1. Introduction
The retrieval of total columnar water vapor path [i.e., precipitable water (PW)] over water surfaces from passive microwave satellite measurements is generally thought to be a tractable problem, largely because of the presence of the weak 22.235-GHz water vapor line in the low centimeter spectrum, which does not generally saturate for the characteristic range of vertically oriented atmospheric water vapor paths. Since the advent of the Special Sensor Microwave Imager (SSM/I) in July 1987, a number of PW retrieval algorithms have been developed for various combinations of passive microwave (PMW) brightness temperatures (TB) at frequencies in the vicinity of the 22.235-GHz line, the actual SSM/I channel-3 frequency. These consist of various statistical and physical methods in which the former use observed PW or related measurements to produce empirically derived coefficients within the algorithm’s PW formulation, while the latter are formulated in terms of substantiated physical principles bereft of empiricism.
Comparisons of several SSM/I algorithms and less precise infrared algorithms used over oceans based on bias–rms statistics and regression analyses suggest that current PW algorithms are in good agreement (e.g., Jackson and Stephens 1995; Wentz 1997). Given this perspective, the emphasis in the current debate appears to be mostly concerned with the relative merits and weaknesses of the statistical versus physical PMW approaches.
In this study we present evidence that a mix of seven current statistical and physical SSM/I algorithms are in considerable disagreement when viewed in a monthly and zonally averaged framework; that is, relative differences of order 50% are not uncommon. Much of the disagreement arises from what might be called tangential environmental factors in relationship to the main contributor (i.e., water vapor path) to the PMW radiation signals used for retrieval. These factors pertain to the nonrepresentative sampling of radiosonde training data used for generating regression coefficients in statistical algorithms, as well as the influence of sea surface temperature (SST), surface wind speed (Us), cloud liquid water path (LWP), and water vapor mixing ratio profile structure [q(z)] on frequency-dependent TB values in both statistical and physical algorithms. The tangential factors lead to errors of different properties, resulting in significant differences in a latitude-dependent intercomparison framework. A diagram from Jackson and Stephens (1995) involving intercomparison of four of the seven algorithms considered here, their Fig. 8b, revealed the nature of this problem but did not underscore its significance.
It is not surprising that differences exist between the PMW algorithms given that there are known significant PW differences between physically founded infrared and 22-GHz passive microwave techniques (e.g., Stephens et al. 1994; Sohn et al. 1998; Scott et al. 1999). (A synopsis of this issue is found in section b of the appendix.) Therefore, it is important in a climate dynamics context that water vapor retrieval remains an ongoing research problem. As discussed below, a number of processes related to how water vapor impacts the climate system should be tested in the framework of how well model representations of the water vapor distribution can be verified by satellite observations (e.g., Chen et al. 1996). Such research is relevant to the scientific goals of the internationally organized Global Energy and Water Cycle Experiment (GEWEX), especially improving the measuring and modeling of water vapor concentrations and transports at climate scales [National Research Council (NRC) 1999].
Section 2 discusses the key issues motivating the need for highly accurate satellite retrieval of water vapor and precipitable water, while section 3 describes the methodology and datasets and section 4 presents results and interpretation. Section 5 offers final discussion and conclusions, including recommendations on how current algorithm discrepancies could be mitigated in the future. The appendix provides summary descriptions of the seven algorithms.
2. Background and motivation
This intercomparison analysis of SSM/I PW retrieval algorithms embodies two foremost scientific objectives. The first is to draw attention to quantifiable algorithm differences, including their most likely causes, and also to assess the scientific significance of the differences. This is complicated by the fact that water vapor distribution over the globe undergoes significant seasonal and interannual variations (Bates 1991; Gaffen et al. 1991, 1992). The second is to shed light on how the differences can be reduced or eliminated based on algorithm refinements. Of course, one of the endemic problems with the latter objective is that, until an unbiased space-based calibration system is developed (perhaps a stabilized Raman lidar system), there are few calibration-quality measurements available for verification.
Obtaining algorithm near agreement is not necessarily a sign that the problem is well in hand. There is always the possibility that all algorithms are biased because of an intrinsic flaw in either the measurements or the assumed underlying line absorption physics in the millimeter–centimeter electromagnetic (EM) spectrum. Nevertheless, algorithm near agreement might be viewed as a necessary condition, if not sufficient condition, before the retrieval problem can be said to have been solved.
In terms of the application of satellite-based global PW fields for seeking a better understanding of the global hydrological cycle and its transient behavior, latitude-dependent gradient errors are a serious issue. This is because meridional moisture transport is one of two main climatic horizontal transport modes (the other is zonal: across ocean–land boundaries).
Accurately representing meridional gradients of PW is a fundamental concern in evaluating atmospheric redistribution of water vapor (e.g., Jedlovec et al. 2000). An error sensitivity analysis of this process is presented in Fig. 1, illustrating how meridionally distributed PW errors affect water vapor transports directly and indirectly through altering infrared radiative cooling rates. All calculations going into this figure were done using monthly National Centers for Environmental Prediction (NCEP) data, according to the following procedures. 1) Meridional PW distribution in the top panel is based on monthly mean vertically integrated NCEP specific humidity profiles from January 1990, with associated outgoing longwave radiation (OLR) distribution obtained from vertically resolved infrared radiative transfer simulations using January 1990 monthly mean NCEP temperature and specific humidity profiles. 2) Meridional ΔPW distribution in the middle panel is given by anomaly PW distribution between results of two PW retrieval algorithms under study, that is, Wentz (1995) and Greenwald et al. (1993), which is then vertically distributed using power-law relationship by adding anomalies to monthly mean NCEP specific humidity profiles, enabling associated anomaly OLR distribution (ΔOLR) to be obtained by radiative transfer. And finally, 3) moisture transport [(υq)] meridional distribution in the lower panel is obtained by vertically integrating meridional moisture fluxes from 7-yr mean NCEP specific humidity and wind profiles using vertically resolved diagnostic procedure, with associated moisture transport anomaly [Δ(υq)] meridional distribution obtained by first producing modified (υq) distribution with anomaly PW distribution from step 2 (vertically distributed using power-law relationship) in conjunction with 7-yr mean meridional distribution of NCEP specific humidity profile, and then differencing resultant vertically integrated modified (υq) flux from 7-yr mean (υq) flux.
Since such transports and the associated circulations are first-order components of the earth’s general circulation and global climate (e.g., Peixoto and Oort 1983, 1992; Bryan and Oort 1984; Chen and Pfaendtner 1993), processes yet to be accurately represented in the operational global analyses (e.g., see Liu et al. 1992; Sohn 1994; Salathé et al. 1995), it is important to consider how such differences in the satellite algorithms arise and what can be done to mitigate them. In this context, the positive or negative impact of data assimilation of satellite-retrieved water vapor on numerical weather prediction (NWP) models is concomitant with the systematic biases in both meridional and zonal gradients contaminating the retrievals.
There are also ongoing debates concerning the role of water vapor as a greenhouse gas (Houghton et al. 1990, 1995), and the extent to which temperature, SST, radiation, cloudiness, and stability changes induce feedbacks into the water vapor mixing ratio, its vertical distribution, and the associated radiative cooling (e.g., Raval and Ramanathan 1989; Betts 1990; Cess et al. 1990; Lindzen 1990; Kinter and Shukla 1990; Stephens 1990; Sun and Lindzen 1993; Chou 1994; Pierrehumbert 1995). Satellite observations of water vapor, particularly upper-tropospheric and stratospheric water vapor, are central to this debate (e.g., McCormick et al. 1993; Soden and Fu 1995; Stephens et al. 1996; Blankenship et al. 2000). This is because they offer the only consistent means to interpret observationally the role of variable atmospheric hydrology on the large-scale climate system (Stephens and Greenwald 1991; Wu et al. 1993; Schmetz and van de Berg 1994; Zhao 1994; Soden and Bretherton 1996; Sohn 1996; Spencer and Braswell 1997), and thus the only effective means to decipher water vapor’s role on possible warming–cooling trends that are being extracted from the temperature records.
Satellite water vapor observations are also the only consistent and independent source of information concerning whether the simulated feedbacks manifested in NWP and GCM models are actually realistic (see Rind et al. 1991; Soden and Bretherton 1994; Zhang et al. 1994; Chahine 1995). Notably, as is the case in a number of scientific controversies, the signal-to-noise ratio in the observations is passed along to the signal-to-noise ratio in the debates.
One of the difficulties in assessing how well climate models perform against the observations is that PW is not the ideal water vapor variable to use in assessing models. Whereas some have suggested upper-tropospheric humidity as a better choice, the optimal variable is the water vapor profile vector. However, current space-based measuring systems for retrieving water vapor profiles, in both the infrared and the passive microwave, are inadequate for climate studies (see NRC 1999). Although this issue is beyond the scope of this article, and notwithstanding the attempts to use PMW techniques to conduct vertical profiling (e.g., Wagner et al. 1990; Lutz et al. 1991; Blankenship et al. 2000), it is fair to say that the only tropospheric water vapor variable currently retrievable that has sufficient accuracy and stability for decadal-scale climate variability studies is PW. That this is so is apparent from the long-term emphasis on using PWs to estimate through underlying but empirically based correlations, water vapor concentrations within specific vertical layers such as the atmospheric boundary layer (ABL); see Reitan (1963), Smith (1966), Liu and Niler (1984), Liu (1986), Liu et al. (1991), and Schulz et al. (1993).
An important result of our analysis to uncover why near agreement has not yet occurred between various PMW algorithms is that the main uncertainty sources can be readily identified. These sources and uncertainty magnitudes are directly relevant to the ongoing debates.
3. Methodology and datasets
The seven SSM/I algorithms under study have all been exercised regularly since their development and are representative of the spectrum of SSM/I PW algorithms currently in use. This set consists of 1) the Alishouse et al. (1990) statistical algorithm (hereafter ALI); 2) the Greenwald et al. (1993) physical algorithm (GRE); 3) the Lojou et al. (1994) statistical algorithm (LOJ); 4) the Petty (1994) statistical algorithm (PET); 5) the Schlüssel and Emery (1990) statistical algorithm (S&E); 6) the Wentz (1995) SSM/I vapor algorithm intercomparison-validation project (VIP) optimal statistical algorithm (WEN-OS); and 7) the Wentz (1992) physical algorithm (WEN-P). Algorithm descriptions are found in the appendix.
a. Assessing algorithm differences in various frameworks
The analysis begins by examining differences between the seven satellite algorithms’ open ocean PW results in three types of frameworks: 1) monthly averaged globally composited point differences relative to radiosonde-derived PWs presented in scatter diagrams; 2) monthly averaged regional differences presented in gridded global maps; and 3) monthly/zonally averaged differences presented in meridional profile diagrams. This first phase analysis demonstrates that scatter diagrams disguise systematic regional biases partly revealed in gridded difference maps and readily apparent in the zonally averaged framework.
This analysis involves two SSM/I TB datasets. The first, used for the part 1 framework analysis, consists of a large compilation of 237 726 PW–TB matchups based on 46 223 open ocean radiosonde profiles and coincident SSM/I pixels acquired by Wentz (1995, 1997). The source of the radiosonde data was the National Center for Atmospheric Research (NCAR) archive of National Meteorological Center (NMC) upper air reports covering a 4-yr period (1987–90). Matchups used a collocation window of 6 h and 60 km. Wentz used this dataset for seeking an optimal statistical PW algorithm (Wentz 1995), as part of the WetNet-sponsored SSM/I vapor algorithm intercomparison-validation project, and for improving his multiparameter physical ocean algorithm for water vapor, cloud liquid water path, and the surface wind vector (see Wentz 1997). Figure 2 identifies the sounding stations used to develop the VIP dataset. Note that a subset of these stations was used by Alishouse et al. (1990) in their PW algorithm development. The second SSM/I dataset, used in parts 2 and 3 of the framework analyses, consists of individual TB overpass maps for the months of January 1990 and July 1990 obtained from the Global Hydrology Resource Center (online at http://ghrc.msfc.nasa.gov) at the National Aeronautics and Space Administration–University of Alabama in Huntsville (NASA–UAH) Global Hydrology and Climate Center (GHCC).
b. Regenerating statistical algorithms
In the second phase of the analysis, we demonstrate that by adapting to a common radiosonde-based training dataset for the statistical algorithms, part of the systematic meridional bias can be removed. This suggests that an important factor giving rise to systematic spatial structure differences in SSM/I statistical algorithms is the variability in the completeness (robustness) of the training data that had been used to establish the coefficients of the predictor variables used in the various statistical formulations. It is also worth noting here that another source of the disparity relates to the long and troubled history associated with transducers used on conventional radiosondes to measure relative or absolute humidity.
Moreover, an additional important result in the second phase of the analysis is that there are coherent residual differences between the various statistical algorithms that are caused by a basic difference in how the training datasets were prepared for the various algorithm formulations. The two basic approaches involve associating the radiosonde profiles to TB values either through PMW radiative transfer (RTE) modeling, or through direct acquisition of measured-collocated TB. This is an important distinction because simulated TB values cannot completely account for actual SST, Us, cloud LWP, and q(z) structure conditions governing the behavior of measured TB; thus the coefficients of the predictor variables for the two types of algorithms are disjoint vis-à-vis the SST, Us, cloud LWP, and q(z) effects. On the other hand, simulations are not vulnerable to measurement errors or geographic sampling biases associated with radiosondes.
c. Testing retrieval responses to SST, Us, cloud LWP, and q(z) variations
Note that the primary mechanisms by which the SST, Us, cloud LWP, and q(z) structure factors impact top-of-atmosphere (TOA) TB are 1) by variable SST altering boundary emission from seawater through both altered thermometric state and thermal emissivity; 2) by variable Us altering boundary emission through altered sea surface roughness and thus variable sea spray and foam conditions; 3) by variable cloud LWP altering atmospheric transmission through Rayleigh absorptance and scattering interactions; and 4) by variable q(z) structure introducing frequency-dependent differences in the vertical weighting functions at the three relevant SSM/I frequencies, coupled with the control of weighting functions on the sensitivity to high-concentration ABL moisture and on the temperature influence in the H2O mass absorptance coefficients. Notably, one measure of a PW algorithm’s reliability is its “degree of orthogonality” to the tangential factors that influence the input brightness temperatures, an issue that is more effectively addressed through the physical algorithm approach.
It is important to recognize that the imposed SST, Us, cloud LWP, and q(z) structure conditions influence the various SSM/I frequencies in different ways so that the orthogonality problem is synergistic with the set of SSM/I frequencies used in a given algorithm, regardless of the specific form of, for example, a regression formula. We also note that there are other less important tangential environmental factors not considered, such as the vertical temperature structure, which has been addressed in a study by Sun (1993) pertaining to the algorithm of Wentz (1992).
In the third phase of the analysis we first assess the uncertainties in both the statistical and physical retrievals resulting from discrepancies in specifying the SST, Us, and cloud LWP conditions. This is done through a set of RTE modeling experiments in conjunction with the 237 726 sets of five-channel SSM/I TB (19V, 19H, 22V, 37V, 37H) collocated with the VIP profiles. The passive microwave RTE model used for these calculations is described by Smith et al. (1994b). Wind-induced roughening of a Fresnel sea surface, including a sea foam component, is based on the work of Stogryn (1972), Wilheit (1979), Monahan and O’Muircheartaigh (1980), and Lojou et al. (1994), and is discussed in a paper by Yang and Smith (1999). The important component of the RTE model for purposes of this analysis is how water vapor absorption is treated in and around the 22.235-GHz line.
As discussed in Smith et al. (1994b), the RTE model uses the absorption physics developed by Liebe (1985) and Liebe et al. (1992), which treats O2 and H2O absorption (including continuum effects) as a combined process across the millimeter–centimeter EM spectrum. For the cloud LWP absorptance calculations, normalized volume absorption coefficients at 19, 22, and 37 GHz for liquid cloud media of unit cloud liquid water content are set to 0.041, 0.058, and 0.138 km–1 (g m–3)–1, as suggested by Ulaby et al. (1981, p. 313). Note that Prigent et al. (1997) have found that Liebe’s model tends to underestimate absorption on the wing of the 22.235-GHz water vapor line, which represents a part of the remaining incompleteness in spectroscopy of the centimeter–millimeter spectrum, although it does not meaningfully impact this intercomparison study.
This third phase of the analysis concludes by examining a more subtle factor than can give rise to differences among the algorithms, that being associated with vertical structure variations inherent to q(z) profiles. As with the SST, Us, and cloud LWP effects on TOA TB, the differences in q(z) structure have different effects at different PMW frequencies. Thus we also seek to assess the possible magnitude of q(z) structure variations on the retrieval algorithms.
d. Compiling representative T–q profile variability
To create a database containing representative profile structure variations in the marine boundary layer for the third-phase RTE calculations, we have developed four T–q profile datasets. The first is derived from the VIP profiles, categorizing by hemisphere (northern and southern) and season (summer and winter). For this purpose the year is divided into two 3-month periods, that is, June–August (JJA) and December–February (DJF). The VIP profile composites for Northern Hemisphere (NH) summer, NH winter, Southern Hemisphere (SH) summer, and SH winter consist of 1795, 1307, 1273, and 1485 profile samples, respectively, in which only profiles that indicate a cloud LWP of less than 0.5 kg m–2 are included [the SSM/I algorithm of Lojou et al. (1994) is used to estimate the cloud LWPs from the associated SSM/I TB]. The second dataset consists of representative sub-Arctic summer and winter profiles. These climatological profiles are from McClatchey et al. (1972), based on compilations of Valley (1965) and Sissenwine et al. (1968).
The third dataset is also derived from Wentz (1995), first filtered for cloud LWPs below 0.5 kg m–2. This consists of selecting and averaging profiles for four regions (referred to as regions 1–4) within 10° latitude zones (all ocean longitudes) centered at the four special radiosonde sites reported by Albrecht et al. (1995) that were associated with three recent cloud-focused field experiments. The four sites were as follows: 1) 33.43°N, 119.57°W—San Nicolas island 100 km off southern California coast [First International Satellite Cloud Climatology Project Regional Experiment (FIRE)]; 2) 36.99°N, 25.17°W—Santa Maria Island in Azores [Atlantic Stratocumulus Tropical Experiment (ASTEX)]; 3) 28.0°N, 24.0°W—ship Valdivia located 1000 km southwest of Santa Maria Island (experiment ASTEX); and 4) 0°, 140.0°W—ship Moana Wave located on the equator [Tropical Instability Wave Experiment (TIWE)]. The sample sizes for the four regional profile averages were 1271, 1271, 5477, and 1301, respectively.
The fourth dataset merges the four ABL-averaged profiles from the Albrecht et al. (1995) study with the VIP regional zone-averaged profiles developed for the third dataset. This dataset provides more details on marine boundary layer moisture structure for tropical and subtropical regimes. The sample sizes were 65, 120, 58, and 44, respectively.
The Albrecht et al. (1995) study provides T–q profiles within and some 100 mb above the ABL. To create the vertically extended merged profiles, we use the statistical moments obtained from the upper atmospheric portions of the associated VIP profiles, consisting of the four zone-averaged regional profiles and their standard deviations (σ) as determined by the profiles selected for the zone averages. To merge the lower and upper portions of the profiles, the difference DIFztop between an Albrecht et al. (1995) T(z) or q(z) average profile at the ABL top (designated as
A summary of the number of samples and the PWs associated with each of the composites for all four profile datasets is given in Table 1. Figures 3–7 provide illustrations of 12 of the 14 composite T(z) and q(z) profiles associated with the four datasets; diagrams of the sub-Arctic summer and winter climatological profiles can be found in McClatchey et al. (1972). For the first dataset, Fig. 3 shows the four hemisphere/season-averaged T–q profiles along with the associated standard deviation profiles. In addition, Fig. 4 shows a family of five q(z) curves for each of the four hemisphere–season categories with the center curves representing the initial average and the two pairs of perturbation profiles on either side created by constant scaling of the initial q(z) profiles by ±25% and 50%. Note the averaged and perturbation profiles are used for the RTE-based retrieval response experiments on q(z) structure variations. Figures 5 and 6 illustrate the same information as Figs. 3 and 4 for the four zonally averaged profiles of the third dataset, while Fig. 7 does the same for the fourth merged profile dataset.
4. Results of analysis
a. Monthly averaged globally composited point differences
To begin the analysis we view the intercomparisons in the monthly averaged globally composited framework, adopting the familiar scatter diagram to illustrate the degree of similarity. Figure 8 presents the scatter diagrams while Table 2 tabulates the associated summary statistics based on intercomparing radiosonde and satellite PWs from the seven SSM/I algorithms for July 1987 to December 1990 (42 months) over the global oceans. In Fig. 8, the PW pairings represent 1-month averages for each of 56 radiosonde stations from all 42 months in the radiosonde profile–SSM/I TB paired time series. In the averages, only profiles associated with cloud LWPs below 0.5 kg m–2 are used. Thus scatter diagrams contain up to 2352 points. Most obvious systematic differences are for relatively small or large PWs (<20 or >50 kg m–2).
Insofar as the statistical methods, the ALI algorithm clearly underestimates at large PWs while slightly overestimating for small PWs. These same features are more extreme for the LOJ algorithm, and whereas the PET, S&E, and WEN-OS algorithms do not indicate high-end biases, S&E has a low-end bias and all three exhibit spreads similar to ALI and LOJ. In this framework, PET and WEN-OS are the best-behaved statistical algorithms in a relative sense.
For the physically based GRE and WEN-P algorithms, their scatter diagrams indicate greater spreads than the two best-behaved statistical algorithms. With respect to one another, the GRE algorithm is in better agreement at the high end whereas the WEN-P algorithm is in better agreement at the low end. Generally, the high-end spreads of both of these algorithms are greater than any single statistical algorithm. The lower right panel in Fig. 8 pairs the WEN-P results with the WEN-OS results, drawing attention to the fact that results from a physical algorithm with respect to the optimal statistical algorithm developed by the same author produces a scatter pattern in which, although the physical algorithm is biased high, the degree of spread is smaller than all radiosonde–satellite pairings. Underestimation at the high end with respect to the radiosonde for any of these algorithms may be related to inexact accounting of the nonlinear absorption properties of the SSM/I 22.235-GHz channel for large water vapor paths.
From Table 2 it is seen that the ALI and PET statistical algorithms generate the smallest biases considering all individual season and annual categories, with WEN-OS producing the best all around rms statistics. Compared to the two best performing statistical methods, the two physical methods indicate larger bias and rms factors. Consistent with the findings of Wentz (1995), the WEN-OS algorithm generates nearly comparable performance to the ALI and PET algorithms and moderately better performance than the S&E, GRE, and WEN-P algorithms.
b. Monthly averaged regional differences
The above verification framework is imprudent in view of how regional algorithm differences manifest themselves. In Figs. 9–11, we illustrate the magnitudes of monthly regional differences over the global oceans based on applying the seven published algorithms to the January and July 1990 global SSM/I datasets. In these figures, the contoured difference maps are based on subtracting the WEN-OS PW fields from each of the other six algorithm PW fields. Figure 9 illustrates the January–July monthly averaged PW comparison fields for WEN-OS, while Figs. 10 and 11 illustrate the difference fields for the other six algorithms.
There are significant meridionally and zonally aligned differences [i.e., north–south (N–S) and east–west (E–W) gradient structures] in all members of the January 1990 and July 1990 global difference maps. This is most evident in the meridional framework in which there are standing systematic differences in the various algorithms when zonal averages are taken (discussed further in the next section). Considering the analysis shown in Fig. 1, these results highlight a major problem intrinsic to satellite-based water vapor retrieval. That is, when considering the issue of water vapor transports, systematic errors in zonal averages of magnitudes indicated here (i.e., up to 5 kg m–2) produce direct transport errors at particular latitudes exceeding 60% of the mean transports themselves. According to the middle panel of Fig. 1, this magnitude of PW error produces meridionally distributed tropospheric IR radiative divergence errors of order ±10 W m–2, producing IR cooling rate gradient errors of order 0.2°C day–1. This imposes a meridionally distributed diabatic heating error on the thermodynamic energy equation. In a modeling context, such a gradient error in the thermodynamics would feed back on the dynamics, particularly the mean meridional mass circulation, which would then transfer error into the water conservation equation, thus introducing an additional source of error into the water vapor transports.
Considering both the January and July periods, the algorithm in best agreement with WEN-OS is S&E, exhibiting maximum differences of ∼±1.5 kg m–2 with well-balanced positive–negative bias distributions (Figs. 10e and 11e). Nevertheless, as small as these differences are, they exhibit a well-organized meridionally aligned gradient. Besides the S&E algorithm, the ALI algorithm is also in reasonable agreement with WEN-OS; however, its January difference field shows a well-organized large-scale meridional gradient with maximum differences exceeding positive 2 kg m–2 (Fig. 11a). The close agreement of the LOJ algorithm (Fig. 10c) has to be considered cautiously since this algorithm required adjustment of the calibrated SSM/I brightness temperatures to obtain consistent PW retrievals. The WEN-P algorithm exhibits meridionally aligned differences exceeding positive 6 kg m–2 in both the North Atlantic and Pacific in July (Fig. 11f), while GRE, PET, and WEN-P exhibit consistent large-scale organized meridionally aligned difference gradients exceeding 3 kg m–2.
c. Monthly/zonally averaged global differences
The use of a zonally averaged framework to summarize the contents of Figs. 9–11 quantifies systematic differences along the meridional axis. Figure 12 presents the meridional profiles for January (top panels) and July (bottom panels). The diagram includes the averaged WEN-OS PW results used as the reference profiles (refer to right ordinates) along with the zonal mean difference profiles for five of the six remaining algorithms (refer to left ordinates). The LOJ method has been eliminated from further analysis because its method of solution was tied to vagaries inherent to the European Centre for Medium-Range Weather Forecasts (ECMWF) forecast model and required modification of calibrated SSM/I brightness temperatures (see section c in appendix). The left two panels of Fig. 12 emphasize the lack of agreement between the six different algorithms and point back to various underlying issues in the GEWEX program concerning the validity of using satellite-retrieved water vapor quantities for validating climate models.
As anticipated in the previous section, the S&E difference profiles indicate that this algorithm is most consistent with the WEN-OS algorithm. However, although its meridional structures are relatively flat and close to zero compared to the others, it exhibits pronounced meridional structures in both January and July with distinct minima near 15°N and 30°S in January, and 45°N and 0° in July; and with distinct maxima near 60°N, 10°N, and 80°S in January, and 65°N, 10°N, and 65°S in July. The maximum excursion from 0 is the ∼2 kg m–2 difference in July near 45°S. The remaining four algorithms all give rise to much more pronounced positive and negative amplitudes in their systematic difference profiles, with the two physical algorithms (GRE and WEN-P) exhibiting the greatest meridional differences relative to WEN-OS (GRE differences are close to or exceed 4 kg m–2 beyond the 60° parallels while WEN-P differences exceed 5 kg m–2 near 45°N in July). Moreover, the two physical algorithm difference profiles exhibit little consistency between themselves. The latter cannot be said of the other two statistical algorithms (ALI and PET); that is, their difference profiles are in good agreement even though both exhibit pronounced minima and maxima. This is not surprising given that the PET algorithm was developed from a subset of the training data used for the ALI algorithm and is formulated using the same three frequencies as ALI (19, 22, and 37 GHz).
The most salient result observed in the left two panels of Fig. 12, when considering all difference profiles concurrently, is that the minimum and maximum excursions occur at equatorial, subtropical, and midlatitude latitudes, not unlike the zonally averaged profiles of cloudiness and precipitation (see Peixoto and Oort 1992).
d. Reducing systematic differences in statistical algorithms
A simple experiment is conducted to demonstrate how the meridionally distributed systematic differences between the four statistical algorithms can be reduced. This is done by retraining the ALI, PET, and S&E statistical algorithms with the same large rawinsonde observation (raob)–TB dataset originally used to train the WEN-OS algorithm, and then reassessing the meridional profile diagram generated from the three modified regression algorithms. Table 3 compares the original regression coefficients to the modified coefficients to demonstrate that significant differences in the regressions have arisen when a common training dataset is used. The new meridional difference profiles are shown in the right two panels of Fig. 12 in which the profiles of the three modified algorithms exhibit better systematic agreement with the WEN-OS algorithm, relative to the originals. There are still latitude-dependent differences with pronounced and distinct minima and maxima, although the relative agreement between ALI and PET has slightly improved (these two algorithms, which use the same channel input, were in good relative agreement at the outset), and the systematic agreement between the latter two and S&E has improved significantly. Even though all three modified algorithms exhibit a systematic positive bias relative to WEN-OS, a feature now that can be presumed to be related to the differences in which channel inputs the four statistical algorithms use (note WEN-OS does not include a 37-GHz channel unlike the three others), it is evident that an important source of systematic differences in the statistical algorithms is the content of the raob training dataset.
To emphasize this point, variance maps based on the WEN-OS algorithm and the three original statistical algorithms for January and July are compared to equivalent variance maps in which the three modified algorithms are substituted for the originals. These results are shown in Fig. 13. It is evident that the relatively larger magnitude and systematically aligned meridional/zonal variance patterns, so pronounced in Figs. 13a,b, give way to the smaller magnitude and more randomized patterns seen in Figs. 13c,d. Whereas training with a common dataset does not eliminate all differences, and would not be expected to eliminate all differences because the combined SSM/I channel selection–nonlinear formulations for each of these algorithms are distinct from one another, it is evident that better agreement is achieved. This suggests that in developing future statistical PMW algorithms, it would be wise to work from the same large, globally distributed, and carefully screened raob dataset, with as much consistency either in associating measured TB to the raobs if direct observation-based regression is used, or in radiative transfer modeling if indirect RTE-based regression is used.
e. Algorithm response to SST, Us, cloud LWP, and q(z) structure variations
To begin the third phase of the analysis involving RTE model-based response studies related to the tangential environmental factors, we first present the magnitudes of differences in how the six selected algorithms retrieve PW associated with the various constructed T–q profiles (recall that the initial PWs for the composite profiles are given in Table 1). Note that the original versions of the ALI, PET, and S&E algorithms are used in this analysis. Figure 14 presents the results for all four T–q profile datasets. In these calculations, the SSTs have been set to the lowest-level temperatures of the composite profiles (i.e., the surface air temperatures), while Us and cloud LWP have been set to 6 m s–1 and 0, respectively.
For each of the 14 composite profiles for all four datasets, the spreads in the retrieved PWs are significant, ranging from ∼3 to 9 kg m–2. For dataset 2 (the sub-Arctic cases), the spread for summer exceeds 20 kg m–2, overlooking the GRE algorithm, which produces a pathological result involving a negative PW. Note also that the WEN-P algorithm produces a negative result for winter. Although most of the spread in magnitude is due to the two physical algorithms (this without considering the two negative results for the sub-Arctic cases), with GRE generally running low and WEN-P generally running high, this is not always the case. Furthermore, depending on the profile dataset and the amount of moisture in the profiles, the relationships among the retrieved PWs from the four statistical algorithms and the profile PWs vary. As will be discussed, there are serious shortcomings concerning the two physical algorithms related to the pathological results for the sub-Arctic profiles.
The PW spreads for just the four statistical algorithms range from ∼3 to 5 kg m–2. Part of this variability results from using different channel combinations with the variety of linear and nonlinear algorithm formulations. In fact, most of PW differences found in Fig. 14 can be explained with the aid of the SST, Us, cloud LWP, and q(z) structure analyses because these are the main factors giving rise to variations, in synergy with the different combinations of SSM/I channels and different formulations used in the individual algorithms.
1) SST variations
With respect to the three SSM/I frequencies used for the six algorithms (19, 22, 37 GHz), SST variations impart the greatest influence on 19 GHz because its atmospheric attenuation in the presence of water vapor is less than for the two higher frequencies. This is because 19 GHz is relatively distant from the 22.235-GHz H2O absorption line and 37 GHz is more influenced by wing absorptance from both the 22.235-GHz H2O line and the 55–65-GHz O2 band [i.e., the transmittance of the 37-GHz window is significantly less than that at 19 GHz; see Fig. 2 in Smith et al. (1994a)]. Thus, any algorithm using 37 GHz in lieu of a 19 GHz to account for SST variations is less equipped to account for such variations because decoupling atmospheric from surface emission becomes all that much more difficult as atmospheric attenuation increases.
Figure 15 presents the SST analyses in conjunction with three of the four T–q profile datasets (1, 3, and 4). These calculations are made by holding Us constant at 6 m s–1, setting cloud LWP to 0, and varying SST over 20°C temperature ranges centered at the surface air temperatures (±10°C either side of a center SST). Note that there are variable slopes in the PW retrieval lines and variable offsets from the profile-based PWs, according to the particular composite profile and its associated SST range.
The nonzero slope behavior involves unaccounted-for SST effects contaminating the PW retrievals. Thus, the algorithms using only 22 and 37 GHz might be expected to exhibit the largest slopes (in an absolute sense) since they are least equipped to account for SST variations on radiances underpinning the specific algorithms. The only two algorithms formulated with just these two frequencies are S&E and WEN-P, that is, a statistical and physical algorithm. It is evident from Fig. 15 that these two algorithms consistently exhibit the greatest slopes.
It is also evident that the signed slopes of the four statistical algorithms are always positive, whereas those of the two physical algorithms are negative. The former might be explained by first noting that the characteristic range of SSTs associated with the raob training dataset is smaller than the SST ranges considered in the response calculations. Since the regression fits for the statistical algorithms intrinsically account for SST variability, but without having the advantage of confronting SSTs over a large range, it is possible that they uniformly underestimate the relatively weak dependence of upwelling radiance on SST, that is, due to the inverse relationship between surface emissivity and surface temperature.
On the other hand, the two physical algorithms both overestimate the influence of SST variability on PW. For GRE, it is no surprise that some type of slope behavior is present because this algorithm presumably attempts to deduce PW through the use of the 19-GHz degree of polarization in absence of a TB(22V) measurement. As noted in the appendix, GRE parameterizes an effective H2O mass absorption coefficient as a function of SST, which is a dubious assumption and, in fact, produces the pathological behavior found in Fig. 14 for the sub-Arctic case. (For these SST conditions, the GRE algorithm produces water vapor transmittances exceeding the physical limit of 1.0.) Moreover, in parameterizing 19-GHz differential emissivity with the TB(19V) and TB(19H), it uses an effective atmospheric emission temperature (Ta) that assumes fixed thermodynamics and stability. This is a poor assumption when considering modified air masses advecting over warm or cold ocean surfaces, possible situations in the vicinity of any coast and even in subtropical, midlatitude, or high-latitude open oceans under certain weather conditions. (This assumption also contributes to the problem with the sub-Arctic summer case in Fig. 14.) However, the most critical simplification in the GRE algorithm is that it actually is not dependent on 19-GHz degree of polarization after parameterizing for 19-GHz differential emissivity, which can be shown with a bit of mathematical analysis in their final PW formulation. The cumulative effect of all simplifications is a negative slope with respect to SST, and nonphysical behavior for extreme conditions beyond where the algorithm is reliable.
For WEN-P, since this algorithm does not include any 19-GHz quantity as an independent variable, it has to account for SST variations through its precision in determining the 37-GHz atmospheric transmittance (T37). As noted above, this becomes more difficult as the transmittance of a window frequency decreases. Moreover, as described in the appendix, the WEN-P algorithm solves a pair of 37-GHz TB equations given in terms of Us and T37, using these solutions to determine 22-GHz transmittance (T22). This means that both T22 and T37 are influenced by uncertainties in decoupling surface and atmospheric emission and in how surface emission and Us are related. Both result in overestimating the effects of SST variability on the retrieved PWs and, thus, the negative slope behavior in Fig. 15. Also, as for the GRE algorithm, pathological results occur for the sub-Arctic winter case, due to the lack of using a 19-GHz channel to optimize the sensitivity to surface boundary conditions produced by variable SST and nonzero Us conditions. (Here, the WEN-P solution results in a ∼40 m s–1 Us, compared to the specified value of 6 m s–1, along with the negative PW.)
The above discussion on the SST analysis, also clarifies why in Fig. 15 the two physical algorithms exhibit the greatest differences with respect to the profile PWs, and why the S&E algorithm exhibits the greatest differences relative to the other three statistical algorithms. Notably, these results are consistent with the results presented in Fig. 14. In general, the PWs of the statistical algorithms exhibit the smallest offsets relative to the profile-based PWs because they are most closely associated through statistical regression to the peculiarities and intrinsic errors in radiosonde measurements. By the same token, a 22–37-GHz-type algorithm, regardless of whether it is statistical or physical, exhibits a propensity for large departures whenever actual SST variation exceeds the magnitude of variation assumed in developing the algorithm. We should not conclude from these results that statistical algorithms are better performers because we have no means to certify the absolute accuracy of radiosonde-based PW information. However, we must conclude that, relative to SST variability, incorporating 19-GHz measurements into a PW retrieval scheme is advisable on the basis of the slope analysis presented in Fig. 15.
2) Us variations
Surface wind effects are more straightforward to interpret, neglecting indirect factors such as the details of the underlying surface roughness model. As with SST and for the same reasons, the frequency most responsive to Us effects on surface emission is 19 GHz. Figure 16 presents the Us analyses in conjunction with T–q profile datasets 1 and 4. These RTE calculations are made by fixing the SSTs to the associated surface air temperatures, setting cloud LWP to 0, and varying Us over the range 0–20 m s–1 for both datasets. As with the SST analysis, it is evident that there are variable slopes and offsets in the results relative to the profile-based PWs, in accordance with the particular profile dataset. Inspection of the figure indicates that if offsets are ignored and only slope behaviors considered, the least well-behaved algorithms are GRE, S&E, and WEN-P, consistent with the results found in the SST analysis. The additional point of interest is that the WEN-P algorithm seems to have the greatest difficulty in accommodating Us variations into its PW retrievals. Therefore, we can draw similar conclusions to those presented in the previous section as to the advisability of including 19-GHz measurements in passive microwave PW algorithms.
3) Cloud LWP variations
The effects of variations in cloud LWP are also straightforward to interpret. Sensitivity to the presence of cloud is greater at 37 than at 19 GHz, and although Rayleigh-size cloud droplets have less total ΔTB influence on 19-GHz TB than they do on 37-GHz TB, the foremost problem in PW retrieval is in accounting for the possible residual presence of cloud water, not in seeking the most transparent frequency. The presumption in this analysis, and for algorithm applications in general, is that clouds with large LWPs (say, exceeding 0.5 kg m–2) or clouds containing precipitation-sized droplets and/or significant concentrations of ice are not to be considered. Note that all published SSM/I PW algorithms include some type of screening procedure that seeks to ensure relatively cloud-free or Rayleigh-type cloud particle conditions before commencing a retrieval, although not all algorithms prevent precipitation-sized particles from contaminating the analyzed scenes, producing another potential source of algorithm differences.
Figure 17 presents the cloud LWP analyses in conjunction with the first and fourth T–q profile datasets. These calculations are made by fixing the SSTs to associated surface air temperatures and holding Us constant at 6 m s–1, while varying cloud LWP over the range 0–1 kg m–2. As before, there are variable slopes and offsets in the results relative to the profile-based PWs. From the above considerations, the algorithms that would be least well behaved insofar as slope magnitude would not include a 37-GHz measurement. The only algorithms not using 37-GHz TB are GRE and WEN-OS. Figure 17 indicates that when the slopes are considered, the least well-behaved algorithm is WEN-OS, and although the GRE algorithm performs better, it along with S&E exhibit difficulty in accounting for cloud LWP variability. Thus, 37 GHz is more effective than 19 GHz in accounting for the presence and variability of residual cloud LWP.
4) The q(z) structure variations
Finally, we consider variations in the q(z) structures. As alluded to previously, since the cumulative transmittance profiles at each of the three salient frequencies are different, variation of the q concentrations along a profile alters the frequency-dependent weighting functions (i.e., the vertical derivatives of the cumulative transmittance functions as accumulated from TOA to surface). In turn, this alters the sensitivity to ABL moisture where water vapor concentrations are largest, as well as the influence of temperature on the frequency-dependent H2O mass absorption coefficients, most potent at the vertically dispersed peaks of the weighting functions.
Figure 18 presents the q(z) structure analyses in conjunction with three of the four T–q profile datasets (1, 3, and 4). The RTE calculations are made by fixing the SSTs to the associated surface air temperatures, holding Us constant at 6 m s–1, setting cloud LWP to 0, and creating q(z) structure variability by constant scaling of the initial q(z) profiles over a range of ±25% and 50% (as discussed in section 3d). These lead to equivalent percentage PW ranges that are dependent on the initial PWs of the individual profiles. Again there are variable slopes and offsets in the results, according to the particular profile dataset and its associated PW ranges. However, unlike the other three sets of response analyses, the only algorithm exhibiting any significant slope departure (in this case from the diagonals of the various diagrams) is the WEN-P physical algorithm. Of note, these departures are only significant for the first hemispheric–seasonal profile dataset whose q(z) structures are more linear with respect to pressure than the more familiar q(z) profiles in the third and fourth T–q profile datasets whose PWs are more concentrated in the ABL (see Figs. 3b,d,e).
5. Discussion and conclusions
The foremost conclusion of this study is that a selection of current, well-accepted, and published SSM/I algorithms being used to retrieve PW are at odds with one another in representing the regional distribution of water vapor. These differences are in addition to more serious ones separating PMW from IR algorithms. The underlying differences are significant, particularly when using SSM/I retrievals for representing the detailed water-energy processes of the global climate system. We have emphasized how the uncertainties suggested by this study will impact the estimation of meridional transports of water vapor, and the meridional structure of radiative cooling. The former process is basic to the global water cycle (specifically to water vapor divergence), while the latter is basic to the earth radiation budget. Moreover, the radiative process through feedback on the thermodynamic state feeds back onto water vapor transport via the control of temperature on momentum and water conservation. Notably, there are other compelling climatic issues within the context of GEWEX not addressed here for which the implication of systematic regional errors in PMW moisture retrieval also represents a problem.
That said, there are straightforward explanations for the discrepancies and evidence that portions of the discrepancies could be removed through more exacting efforts to construct algorithms. The first issue in this regard pertains to statistical algorithms. We have shown how four such algorithms in initial disagreement, that is, those of Alishouse et al. (1990), Petty (1994), Schlüssel and Emery (1990), and Wentz (1995), can be made to improve their agreement by using a common radiosonde training dataset. However, development of any statistical algorithm based on radiosonde measurements always involves two types of residual uncertainty. The first is that the measurements are imperfect and may be so in complex systematic ways. The second is that statistical regression, regardless of how the TB are manipulated into independent variables, always leads to a certain degree of smoothing, which in turn disperses error characteristics as the independent variables are changed. This is why there has always been some interest in deriving statistical algorithms from regionally based training datasets and patching the results together at the seams. However, this is a dubious approach when seeking to develop global satellite datasets that resolve subtle climatic processes related to water vapor concentration and transport, because it guarantees algorithm artifacts at regional seams.
However, even with all of the criticism that can be leveled at statistical algorithms, there is evidence presented in this study that the statistical algorithms are currently outperforming the physical algorithms. That gets to the second issue pertaining to bringing about better agreement. The reason that complete agreement cannot be obtained between statistical algorithms using different SSM/I channels is that the tangential environmental factors affecting the top-of-atmosphere brightness temperatures impart varying degrees of influence at the different measuring frequencies. These varying TB effects then lead to different smoothing effects during the regressions. Therefore, even by appealing to a common training dataset, we have shown that not all discrepancies can be removed between statistical algorithms.
In concert with these difficulties are the performances of the physical algorithms, which are not all that well behaved, depending on how one views the credibility of the radiosonde measurements. Regardless of the outcome of that debate, much residual uncertainty remains in the physical algorithms when their results are cast into a regional framework and one sees that they possess large standing meridional and zonal PW gradients among themselves.
The real question is, why do physical algorithms produce inconsistent retrievals? Our answer here is twofold. First, the two physical algorithms we have studied are not yet sufficiently orthogonal to variations in the tangential environmental factors to produce well-behaved retrievals. All of the response analyses presented in section 4 lead to that conclusion in varying degrees. Second, neither of the physical algorithms is actually physical in the true sense of the word. There are various empirical, heuristic, and apocryphal simplifying assumptions inherent to both the GRE and WEN-P algorithms [as noted in the appendix and discussed in sections 4e and 4e(1)] that diminish their ability to resolve all of the complexities in the tangential environmental factors. In fact, there are simplifying assumptions inherent to these two algorithms that lead to serious errors because the underlying derivatives of PW with respect to SST and Us in the respective algorithm formulations break down beyond certain PW, SST, and Us limits. In essence, the physical algorithms under study contain the seeds of their own demise.
Therefore, for physical algorithms to improve, they have to unburden themselves from approximations that have the effect of generating unrealistic derivatives of the target variable “PW,” with respect to the tangential environmental factors such as SST, Us, and cloud LWP. This is not intended as a criticism of the algorithms under study or the intellectual effort that has gone into their development. Instead, it is a suggestion to the community of interest that if we are to proceed beyond the current situation of what might be called “algorithm estrangement,” it will be necessary to improve the underlying radiative transfer models for moist environments and in specifying surface boundary conditions for physical retrieval solutions.
Based on the response analyses, we conclude that the best PMW algorithm design in the presence of variations in the tangential environmental factors combines the 22-GHz with the 19- and 37-GHz channels on either side of the 22.235-GHz water vapor absorption line, because 19 GHz is most sensitive to surface boundary conditions while 37 GHz is most sensitive to the presence of residual cloud liquid water. This is not just intuitive, but is borne out by the results of the response analyses, which are contrary to the notion of some that 19 and 37 GHz are redundant in isolating the differential passive microwave signal of total path water vapor absorption. Whereas the results of the Wentz (1995) VIP study do not bear out our view completely [WEN-OS contains only TB(19V) and TB(22V)], it should be noted that his selection of an optimal algorithm was partly subjective based on considering both rms and systematic errors and that the 19-22-37 algorithms he examined produced error statistics of nearly equal skill. (Wentz’s results also highlight another problem with statistical algorithms; i.e., they do not unambiguously resolve physical controls on measured TB.)
For climate purposes, that is, monitoring water vapor concentrations and transports and validating climate simulations, it is important to improve physical algorithms since they are potentially better suited to identifying subtle changes in atmospheric moisture that lie outside the environmental ensembles used to generate statistical algorithms. In fact, such improvements are under way by various researchers in conjunction with new and advanced PMW radiometers. By the same token, to the extent that the magnitude of likely climate changes might be similar to seasonal changes, it may be that statistical algorithms are sufficient for some climate change research. However, it is certain that statistical algorithms cannot account for subtle concomitant changes that occur in water vapor profile structures and tangential environmental factors arising from changes not intrinsic to the current climate state. This is because statistical algorithms relentlessly smooth over the effects of the tangential environmental factors. Because of the complexity of the remote sensing problem, if physically based water vapor retrieval is to emerge supreme for climate change detection, it must incorporate more advanced methods interrelating the many environmental processes that determine atmospheric moisture. Above all, such algorithms depend on high stability in the measuring process because important climate changes may produce only minor changes in ambient moisture concentrations.
Acknowledgments
The authors express their appreciation to Mr. Do-Hyung Kim of Seoul National University for his assistance with the analysis, to Drs. John Bates, Peter Schlüssel, and Thomas Wilheit for their scientific advice, and to Dr. Frank Wentz for providing his VIP radiosonde dataset. This research has been supported by Grant NAG5-6665 of the National Aeronautics and Space Administration funded through the Code Y program for “satellite remote sensing measurement accuracy, variability, and validation studies” administered by Dr. James Dodge. The first author has also been supported by the SRC program within the Science and Engineering Foundation of Korea and the Korean Ministry of Education’s BK21 program.
REFERENCES
Albrecht, B. A., M. P. Jensen, and W. J. Syrett, 1995: Marine boundary layer structure and fractional cloudiness. J. Geophys. Res., 100 , 14209–14222.
Alishouse, J. C., S. A. Snyder, J. Vongsathorn, and R. R. Ferraro, 1990: Determination of oceanic total precipitable water from the SSM/I. IEEE Trans. Geosci. Remote Sens., 28 , 811–816.
Bates, J. J., 1991: High-frequency variability of Special Sensor Microwave/Imager derived wind speed and moisture during an intraseasonal oscillation. J. Geophys. Res., 96 , 3411–3423.
Betts, A. K., 1990: Greenhouse warming and the tropical water vapor budget. Bull. Amer. Meteor. Soc., 71 , 1465–1467.
Blankenship, C. B., A. K. Al-Khalaf, and T. T. Wilheit, 2000: Retrieval of water vapor profiles using SSM/T-2 and SSM/I data. J. Atmos. Sci., 57 , 939–955.
Bryan, F., and A. Oort, 1984: Seasonal variation of the global water balance based on aerological data. J. Geophys. Res., 89 , 11717–11730.
Cess, R. D., and Coauthors. 1990: Intercomparison and interpretation of climate feedback processes in 19 atmospheric general circulation models. J. Geophys. Res., 95 , 16601–16615.
Chahine, M. T., 1995: Observation of local cloud and moisture feedbacks over high ocean and desert surface temperatures. J. Geophys. Res., 100 , 8919–8927.
Chen, C-T., E. Roeckner, and B. J. Soden, 1996: A comparison of satellite observations and model simulations of column-integrated moisture and upper tropospheric humidity. J. Climate, 9 , 1561–1585.
Chen, T-C., and J. Pfaendtner, 1993: On the atmospheric branch of the hydrological cycle. J. Climate, 6 , 161–167.
Chou, M. D., 1994: Coolness in the tropical Pacific during an El Niño episode. J. Climate, 7 , 1684–1692.
Gaffen, D. J., T. P. Barnett, and W. P. Elliot, 1991: Space and time scales of global tropospheric moisture. J. Climate, 4 , 989–1008.
Gaffen, D. J., A. Robock, and W. P. Elliot, 1992: Annual cycles of tropospheric water vapor. J. Geophys. Res., 97 , 18185–18193.
Greenwald, T. J., G. L. Stephens, T. H. Vonder Haar, and D. L. Jackson, 1993: A physical retrieval of cloud liquid water over the global oceans using Special Sensor Microwave/Imager (SSM/I) observations. J. Geophys. Res., 98 , 18471–18488.
Greenwald, T. J., G. L. Stephens, S. A. Christopher, and T. H. Vonder Haar, 1995: Observations of the global characteristics and regional radiative effects of marine cloud liquid water. J. Climate, 8 , 2928–2946.
Hollinger, J. P., Ed.,. 1989: DMSP Special Sensor Microwave/Imager calibration/validation. CalVal Final Rep. Vol. I, Naval Research Laboratory, 182 pp.
Hollinger, J. P., Ed.,. 1991: DMSP Special Sensor Microwave/Imager calibration/validation. CalVal Final Rep. Vol. II, Naval Research Laboratory, 298 pp.
Houghton, J. T., J. J. Jenkins, and J. J. Ephraums, Eds.,. 1990: Climate Change—The IPCC Scientific Assessment. Cambridge University Press, 365 pp.
Houghton, J. T., J. J. Jenkins, and J. J. Ephraums, Eds.,. 1995: Climate Change 1995—The Second IPCC Assessment. Cambridge University Press, 572 pp.
Jackson, D. L., and G. L. Stephens, 1995: A study of SSM/I-derived columnar water vapor over the global oceans. J. Climate, 8 , 2025–2038.
Jedlovec, G. J., J. A. Lerner, and R. J. Atkinson, 2000: An upper-level water vapor transport index from geostationary satellites. J. Appl. Meteor., 39 , 15–41.
Kinter, J. L., and J. Shukla, 1990: The global hydrologic and energy cycles: Suggestions for studies in the pre-Global Energy and Water Cycle Experiment (GEWEX) period. Bull. Amer. Meteor. Soc., 71 , 181–189.
Liebe, H. J., 1985: An updated model for millimeter wave propagation in moist air. Radio Sci., 20 , 1069–1089.
Liebe, H. J., P. Rosenkranz, and G. Hufford, 1992: Atmospheric 60 GHz oxygen spectrum: New laboratory measurements and line parameters. J. Quart. Spectrosc. Radiat. Transfer, 48 , 629–643.
Lindzen, R. S., 1990: Some coolness concerning global warming. Bull. Amer. Meteor. Soc., 71 , 288–299.
Liu, W. T., 1986: Statistical relation between monthly mean precipitable water and surface-level humidity over global oceans. Mon. Wea. Rev., 114 , 1591–1602.
Liu, W. T., and P. P. Niler, 1984: Determination of monthly mean humidity in the atmospheric surface layer over oceans from satellite data. J. Phys. Oceanogr., 14 , 1451–1457.
Liu, W. T., W. Tang, and P. P. Niler, 1991: Humidity profiles over the ocean. J. Climate, 4 , 1023–1034.
Liu, W. T., W. Tang, and F. J. Wentz, 1992: Precipitable water and surface humidity over global oceans from Special Sensor Microwave Imager and European Center for Medium Range Weather Forecasts. J. Geophys. Res., 97 , 2251–2264.
Lojou, J-Y., R. Benard, and L. Eymard, 1994: A simple method for testing brightness temperatures from satellite microwave radiometers. J. Atmos. Oceanic Technol., 11 , 387–400.
Lutz, R., T. T. Wilheit, J. R. Wang, and R. K. Kakar, 1991: Retrieval of atmospheric water vapor profiles using radiometric measurements at 183 and 90 GHz. IEEE Trans. Geosci. Remote Sens., 29 , 602–609.
McClatchey, R. A., R. W. Fenn, J. E. A. Selby, F. E. Volz, and J. S. Garing, 1972: Optical properties of the atmosphere. AFCRL Rep. 72-0497, Air Force Geophysics Laboratory, Hanscom AFB, MA, 113 pp.
McCormick, M. P., E. W. Chou, L. R. McMaster, W. P. Chu, J. C. Larsen, D. Rind, and S. Oltmans, 1993: Annual variations of water vapor in the stratosphere and upper troposphere observed by the Stratospheric Aerosol and Gas Experiment II. J. Geophys. Res., 98 , 4867–4874.
Monahan, E. C., and I. O’Muircheartaigh, 1980: Optimal power-law description of oceanic whitecap coverage dependence on wind speed. J. Phys. Oceanogr., 10 , 2094–2099.
NRC, 1999: The GEWEX Global Water Vapor Project (GVaP)—U.S. Opportunities. National Research Council, National Academy Press, 17 pp.
Peixôto, J. P., and A. H. Oort, 1983: The atmospheric branch of the hydrological cycle and climate. Variations in the Global Water Budget, A. Steet-Perrott, M. Beran, and R. Ratcliffe, Eds., Reidel, 5–65.
Peixôto, J. P., and A. H. Oort, 1992: Physics of Climate. American Institute of Physics, 520 pp.
Petty, G. W., 1990: On the response of the Special Sensor Microwave/Imager to the marine environment—Implications for atmospheric parameter retrievals. Ph.D. dissertation, University of Washington, 291 pp.
Petty, G. W., 1994: Physical retrievals of over-ocean rain rate from multichannel microwave imaging. Part II: Algorithm implementation. Meteor. Atmos. Phys., 54 , 101–122.
Pierrehumbert, R. T., 1995: Thermostats, radiator fins, and the local runaway greenhouse. J. Atmos. Sci., 52 , 1784–1806.
Prigent, C., L. Phalippou, and S. English, 1997: Variational inversion of the SSM/I observations during the ASTEX campaign. J. Appl. Meteor., 36 , 493–508.
Raval, A., and V. Ramanathan, 1989: Observational determination of the greenhouse effect. Nature, 342 , 758–761.
Reitan, C. H., 1963: Surface dew point and water vapor aloft. J. Appl. Meteor., 2 , 776–778.
Rind, D., E-W. Chiou, W. Chu, J. Larsen, S. Oltmans, J. Lerner, M. P. McCormick, and L. McMaster, 1991: Positive water vapor feedback in climate models confirmed by satellite data. Nature, 349 , 500–503.
Salathé, E. P., D. Chesters, and Y. C. Sud, 1995: Evaluation of the upper-tropospheric moisture climatology in a general circulation model using TOVS radiance observations. J. Climate, 8 , 2404–2414.
Schlüssel, P., and W. J. Emery, 1990: Atmospheric water vapour over oceans from SSM/I measurements. Int. J. Remote Sens., 11 , 753–766.
Schmetz, J., and L. van de Berg, 1994: Upper tropospheric humidity observations from METEOSAT compared with short term forecast fields. Geophys. Res. Lett., 21 , 573–576.
Schulz, J., P. Schlüssel, and H. Grassl, 1993: Water vapour in the atmospheric boundary layer over oceans from SSM/I measurements. Int. J. Remote Sens., 14 , 2773–2789.
Scott, N. A., and Coauthors. 1999: Characteristics of the TOVS Pathfinder Path-B dataset. Bull. Amer. Meteor. Soc., 80 , 2679–2701.
Sissenwine, N., D. D. Grantham, and H. A. Salmela, 1968: Mid-latitude humidity to 32 km. J. Atmos. Sci., 25 , 1129–1140.
Smith, E. A., C. Kummerow, and A. Mugnai, 1994a: The emergence of inversion-type precipitation profile algorithms for estimation of precipitation from satellite microwave measurements. Remote Sens. Rev., 11 , 211–242.
Smith, E. A., X. Xiang, A. Mugnai, and G. J. Tripoli, 1994b: Design of an inversion-based precipitation profile retrieval algorithm using an explicit cloud model for initial guess microphysics. Meteor. Atmos. Phys., 54 , 53–78.
Smith, W. L., 1966: Note on the relationship between total precipitable water and surface dew point. J. Appl. Meteor., 5 , 726–727.
Soden, B. J., and F. P. Bretherton, 1994: Evaluation of water vapor distribution in general circulation models using satellite observations. J. Geophys. Res., 99 , 1187–1210.
Soden, B. J., and R. Fu, 1995: A satellite analysis of deep convection, upper tropospheric humidity, and the greenhouse effect. J. Climate, 8 , 2333–2351.
Soden, B. J., and F. P. Bretherton, 1996: Interpretation of TOVS water vapor radiances in terms of layer-average relative humidities: Method and climatology for the upper, middle, and lower troposphere. J. Geophys. Res., 101 , 9333–9343.
Sohn, B. J., 1994: Diagnosing temperature–moisture biases in ECMWF analyses based on clear sky longwave simulations constrained by SSM/I and MSU measurements and comparisons to ERBE estimates. J. Climate, 7 , 1707–1718.
Sohn, B. J., 1996: Analysis of the clear-sky greenhouse effect inferred from satellite data. J. Korean Earth Sci. Soc., 17 , 127–136.
Sohn, B. J., D-H. Kim, and H. S. Chung, 1998: Examing total precipitable water from SSM/I and TOVS over the East Asian summer monsoon region. Korean J. Atmos. Sci., 1 , 79–93.
Spencer, R. W., and W. D. Braswell, 1997: How dry is the tropical free troposphere? Implications for global warming theory. Bull. Amer. Meteor. Soc., 78 , 1097–1106.
Stephens, G. L., 1990: On the relationship between water vapor over the oceans and sea surface temperature. J. Climate, 3 , 634–645.
Stephens, G. L., and T. J. Greenwald, 1991: The Earth’s radiation budget and its relation to atmospheric hydrology. I. Observations of the clear sky greenhouse effect. J. Geophys. Res., 96 , 15311–15324.
Stephens, G. L., D. L. Jackson, and J. J. Bates, 1994: A comparison of SSM/I and TOVS column water vapor data over the global oceans. Meteor. Atmos. Phys., 54 , 183–201.
Stephens, G. L., D. L. Jackson, and I. Wittmeyer, 1996: Global observations of upper-tropospheric water vapor derived from TOVS radiance data. J. Climate, 9 , 305–326.
Stogryn, A., 1972: The emissivity of sea foam at microwave frequencies. J. Geophys. Res., 77 , 1658–1666.
Sun, D-Z., and R. S. Lindzen, 1993: Distribution of tropospheric water vapor. J. Atmos. Sci., 50 , 1643–1660.
Sun, J., 1993: Effects of vertical distribution of water vapor and temperature on total column water vapor retrieval error. J. Geophys. Res., 98C , 7069–7079.
Tjemkes, S. A., G. L. Stephens, and D. L. Jackson, 1991: Spaceborne observation of columnar water vapor: SSM/I observations and algorithm. J. Geophys. Res., 96 , 10941–10954.
Ulaby, F. T., R. K. Moore, and A. K. Fung, 1981: Microwave Remote Sensing Fundamentals and Radiometry. Vol. 1, Microwave Remote Sensing—Active and Passive. Artech House, 456 pp.
Valley, S. L., Ed.,. 1965: Handbook of Geophysics and Space Environment. U.S. Air Force Geophysics Directorate, 700 pp.
Wagner, D., E. Ruprecht, and C. Simmer, 1990: A combination of microwave observations from satellites and an EOF analysis to retrieve vertical humidity profiles over the ocean. J. Appl. Meteor., 29 , 1142–1157.
Wentz, F. J., 1992: Measurement of oceanic wind vector using satellite microwave radiometers. IEEE Trans. Geosci. Remote Sens., 30 , 960–972.
Wentz, F. J., 1995: The intercomparison of 53 SSM/I water vapor algorithms. Remote Sensing Systems Tech. Rep. on WetNet Water Vapor Intercomparison Project (VIP), Remote Sensing Systems, Santa Rosa, CA, 19 pp.
Wentz, F. J., 1997: A well-calibrated ocean algorithm for Special Sensor Microwave/Imager. J. Geophys. Res., 102 , 8703–8718.
Wentz, F. J., and R. W. Spencer, 1998: SSM/I rain retrievals within a unified all-weather ocean algorithm. J. Atmos. Sci., 55 , 1613–1627.
Wilheit, T. T., 1979: A model for the microwave emissivity of the ocean’s surface as a function of wind speed. IEEE Trans. Geosci. Electron, 17 , 244–249.
Wu, X., J. J. Bates, and S. J. S. Khalsa, 1993: A climatology of the water vapor band brightness temperatures from NOAA operational satellites. J. Climate, 6 , 1282–1300.
Yang, S., and E. A. Smith, 1999: Moisture budget analysis of TOGA COARE area using SSM/I-retrieved latent heating and large-scale Q2 estimates. J. Atmos. Oceanic Technol., 16 , 633–655.
Zhang, M. M., J. J. Hack, J. T. Kiehl, and R. D. Cess, 1994: Diagnostic study of climate feedback processes in atmospheric general circulation models. J. Geophys. Res., 99 , 5525–5537.
Zhao, B., 1994: Study of TOVS applications in monitoring atmospheric temperature, water vapor, and cloudiness in East Asia. Meteor. Atmos. Phys., 54 , 261–270.
APPENDIX
Algorithm Descriptions
ALI statistical algorithm
GRE physical algorithm
The GRE physical algorithm is actually a modified version of an algorithm originally developed by Tjemkes et al. (1991) and Stephens et al. (1994) and then used in conjunction with a cloud LWP retrieval algorithm developed by Greenwald et al. (1993). Another Greenwald et al. (1995) study, to which Jackson and Stephens (1995) refer as the source of the GRE algorithm, involves an extension of both the Tjemkes et al. (1991) PW and Greenwald et al. (1993) algorithms. It was first used for analysis of a 53-month SSM/I dataset in which the main focus was comparing SSM/I-derived cloud LWP retrievals to satellite-derived earth radiation budget and cloudiness data. The latter data were derived from Earth Radiation Budget Experiment (ERBE) and International Satellite Cloud Climatology Project (ISCCP) observations.
Note that the polarized emittance difference (εs19H – εs19V) is parameterized by the measurement-based ratio [TB(19H) – Ta]/[TB(19V) – Ta]. The Ta expression assumes fixed values for atmospheric lapse rate (Γ) and water vapor scale height (Hw), formulated as Ta = Ts + ΓHw(1 –
Although not stated in the Tjemkes et al. (1991), Greenwald et al. (1993), and Stephens et al. (1994) papers, their parameterization of εs19H – εs19V eliminates PW dependence on 19-GHz degree of polarization and reduces the PW expression to a regression equation in TB(19V) of the form PW = a0 + a1 ln[Ta – TB(19V)]. In this expression, Ta serves as the temperature bias for a “log brightness temperature” type variable, while a0 contains the influence of εs19V along with both TB(37V) and TB(37H) formulated within a surface wind speed term used to adjust εs19V for wind roughening effects. Thus, the GRE algorithm has the oddity of estimating PW using 19 and 37 GHz in absence of the actual water vapor sensitive frequency of 22 GHz, and without the degree of polarization information allegedly used to produce water vapor sensitivity through its control on polarization reduction through absorption–emission.
The study specifically focused on an assessment of the GRE PW retrievals was provided by Stephens et al. (1994) in a comparison of global-ocean monthly averaged PWs retrieved from a 1989 SSM/I dataset to coincident averages retrieved from a Television Infrared Observational Satellite (TIROS) Operational Vertical Sounder (TOVS) dataset. This study concluded that for the total data sample, the TOVS retrievals were drier than the SSM/I retrievals (∼0.25 g cm–2) because of inconsistencies between the two algorithms in how cloudy pixels were flagged (a bias that might be removed by using a consistent cloud-clearing approach). However, for large-scale subsidence regions undergoing steady drying (such as subtropical highs), or in the vicinity of deep convection where the entire troposphere is relatively moist (such as organized mesoscale systems), the TOVS retrievals are, respectively, systematically moister or drier than the SSM/I retrievals. This is because of the smaller radiometric (brightness temperature) contrast between a moisture-covered ocean surface and an unobscured (dry atmosphere) ocean surface occurring at infrared wavelengths, relative to millimeter–centimeter wavelengths.
LOJ statistical algorithm
PET statistical algorithm
The S&E statistical algorithm
WEN-OS statistical algorithm
WEN-P physical algorithm
Sensitivity of meridional water vapor transport and tropospheric infrared cooling to characteristic errors in representing latitude-dependent PW gradients. (top) Zonally and monthly averaged PW profile from NCEP reanalysis dataset for Jan 1990 along with corresponding clear-sky OLR profile based on infrared radiative transfer calculations for monthly averaged NCEP temperature–moisture profile conditions. (middle) Associated PW differences (ΔPW anomaly error) between PW retrievals for two SSM/I algorithms used in study (GRE and WEN-OS, see section 3) along with corresponding OLR anomaly error (ΔOLR) in calculation of OLR profile. (bottom) Associated vertically and monthly averaged northward oceanic meridional moisture transport [(
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
Radiosonde site map where open circles indicate 55 of 56 sites used for developing VIP training dataset consisting of 53 raob sites and 2 fixed weather ship sites (located off west coast of southern India). (Mobile weather ship site in Atlantic ocean used for VIP analysis is not indicated.) Here 19 sites with crosses indicate raob station subset selected for Alishouse et al. (1990) analysis
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
Thick lines show composite T–q profiles for first profile dataset from VIP averages stratified by hemisphere and season, with thin lines showing ±1σ limits.
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
Augmented composite q profiles for first profile dataset. Center profiles (thick lines) within each family of five are initial averages, while two pairs of perturbation profiles (thin lines) on either side are created by constant scaling of initial q(z) profiles by ±25% and ±50%.
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
Thick lines show composite T–q profiles for third profile dataset (i.e., VIP zonal averages for regions 1–4), with thin lines indicating ±1σ limits.
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
Augmented composite q profiles for third profile dataset. Center profiles (thick lines) within each family of five are initial averages, while two pairs of perturbation profiles (thin lines) on either side are created by constant scaling of initial q(z) profiles by ±25% and ±50%.
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
(a), (b) Composite T–q profiles for fourth merged profile dataset, i.e., Albrecht et al. (1995) marine boundary layer profiles merged with VIP zone-averaged profiles above ABL top. (c)–(f) Augmented composite q profiles for fourth merged profile dataset. Center profiles (thick lines) within each family of five are initial averages, while two pairs of perturbation profiles (thin lines) on either side are created by constant scaling of initial q(z) profiles by ±25% and ±50%.
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
First seven panels are scatter diagrams between radiosonde-based (abscissas) and satellite-based (ordinates) PWs from seven SSM/I algorithms for Jul 1987 to Dec 1990 (42 months) over global oceans. Lower-right panel is scatter diagram between WEN-OS (abscissa) and WEN-P (ordinate) PWs. Points represent 1-month averages for each of 56 radiosonde stations from all 42 months in radiosonde–SSM/I TB time series, but for which radiosonde profiles whose associated SSM/I brightness temperatures indicate cloud LWPs greater than 0.5 kg m–2 are eliminated
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
Global distributions of PW (kg m–2) for (top) Jan 1990 and (bottom) Jul 1990 based on WEN-OS. Contour interval is 10 kg m–2
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
Global PW difference maps (kg m–2) for Jan 1990 between WEN-OS and (a) ALI, (b) GRE, (c) LOJ, (d) PET, (e) S&E, and (f) WEN-P. Contour interval is 1 kg m–2 for (a), (b), (c), and (f), and 0.5 kg m–2 for (d) and (e). Larger positive differences are shaded; negative differences are dashed.
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
(b) Same as Fig. 10 except for Jul 1990
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
(top left), (lower left) Meridional profiles of monthly zonally averaged WEN-OS results for Jan 1990 and Jul 1990, respectively, indicated with thick solid lines and right ordinates, along with meridional profiles of differences with respect to five original algorithms indicated with thin solid or broken lines and left ordinates (excludes LOJ). (top right), (bottom right) Similar to (top left) and (lower left) except with results for three modified statistical retrieval algorithms but without repeat plotting of difference profiles for GRE and WEN-P physical algorithms
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
(a), (b) Variance maps of PW (kg2 m–4) based on WEN-OS and three original statistical retrieval algorithms (excludes LOJ) for Jan 1990 and Jul 1990; (c), (d) similar results for three modified algorithms along with original WEN-OS algorithm
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
Summary of intercomparison of radiosonde-based and SSM/I algorithm–based PWs (for six algorithms) using PMW radiative transfer model to associate TB to constructed profiles for all four T–q profile datasets
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
SST sensitivity analysis for (left) first, (middle) third, and (right) fourth T–q profile datasets. Dashed horizontal lines indicate magnitudes of PWs associated with four different q(z) profiles.
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
The Us sensitivity analysis for (left) first and (right) fourth T–q profile datasets. Dashed horizontal lines indicate magnitudes of PWs associated with four different q(z) profiles.
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
Cloud LWP sensitivity analysis for (left) first and (right) fourth T–q profile datasets. Dashed horizontal lines indicate magnitudes of PWs associated with four different q(z) profiles.
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
The q(z) structure sensitivity analysis for (left) first, (middle) third, and (right) fourth T–q profile datasets. Abscissas give perturbation scales for PW
Citation: Journal of Climate 16, 20; 10.1175/1520-0442(2003)016<3229:ESODII>2.0.CO;2
Number of profile samples and PWs associated with each of 14 composite profiles for all four T–q profile datasets. Note PWs for dataset 4 refer only to boundary layer portions. Profile sample counts are not given for sub-Arctic summer and winter cases. [See original references of Valley (1965) and Sissenwine et al. (1968) for explanation of how sub-Arctic profiles were generated]
Intercomparison statistics (kg m–2) between radiosonde-based and satellite-based PWs from seven SSM/I algorithms for Jul 1987 to Dec 1990 (42 months) over global oceans. Statistics are derived from point pairings, as graphically illustrated in Fig. 8 (consisting of mean, standard deviation σ, bias, and unadjusted rms) and given for four calendar season and annual periods: DJF, Mar–May (MAM), JJA, and Sep–Nov (SON)
Tabulation of original (orig) and modified (mod) regression coefficients for ALI, PET, and S&E algorithms along with WEN-OS original coefficients; independent variables are also indicated for each algorithm