4.4 W m −2 ) overestimate the buoy observations in the Atlantic.The CMv4 RMSE (26.18 W m −2 ) is also lower than that of CMv3 (35.84 W m −2 ).The variability represented by the standard deviation is also closer to the observed values in CMv4 than in CMv3.Although Q L in CMv4 and CMv3 underestimated the observations with a positive bias, CMv4 has better estimates with lower bias.All other statistical values of Q L in CMv4 also agree well with the observations.

Tropical Pacific Ocean
The availability of in situ data in the tropical Pacific is greater than that in the Atlantic and Indian Oceans.Figure 8 shows the frequency distribution of both Q S and Q L collocated with in situ observations for different ranges in the tropical Pacific.We used concurrent data; hence, the total numbers are the same for in situ and satellite products.The Q S shows a large range with a major distribution ranging from 50 to 350 W m −2 .On the other hand, Q L from both observations and CM show values between 300 and 400 W m −2 .Similar to the Indian Ocean, the Q S from CM significantly differ from the observations in lower and higher ranges, while the Q L from CM shows no such difference from the observations in the tropical Pacific.
Figure 9 shows the temporal variation in Q S from both CMv3 (red) and CMv4 (blue) compared with TAO/TRITON (gray) during 2000 to 2015 over the selected locations in the tropical Pacific Ocean.Gaps in the observation time series (gray lines) are due to the unavailability of TRITON data during those periods.The satellite-derived products have captured the observed strong intraseasonal variability of Q S in the western tropical Pacific Ocean (Fig. 9A and B: 170°W and  165°E).These 2 buoys are in the west Pacific warm pool region, and the intraseasonal oscillation variability could be due to strong convection and large variability in the amount of cloud fraction over the warm-pool region [70].Thus, the Q S over the eastern tropical Pacific varies between 150 and 350 W m −2 .CMv4 agrees with TRITON data with better mean and standard deviation values than CMv3 over the eastern Pacific with strong seasonal variations.In particular, both 110°W and 140°W stations persistently receive large Q S above 200 W m −2 .Additionally, the variability in Q S over both the western (165°E) and eastern (110°W) Pacific shows a strong ENSO impact.This can be seen in the dipped Q S values.Q S dropped below 200 W m −2 for a brief period over these stations (Fig. 9C and D) during the ENSO events of 2002 and 2009 to 2010 and could be associated with strong convection and cloud formation, which are opaque to the downwelling shortwave radiation [71].Q S from CMv4 shows better correlation (r = 0.90) with observations than CMv3 (r = 0.79) (Fig. S3E and F).Additionally, a low RMSE (28.91 W m −2 ) and bias (0.67 W m −2 ) in CMv4 highlight its improved performance in the tropical Pacific as well (Table 4).
Q L from CMv4 in the tropical Pacific shows better performance than CMv3.However, marginal improvements were observed in parameters such as standard deviation, RMSE, and correlation (Table 4).Satellite data did not capture higher observations and show large biases, as shown in Fig. S4E and F. Nevertheless, Q L in CMv4 overestimated the observations with a negative bias (−5 W m −2 ), while CMv3 underestimated the TRITON data with a positive bias (3.26 W m −2 ).All statistics of both Q S and Q L compared with the TRITON data are shown in Table 4.The time series of Q L collocated with observations over various locations in the tropical Pacific during 2006 to 2015 is shown in Fig. 10.Q L shows no large seasonal variability; rather, intraseasonal variability is embedded in the interannual variability at all locations.Both versions of satellite data could capture these variations at various locations in the tropical Pacific (Fig. 10).Q L over 170°W, 165°E, and 140°W shows the impact of the strong negative phase of ENSO during 2010 to

Spatial variability of downwelling radiation
The spatial variability of both versions of CM is studied during 2000 to 2017 to compare and study the enhancements in CMv4 with CMv3. Figure 11 shows the seasonal mean Q S from CMv3 (A to D) and CMv4 (E to H).While both CMv4 and CMv3 show similar seasonal patterns, CMv4 shows lower mean values than CMv3 during all seasons.High Q S in each hemisphere have a strong coupling with a shift in the thermal equator.The northward shift in the ITCZ in spring causes Q S to show large values in the Northern Hemisphere during MAM and JJA.While the ITCZ is in the south, the Southern Hemisphere receives large Q S .The southcentral Indian Ocean, central Pacific, southeastern Pacific, and southwestern Atlantic receive high insolation stretching along 0-30°N during austral summer.Similarly, Q S shows high values in the Northern Hemisphere along 10-25°N during spring and summer.On the other hand, Q S in different seasons shows large spatial variability (standard deviation) over regions with low mean values (Fig. S5).Thus, the equatorial Indian Ocean, Bay of Bengal, southeastern, and equatorial Pacific show large standard deviations (~60 W m −2 ) during spring, summer, and autumn.Contrary to Q S , Q L in CMv3 and CMv4 shows large values in the tropical oceans between 10°S and 10°N irrespective of the season (Fig. 12).These large values in the equatorial belt correspond to a large cloud fraction due to high convection caused by intense insolation [72].Mainly, Q L is high over the Indo-Pacific warm pool region during all seasons.Q L from both CMv3 and CMv4 in the northern Indian Ocean shows large values in summer due to reflected longwave radiation from the high cloud fraction during ISMR (JJA season) [73].However, improvements in CMv4 lead to higher Q L values (>420 W m −2 ) in the tropical oceans than those in CMv3.Thus, the difference in Q L between both versions is approximately 20 W m −2 in the tropical oceans.This bias in 2 versions highlights the need to perform necessary corrections of the same order in the hybrid or developed reanalysis products, forced OGCMs, and climate models (such as CMIP5 and CMIP6) evaluated using CMv3 data.However, there is no notable change in the spatial variability of Q L in tropical oceans in either CMv3 or CMv4, which is approximately 5 W m −2 (Fig. S6).Yet, a narrow stretch of the tropical eastern Pacific, which is more prone to ENSO evolution and its annual variability, shows a standard deviation of 15 W m −2 in both versions during all seasons.
Similarly, the climatologies of CMv3 and CMv4 show a significant spatial difference between Q S and Q L , as shown in Fig. 13.The Q S values in CMv4 are lower than those in CMv3 over the central-eastern Pacific, Arabian Sea, and Indonesian Throw Flow regions.As these regions are vital in controlling tropical weather patterns such as the evolution of ENSO, the advancement of the summer monsoon, and the development of IOD, the corrections in Q S could enhance the study of these atmosphericocean phenomena in the tropical oceans.Similarly, Q L in CMv4 shows slightly higher values than that in CMv3 over the  Indo-Pacific warm pool region and central Pacific.The climatological difference between CMv3 and CMv4 (CMv3-CMv4) for Q S and Q L is shown in Fig. 14.There are sharp gradients in the central Pacific, central Atlantic, and eastern Indian Ocean.These sharp gradients are due to the data product issue in CMv3.The CM products utilize 5 contiguous geostationary satellite imager hourly cloud retrievals to compute the surface fluxes.The boundaries between the satellite imagers are at 100°E, 180°E, 105°W, and 40°W longitudes.SYN1deg-hour, SYN1deg-day, and SYN1deg-month have the same geostationary surface flux artifacts.The use of the SYN1deg dataset at 3-hourly and daily timescales to find the bias between the 2 versions produces sharp gradients in the spatial patterns in Fig. 14.These discontinuities appear regardless of whether hourly, daily, or monthly SYN1deg data are used in the different plots of CMv4 with CMv3.The differences between the 2 products are not uniform.The CMv4 Q S is higher than CMv3 over the Pacific (except south of the central Pacific) and is approximately 10 W m −2 (Fig. 14).On the other hand, the northwest Atlantic and the north Indian Ocean show higher CMv3 Q S (Fig. 14A) compared to CMv4.Otherwise, CMv4 Q S is higher than CMv3 throughout the global tropics.The spatial annual climatological difference for Q L between both versions shows a mirrored pattern of Q S with high values in the southern parts of the tropical oceans and low values in the north.While the Pacific's difference lies at −10 W m −2 , parts of the western and southern Indian Ocean, south of the central Pacific, off the Chilean and African coasts show the least difference between the 2 products (Fig. 14B).A long-term linear trend analysis is performed to determine whether any secular changes are reflected in the downwelling radiation data in the present global warming scenario.Figure 15 shows the annual trends of both Q S and Q L in the global tropical oceans from CMv3 and CMv4 during 2001 to 2017.Q S from both versions of CM shows a similar annual trend in the tropical oceans, with an annual increase of 1 W m −2 in the Southern Hemisphere and a mean annual decrease of −0.7 W m −2 in the Northern Hemisphere.However, CMv3 shows a larger negative trend (Fig. 15A) in the northern Pacific and Atlantic and a positive trend in the southern Pacific and Atlantic than the Q S trend in CMv4.On the other hand, parts of the eastern Pacific in both the Southern Hemisphere and the Northern Hemisphere, including the southern Indian Ocean, show a significant increase in Q S in the CMv4 (Fig. 15B).
Similarly, the Q S trend in the central tropical Pacific shows a lower magnitude in CMv4 than in CMv3.Improved Q S in the CMv4 over the central Pacific and Indo-Pacific warm pool region may help understand air-sea interactions, walker circulation changes, and deep convection in tropical oceans.A major part of these positive and negative trends in Q S in both CMv3 and CMv4 are significant at 95%, except in the central Atlantic and the north Indian Ocean in both versions.On the other hand, Q L shows no such significant annual trends (Fig. 16A and B) except over a patch north of South America, which shows a positive trend in CMv3.This highlights the stronger annual changes in Q S than in Q L in the global tropical oceans.Q L over the southeastern Atlantic shows an annual trend at 90% significance (figure not shown).However, CMv4 shows no such prominent increasing trend of Q L .Parts of the central and northeastern Pacific, southern Indian Ocean, and southern and northern Atlantic show a positive trend of mean Q L at 0.3 W m −2 (Fig. 16B).The annual trends computed using selected GTMBA stations for Q S and Q L over 3 tropical oceans are in coherence with the annual trends in the satellite products in the study period (Table 5).Although these stations are based on data availability, sparseness/discontinuous observations may lead to discrepancies in observed trends compared to trends in satellite data.Thus, improved and accurate downwelling radiation data over these regions as input to forcing the GCMs could reduce errors in the SST and mixed-layer depths and enhance the understanding of phenomena such as the evolution of ENSO and IOD and their teleconnections in the tropical oceans.

Annual change in SST associated with the annual trend in Q S and Q L in CMv4
Both radiative and turbulent heat fluxes are vital to understanding the air-sea interaction, global circulation patterns, and heat and momentum exchange between the atmosphere and oceans.Notably, a large bias in radiation fluxes could induce bias in the SST, upper ocean heat content, and variability in the global oceans [15,16].These biases in the upper ocean could further lead to discrepancies in weather forecasting, studying climate variability, and understanding air-sea interactions [74][75][76].Similarly, accurate radiative fluxes in the analysis and numerical models could potentially lead to a better understanding of air-sea coupled processes such as ENSO, IOD, and Madden-Julian oscillation in the tropical oceans and convection over time over the warm ocean surfaces and their teleconnections.This could further greatly enhance weather forecasting and climate prediction [76].
SST anomalies play a key role in the upper ocean, lower atmospheric variability, and predictability [75][76][77].SST anomalies can also induce anomalous convection through surface evaporation and low-level moisture convergence [78].The anomalous atmospheric convection through cloud radiation, wind evaporation, wind-induced oceanic mixing, and upwelling causes the SST to change.These atmospheric feedbacks can be detected in the SST tendency.The processes responsible for SST tendency differ region to region over tropical oceans.
errors in and heat flux over the tropical ocean are responsible for the distortion of seasonal SST change in the coupled models [79].Thus, it is essential to evaluate the role of different fluxes on SST and mixed layers.This study estimated the changes in mixed-layer temperature due to the annual trend in Q S and Q L from CMv4.The mixed-layer temperature or SST tendency equation is given as where T m denotes the mixed-layer temperature.K z is the coefficient of vertical diffusion of heat (0.1 × 10 −4 m 2 s −1 ).Res is the residual term.W e is the entrainment rate (m s −1 ), T b is the temperature at the bottom of the mixed layer, ∂T m /∂t is the rate of change in T m , ρ is the seawater density (1,024 kg m −3 ), C p is the heat capacity of seawater (3,993 J kg −1 °C−1 ), and H m is the mixed layer depth.Q 0 is the net surface heat flux (W m −2 ) (1)    17a).On the other hand, although Q L annual positive trends show an increase in annual SST trends in the tropical oceans (Fig. 17b), these low magnitude values (−0.1 to 0.1 °C) could be less significant than low significant Q L trends in Fig. 16A and B.

Conclusions and Summary
This study evaluated satellite-derived downwelling radiation data from 2 versions of CM in the global tropical oceans using in situ observations from the GTMBA program.In situ data from GTMBA were used in this study during 2000 to 2015 to evaluate the CM performance.GTMBA has more data points in the Pacific, followed by the Atlantic and Indian Oceans (Table 1).Both Q S and Q L from CMv4 show improvements compared to CMv3, but the performance of CMv4 is not homogeneous in tropical oceans.While the improvement in Q S is robust in all 3 basins, Q L from CMv4 in the tropical Indian Ocean and tropical Pacific has slightly better performance than CMv3 (Tables 2, 3, and 4).On the other hand, both the correlation and RMSE of Q L from CMv4 show a significant improvement in the Pacific compared to CMv3.However, Q S in the CMv4 has been significantly enhanced over major parts of the tropical oceans.Improved satellite sensors, data retrieval techniques, radiative transfer algorithms, and data processing techniques could have potentially contributed to the enhancements in the Q S data in CMv4 over the tropical oceans.The magnitude and spatial patterns of both Q L and Q S from CMv3 and CMv4 are strongly coupled with the ITCZ and the movement of the thermal equator during all seasons.Although both Q S and Q L show similar seasonal mean and variability (standard deviation) patterns, the seasonal mean and standard deviation of Q S from CMv3 are higher than those from CMv4, while those of Q L from CMv3 are lower than those from CMv4 (Figs. 11  and 12).However, persistent convection and a large cloud fraction over major tropical oceans cause Q L to be high during all seasons.The annual climatological difference between the 2 products (CMv3-CMv4) shows larger Q S values in the Pacific than in other parts of the global tropical oceans (Fig. 14A).
On the other hand, the annual climatological difference between the 2 versions of CM downwelling data (Q L ) shows large values of Q L from CMv4 in the Indian Ocean (Fig. 14B).Similarly, the annual trends of Q S and Q L show large variability in both versions over tropical oceans (Figs. 15 and 16).While the Northern Hemisphere shows a negative Q S annual trend of −0.7 W m −2 , the Southern Hemisphere shows an annual trend of 1 W m −2 .The large positive annual trend (0.5 W m −2 ) in Q L from CMv3 has been further lowered by the CMv4 annual trend over many parts of the tropical oceans.While annual trends in Q S are significant at 95%, Q L shows no such significant trends in both satellite products.The positive annual trends in Q S may lead to the large annual change in SSTs in the tropical oceans.While positive annual trends in Q S cause southern hemispheric SSTs to increase, negative annual trends show no significant impact on northern hemispheric SSTs (Fig. 17).
Similarly, although Q L annual trends cause SSTs to change in the tropics, these changes are lower than those caused by Q S and are insignificant.Thus, CMv4 shows significant enhancements in both Q L and Q S over the regions of the Indo-Pacific warm pool, eastern tropical Pacific, and central Atlantic.However, persisting errors, bias, and low correlation values over regions such as the tropical Indian Ocean need a relook and demand further study.Similarly, low correlation values over the Indian Ocean highlight the complex nature of cloud patterns and air-sea interactions over the region.As clouds, precipitation, aerosols from forest fires, volcanic eruptions, pollution, and desert storms could affect the downwelling radiation, necessary corrections in the data retrieval techniques from spaceborne instruments could potentially increase the accuracy of the data from space.The impact of coupled climate phenomena such as ENSO, Madden-Julian oscillation, and IOD in tropical oceans plays a vital role in modulating the cloud fraction and precipitation daily to seasonal scales [84].In future work, we will investigate the impact of these events on the variability of downwelling fluxes in the Indian Ocean to better understand the low correlations, large bias, and RMSE with in situ observations over this region.The satellite-derived products underestimate the observed daily variability in terms of standard deviation in all 3 ocean basins.Although CMv4 has been improved significantly, it underestimates the observed values.Hence, these findings may help the data product developers to improve it in future releases.Thus, the datasets produced, numerical models run, and climate model outputs validated using CMv3 need corrections to make their results more accurate.As CMv4 shows an apparent enhancement over CMv3, CMv4 could be a better alternative to reanalysis or model-simulated products for climate model evaluation and for use as a forcing field to run standalone ocean models.

Fig. 2 .Fig. 3 .
Fig. 2. Frequency distribution of (A) Q S and (B) Q L in the Indian Ocean from RAMA and CM using daily data during 2000 to 2015.

Fig. 4 .
Fig. 4. Temporal variability in Q L in CMv3 (red) and CMv4 (blue) compared with RAMA (gray) in the Indian Ocean during 2000 to 2015.Data plotted using 7-day running mean (smoothing).Discontinuous gray lines are due to the unavailability of observations during those time steps.(A) and (B) are the locations of the RAMA buoy stations in the Indian Ocean where the location-based analysis was performed with the CMv3 and CMv4.

Fig. 5 .
Fig. 5. Frequency distribution of (A) Q S and (B) Q L in the tropical Atlantic Ocean from PIRATA and CM using daily data during 2000 to 2015.

Fig. 6 .
Fig. 6.Temporal variability of Q S from CMv3 (red) and CMv4 (blue) compared with PIRATA (gray) in the tropical Atlantic Ocean during 2000 to 2015.Data plotted using 7-day running mean (smoothing).Discontinuous gray lines are due to the unavailability of observations during those time steps.(A) to (D) are the locations of the PIRATA buoy stations in the Atlantic Ocean where the location-based analysis was performed with the CMv3 and CMv4.

Fig. 7 .
Fig. 7. Temporal variability in Q L from CMv3 (red) and CMv4 (blue) compared with PIRATA (gray) in the tropical Atlantic Ocean during 2000 to 2015.Data plotted using 7-day running mean (smoothing).Discontinuous gray lines are due to the unavailability of observations during those time steps.(A) to (C) are the locations of the PIRATA buoy stations in the Atlantic Ocean where the location-based analysis was performed with the CMv3 and CMv4.

Fig. 8 .Table 4 .
Fig. 8. Frequency distribution of (A) Q S and (B) Q L in the tropical Pacific Ocean from TAO/TRITON and CM using daily data during 2000 to 2015.

Fig. 9 .
Fig. 9. Temporal variability of Q S (W m −2 ) from CMv3 (red) and CMv4 (blue) compared with TAO/TRITON (gray) in the tropical Pacific Ocean during 2000 to 2015.Data plotted using 7-day running mean (smoothing).Discontinuous gray lines are due to the unavailability of observations during those time steps.(A) to (D) are the locations of the TRITON buoy stations in the Pacific Ocean where the location-based analysis was performed with the CMv3 and CMv4.Downloaded from https://spj.science.orgon February 01, 2023

Fig. 10 .
Fig. 10.Temporal variability in Q L from CMv3 (red) and CMv4 (blue) compared with TAO/TRITON (gray) in the tropical Pacific Ocean during 2000 to 2015.Data plotted using 7-day running mean (smoothing).Discontinuous gray lines are due to the unavailability of observations during those time steps.(A) to (D) are the locations of the TRITON buoy stations in the Pacific Ocean where the location-based analysis was performed with the CMv3 and CMv4.

Fig. 11 .Fig. 12 .Fig. 13 .BFig. 14 .BFig. 15 .BFig. 16 .
Fig. 11.Seasonal mean Q S from CMv3 and CMv4 in the global tropical oceans from March 2000 to February 2017.(A to D) Seasonal mean of QS from CMv3.(E to H) Seasonal mean of QS from CMv4.Downloaded from https://spj.science.orgon February 01, 2023 profile data during 1993 to 2017, i.e., using 25 years of data[62].A reference depth of 10 m, including temperature gradient criteria ∆T = 0.8 °C, was used to compute the MLD from temperature profiles[83].The climatology of the MLD computed is shown in Fig. S7.We use the same duration, i.e., 2000 to 2017, to prepare the climatology.We observe no difference in the SST climatology made for a longer period (1993 to 2017).The Pacific and Atlantic show large MLDs ranging from 70 to 100 m and are along the path of mean wind directions.The Indian Ocean shows a lower MLD with a maximum of 60 m.The maximum annual change in SST caused by the Q S and Q L annual trend in the global oceans is shown in Fig. 17.A positive Q S trend in CMv4 causes southern hemispheric SSTs, particularly over the southern Indian Ocean and southeastern Pacific, to increase between 0.1 and 0.3 °C annually.Nevertheless, negative trends in Q S over the Northern Hemisphere show no such large negative trends in SSTs (Fig.

Fig. 17 .
Fig. 17.Annual change in SST due to annual trends in (A) Q S and (B) Q L from CMv4 in the global tropical oceans during 2000 to 2017.

Table 1 .
Details of GTMBA data used to evaluate CM data during 2000 to 2015.
Downloaded from https://spj.science.orgon February 01, 2023 data of Q S and Q L .Data are obtained from 7 CERES instruments (https://ceres.larc.nasa.gov/instruments/) on the 5 launched satellites (Tropical Rainfall Measuring Mission, Terra, Aqua, Suomi National Polar-orbiting Partnership, and NOAA-20).The Langley Fu-Liou radiative transfer model produces the top of the atmosphere fluxes.Although CM Synoptic (SYN1deg) products (CMv3 and CMv4) provide hourly data (and also have a temporal resolution of 3 h, daily, monthly hourly, and monthly), we use climate-quality global daily (temporarily averaged), 1°×1° (spatially) gridded surface radiant fluxes (Q S and Q L from CMv3 and CMv4) available during 2000 to 2017

Table 2 .
Statistics of Q S and Q L from CMv3 and CMv4 compared with RAMA in the Indian Ocean using daily data during 2000 to 2015.

Table 3 .
Statistics of Q S and Q L from CMv3 and CMv4 compared with PIRATA in the Atlantic Ocean using daily data during 2000-2015.

Table 5 .
Annual trend in the GTMBA data over selected locations in the global tropical oceans.Empty cells correspond to unavailability of sufficient no. of observations.