The Japanese Ocean Flux Data Sets with use of Remote Sensing Observations (J-OFURO) latent heat flux field is compared with the Hamburg Ocean–Atmosphere Parameters and Fluxes from Satellite Data (HOAPS), the Goddard Satellite-Based Surface Turbulent Fluxes (GSSTF), ECMWF, NCEP–NCAR reanalysis (NCEP1), and da Silva et al.'s fields. All products qualitatively reveal a similar pattern in the average fields. Although the latent heat fluxes of J-OFURO and GSSTF are quite similar, they are larger than those of HOAPS in the tropical regions. The difference between J-OFURO and the da Silva data is large, and the temporal correlation is extremely low in the Southern Hemisphere. This suggests that the da Silva product hardly reproduces accurate variability in the data-sparse regions. Also the time correlation between J-OFURO and ECMWF or NCEP1 is considerably lower in the Southern Hemisphere than in the Northern Hemisphere. The ECMWF and NCEP1 fields may be affected by the lack of ship observations there. The present study also compares meridional profiles of the zonal average. The HOAPS and da Silva products significantly underestimate these profiles in the tropical regions compared with the other products. On the other hand, the ECMWF product overestimates these profiles in the equatorial regions.
There is a deficiency in the understanding of ocean and cloud processes in the global climate system. Heat transfer at the sea surface plays an important role in linking the ocean and the atmosphere, and consequently, to the generation of clouds. Therefore, monitoring of the heat transfer between the ocean and the atmosphere is crucial for understanding a global climate system. The heat transfer has four components, that is, shortwave radiation, longwave radiation, latent heat flux, and sensible heat flux. Shortwave radiation transfers heat from the atmosphere to the ocean, while the other three components mainly transfer heat from the ocean to the atmosphere. The magnitude of the heat flux strongly depends on time and location. Generally, shortwave radiation and latent heat flux are principal components of the heat transfer. Although the shortwave radiation is larger than latent heat flux, the latent heat flux is more important for the global climate problem because of two inherent characteristics. One is the large amplitude of interannual and spatial variability. It is in striking contrast to shortwave radiation. Since shortwave radiation basically depends on insolation, the distribution is zonal and the time variation is annual. The latent heat included in water vapor can be freely moved from one place to another. This characteristic is closely related to the redistribution of heat energy in the global climate system and is one of the essential factors for understanding a global climate system. Therefore, it is important to observe latent heat flux at the sea surface.
Recently several datasets of latent heat flux were constructed from satellite data. These data are freely available for use by scientific groups. They are the Hamburg Ocean–Atmosphere Parameters and Fluxes from Satellite Data (HOAPS), the Goddard Satellite-Based Surface Turbulent Fluxes (GSSTF), and the Japanese Ocean Flux Data Sets with Use of Remote Sensing Observations (J-OFURO). Though all of these datasets are constructed mainly by using the Defense Meteorological Satellite Program's (DMSP's) Special Sensor Microwave Imager (SSM/I), they each use a different algorithm. For example, GSSTF uses an algorithm by Chou et al. (1997), while HOAPS and J-OFURO use an algorithm by Schlüssel et al. (1995) for estimation of specific humidity. Therefore, it is expected that characteristics of the datasets are different from each other. On the other hand, numerical weather prediction (NWP) products can also provide latent heat flux estimates. For example, NWP products by the European Centre for Medium-Range Weather Forecasts (ECMWF) and the National Centers for Environmental Prediction (NCEP) are often used for various studies. Although a NWP model can reproduce a realistic state through data assimilation, it is impossible to maintain the continuity of the characteristics over longer periods of time because the models are continuously changed and improved. Therefore, many recent studies use reanalysis data such as the ECMWF Re-Analysis (ERA) data and the NCEP–National Center for Atmospheric Research (NCAR) reanalysis data instead of NWP products.
Though we can evaluate instantaneous latent heat flux data by using buoy data as sea truth data (e.g., Kano and Kubota 2000), it is difficult to evaluate time-averaged grid data because there is no sea truth for these data, since they are estimated indirectly from observations. Therefore, it is important to compare each dataset to the others, clarify the differences between them, and indicate the characteristics of each. Heat flux data are often used as input data for atmospheric and oceanic general circulation models, and therefore, understanding the characteristics of the input datasets is critical.
In the present study, intercomparison of various datasets of monthly latent heat flux is carried out. The compared datasets are HOAPS, GSSTF, J-OFURO, ECMWF, NCEP1 (NCEP–NCAR), and da Silva et al.'s (1994) data. The last dataset is based on in situ data such as ship and buoy data. A major disadvantage of the ship observations is inadequate coverage that is strongly dependent on shipping lanes. The intercomparison is carried out for the period from 1992 to 1994 for all the datasets except for the da Silva data. The intercomparison period is from 1992 to 1993 for the da Silva dataset. Section 2 briefly describes each dataset. Section 3 presents comparison results for all the datasets. The influence of wind and specific humidity on the difference of latent heat flux is discussed in section 4. Finally, a summary is given in section 5.
In the present study we compare the J-OFURO latent heat flux data (Kubota et al. 2002) with the other five datasets. The datasets compared with J-OFURO are the HOAPS, GSSTF, ECMWF analysis, NCEP–NCAR reanalysis (NCEP1), and the da Silva datasets. HOAPS, GSSTF, and J-OFURO are all satellite-derived datasets. We use an older version of GSSTF (version 1; GSSTF1) in the present study, though a new GSSTF product (version 2; GSSTF2) has become available recently. In order to estimate latent heat flux using a bulk formula, one needs three physical parameters, that is, wind speed, sea surface temperature (SST), and specific humidity. Although the original data for determining these physical parameters are obtained from the SSM/I on board the satellites of the DMSP and the Advanced Very High Resolution Radiometer (AVHRR) on board the satellites of the National Oceanic and Atmospheric Administration (NOAA) in every dataset, some of the algorithms for obtaining physical parameters and calculating latent heat flux are different depending on each dataset. HOAPS uses the NOAA–National Aeronautics and Space Administration (NASA) Oceans Pathfinder SST product (Graßl et al. 2000). J-OFURO and GSSTF use the NCEP analysis for SST data (Reynolds and Smith 1994, 1995). For estimation of wind speed the retrieval of Schlüssel and Luthardt (1991) with improvements described by Schlüssel (1995) is used in HOAPS. On the other hand, GSSTF and J-OFURO use wind speed in an SSM/I geophysical product provided by Wentz (1994). The algorithm estimating surface air humidity from SSM/I data is based on Schlüssel et al. (1995) in HOAPS and J-OFURO, while that of GSSTF use the method of Schultz et al. (1993) and Chou et al. (1995, 1997). Another cause for differences in latent heat flux estimates between different products is a difference in the bulk formulation used. All global latent heat flux data are estimated using a bulk formula. Therefore, the choice of the bulk formula will largely affect the resulting heat flux estimate. HOAPS, GSSTF, and J-OFURO use a bulk formula based on Smith (1988), Chou (1993), and Kondo (1975), respectively. Finally the spatial resolution is different depending on the dataset. The spatial resolution of HOAPS and J-OFURO is 1° × 1°, while that of GSSTF is 2° × 2.5°. The above-mentioned differences in the estimation methods for the different satellite-derived product are summarized in Table 1.
ECMWF and NCEP1 data are compared with J-OFURO data in the present study. ECMWF has created and maintains archives of level III-A data and of level III-B data (Re-Analysis). The level III-A archive is subdivided into four classes of datasets. In the present study we use the Extension Dataset in the four datasets. The Extension Dataset contains additional surface data, fluxes, net radiation data, and precipitation derived from 24-h forecast values. All the fields in this dataset contain values accumulated between time step 12 and time step 36 of the forecast. The spatial resolution is 1.125° × 1.125°. On the other hand, the NCEP1 dataset is a reanalysis dataset produced by NCEP–NCAR. Perceived climate jumps associated with changes in the operational data assimilation system are eliminated in this product. The spatial resolution is 2.5° × 2.5°. The reanalysis data assimilation system is described in Kalnay et al. (1996) in more detail. The da Silva et al. (1994) product is based on the Comprehensive Ocean–Atmosphere Data Set (COADS), which is constructed by collecting global weather observations taken near the ocean's surface, primarily from ships and buoys (Woodruff et al. 1987). The COADS provides two kinds of datasets, which are monthly mean summaries in 2° × 2° boxes over the global oceans and raw individual observations. Da Silva et al. (1994) constructed the monthly mean fields objectively analyzed with essentially the same successive correction scheme used by Levitus (1982). The analyzed fields are given with 1° grid spacing. Also it should be noted that the da Silva dataset is available only for the years before 1993.
All data are transformed into monthly data with 1° grid spacing for intercomparison in the present study. The intercomparison period is from 1992–94, except da Silva et al. (1992–93).
Figure 1 shows the mean latent heat flux as estimated by J-OFURO (1992–94) and the mean difference between J-OFURO and other products. The map shows that the largest values occur over the subtropical oceans around 20°, called the oceanic deserts. However, even in midlatitudes we can find large values along the western boundary of the Pacific and Atlantic Oceans. These are caused by the effects of dry and cold monsoon and warm currents such as the Kuroshio and Gulf Stream in winter (Masuzawa 1952). On the other hand, latent heat flux is small over the equatorial oceans due to weak winds in the western part and relatively low SST caused by the effect of equatorial upwelling in the eastern part. Basically latent heat flux in the high latitudes is considerably lower, less than 10 W m−2. Though both HOAPS and J-OFURO are satellite-derived products, the latent heat flux of J-OFURO is larger than that of HOAPS in the subtropics. On the other hand, it is interesting that the average difference between J-OFURO and GSSTF is extremely small. J-OFURO underestimates in the equatorial regions and in the central part of the North Pacific compared with the ECMWF product. J-OFURO overestimates in the eastern part of the subtropics in the South Pacific and in the central part of the subtropics in the North Pacific compared with not only ECMWF but also NCEP1 and da Silva. The overall feature of the mean difference field between J-OFURO and ECMWF, NCEP1, and da Silva products is fairy common. The mean difference field between J-OFURO and da Silva shows a feature with a small spatial scale, closely related to a sampling error. We checked whether the difference is significant or not by a two-tailed F distribution. The shaded regions in Fig. 2 indicate that the difference is significant within a 99% limit. It shows that J-OFURO is significantly different from other products in the above-mentioned regions and in the high-latitudes in the Southern Hemisphere except HOAPS.
Figure 3 shows root-mean-square (rms) difference fields after removing the average difference. It is noted that the rms difference is small compared with the average difference. In particular, the rms difference between J-OFURO and GSSTF is extremely small, less than 10 W m−2 in most places. The rms difference between J-OFURO and other products except ECMWF is large, more than 50 W m−2 in the tropical regions except around the equator. In particular, the difference between J-OFURO and the da Silva product is considerably large even in mid- and high latitudes, though that is smaller than the average difference between them. On the other hand, the rms difference between J-OFURO and ECMWF is large in the western equatorial Pacific and over the western boundary currents such as the Kuroshio and Gulf Stream.
Temporal cross-correlation coefficients (CCC) are calculated at each grid point between the fields. The map of the cross-correlation coefficient is given in Fig. 4. The CCC is extremely high, more than 0.96, between J-OFURO and GSSTF. On the other hand, the CCC between J-OFURO and HOAPS is high in the Northern Hemisphere and low in the Southern Hemisphere, though both of them are satellite-derived products. The asymmetry of the CCC between the Northern and Southern Hemispheres is common for ECMWF and NCEP1. The CCC between J-OFURO and da Silva is considerably smaller, less than 0.5 in most places, compared with other cases. The high CCC regions between J-OFURO and da Silva seem to correspond with regions where the observations are most abundant. This result suggests that the large variation in CCC in Fig. 4 is due to the lack of ship observations in the Southern Hemisphere and the da Silva product, which hardly reproduces time variability in the data-sparse regions. Also the low CCC regions even between J-OFURO and NCEP1 or ECMWF are found over data-sparse regions with low variability such as equatorial regions and the high latitudes in the Southern Hemisphere. This suggests that the effectiveness of NCEP1 and ECMWF may be limited to the Northern Hemisphere and the subtropics in the Southern Hemisphere.
Zonal average profiles of average values are shown in Fig. 5. The overall variation pattern is common for all products. There is a maximum around 15° in each hemisphere. Though the average value of the maximum in each hemisphere is similar, the maximum in winter is larger than the maximum in summer. It is remarkable that the HOAPS product is extremely small in the low latitudes compared with other products. On the other hand, the GSSTF product is larger than other products north of 40°N, especially in boreal winter. The da Silva product remarkably underestimates in the subtropics compared with other data except HOAPS. Also the ECMWF product is unexpectedly larger than other products in the equatorial region. Though NCEP1 shows a similar value to J-OFURO, the meridional contrast in the low latitudes is more remarkable in the J-OFURO product. The overall variation pattern is common for three satellite-derived products. However, there exist quantitative differences depending on space and season. HOAPS underestimates within 30°N in both seasons, while GSSTF overestimates north of 40°N compared with other satellite-derived products. On the other hand, the variation pattern given by J-OFURO is a little bit different from that seen in ECMWF, NCEP1, and da Silva. For example, the latter three products are shown to be considerably flatter, from 10° to 40°N in boreal winter, while J-OFURO and GSSTF have a remarkable maximum near 10°N. It is interesting that HOAPS, even though it is a satellite-derived product, also shows the flat feature. All products give the minimum on the equator. However, the value of ECMWF is larger and that of HOAPS is smaller than the values of other products. The location of the ECMWF maximum in the Southern Hemisphere during the boreal winter is near 5°S, while maxima in other products are between 15° and 20°S.
4. Influence of wind and specific humidity
As mentioned before, there is no sea truth for the monthly gridded latent heat flux because they are not directly observed. However, it is important to compare them and to investigate the cause of the differences between them in order to construct a more accurate product. There are many factors that can cause the differences in latent heat flux estimates between the different products. Table 1 shows that the algorithm used to derive surface physical parameters from the satellite data is different for each product. Thus, we examine the impact of the surface parameters on the difference of latent heat flux fields in the present study. The method is quite simple. We evaluate the difference between J-OFURO and the results after we calculate the latent heat flux again by changing a surface physical parameter from J-OFURO to HOAPS or GSSTF. However, the method calculating specific humidity and wind speed is similar between J-OFURO and HOAPS, and J-OFURO and GSSTF, respectively. Therefore, we investigate the impacts of other parameters, that is, wind speed for HOAPS and specific humidity for GSSTF. The results are shown in Fig. 6. Figure 6 shows meridional profiles of the zonal mean difference between the mean values of J-OFURO and other satellite-derived products. The results of J-OFURO calculated using HOAPS wind data or GSSTF specific humidity instead of the original J-OFURO data are shown in Fig. 6. If we use HOAPS wind data, the difference decreases remarkably in the tropical region. However, the difference slightly increases in high latitudes. On the other hand, the impact of GSSTF specific humidity is not so large. Although the difference decreases as a whole when using GSSTF specific humidity, the value of the decrease varies depending on the latitude. Also there exist several regions where the difference increases. This suggests the existence of other causes for the difference.
Results from the comparison of the latent heat flux from J-OFURO with HOAPS, GSSTF, ECMWF, NCEP1, and da Silva et al. (1994) have been presented. Time and space resolutions for data used are 1 month and 1° × 1°, respectively. The comparison period is from 1992 to 1994 for all products, except for the da Silva product for which the period is 1992–93. The HOAPS and da Silva products are found to underestimate the latent heat flux in the tropical regions compared with the other products. On the other hand, the GSSTF product is found to overestimate the latent heat flux in high latitudes. The J-OFURO product generally gives large values of latent heat flux in the subtropics compared with other products except GSSTF. The features related to the da Silva product are considerably different from other products. For example, the mean difference field between J-OFURO and the da Silva products has small-scale structure. The CCC between the products is extremely low in the Southern Hemisphere compared with the Northern Hemisphere where ocean observations are most abundant. These results suggest that the usefulness of the da Silva product in reproducing true variability may be limited to the Northern Hemisphere midlatitudes and a few other regions. It should be noted that the differences in correlations between data-rich and data-sparse regions are evident in correlations of ECMWF and NCEP1 with J-OFURO. This suggests that accuracy of the latent heat flux from NWP products such as ECMWF and NCEP1 strongly depends on the density of assimilated data. Some of the comparisons may be affected by the difference in periods between da Silva and other products. However, because we compare da Silva data with J-OFURO for the same period, the effects may not be so large compared with other factors.
There are many factors that cause the differences in latent heat flux estimates between the products. We compared the J-OFURO product with that using the HOAPS wind data and the GSSTF specific humidity data. The impact of the HOAPS wind data is found to be considerably large in the tropical regions, while the impact of the GSSTF specific humidity is found to be small except at high latitudes. In the present study different products were gridded and temporally averaged in time and space to a unified resolution of 1 month and 1° × 1°. However, the original time and space resolutions are different for each product. Therefore, the difference in latent heat flux estimates may be affected by the interpolation of data. For example, there is a large difference between GSSTF and other products in high latitudes. The cause may be the large grid size in GSSTF and the smallness of the length of a parallel in high latitudes. Recently version 2 of GSSTF was produced by the surface turbulent fluxes research group at the Goddard Space Flight Center. Since the spatial resolution of this product is 1° × 1°, it will enable the investigation of the cause of the large difference in high latitudes. Also it should be pointed out that we cannot neglect the impact of the difference of the bulk formula. This issue is being investigated in the Ocean Surface Turbulent Flux Project at present (see online at http://paos.colorado.edu/∼curryja/ocean/bulk flux.html).
All satellite-derived products use SSM/I wind data for estimation of latent heat flux. The choice of SSM/I wind data has an advantage of simultaneous observation for specific humidity. However, a microwave scatterometer such as NASA's QuikSCAT provides us with accurate wind data at the sea surface. Therefore, we should investigate the impact of the choice of wind data on latent heat flux in the future.
Finally we cannot know which product is closer to truth since we only carried out intercomparison of the products. In order to do so, it may be necessary to compare each product with sea truth data, for example, buoy data. Our future work will focus on this subject.
We thank Gad Levy and two anonymous reviewers for their valuable comments. This research was partly supported by the Grant-Aid for Science Research on Priority Areas (No. 08241111) of the Ministry of Education, Science, Sports, and Culture, Japan.
Corresponding author address: Dr. Masahisa Kubota, School of Marine Science and Technology, Tokai University, 3-20-1 Orido, Shimizu, Shizuoka 424-8610, Japan. Email: email@example.com