1. Introduction

Accurate measurements of the space–time distribution of precipitation are essential not only for a better understanding of the global water and energy cycle but also for improving weather, climate, and hydrological predictions. With the coverage of traditional ground-based measurements from gauges and radars concentrated over well-populated land regions, a comprehensive description of the space–time variability of global precipitation can only be achieved with observations from space.

Although infrared (IR) and visible (VIS) instruments can provide information such as cloud cover and cloud-top temperatures from which to infer precipitation rates (e.g., Kilonsky and Ramage 1976; Griffith et al. 1978; Arkin 1979; Arkin and Meisner 1987), the backbone of space-based precipitation measurements are microwave sensors that directly respond to the absorption and scattering of cloud hydrometer particles (e.g., Wilheit et al. 1977; Spencer et al. 1989; Liu and Curry 1992; Petty 1994a, b; Kummerow et al. 1996; Olson et al. 1996).

As shown in Wilheit (1986) and Spencer et al. (1989), liquid drops absorb and scatter microwave radiation, with absorption dominating the radiative transfer, whereas ice generally only scatters microwave radiation. Both scattering and absorption tend to increase with frequency and with rain rate. Based on these radiative properties, rain rates in tropical oceans can be estimated through sensing the thermal emission (and scattering) of raindrops over the radiometrically cold ocean at relatively low passive microwave (PMW) frequencies, whereas rain rates over land can be estimated from the scattering signal due to ice particles at relatively high frequencies (e.g., Wilheit et al. 1977; Spencer et al. 1983; Wilheit 1986; Spencer 1986; Grody 1991; Adler et al. 1993).

Among microwave sensors, the quality of rain retrievals varies with sensor characteristics, surface emission, and many other factors. For example, because of the low, relatively uniform ocean surface emissivity, rain rates over the tropical and subtropical oceans can be physically retrieved with good accuracy from conically scanning radiometric imager data using emission-based (or emission–scattering-combined) rainfall algorithms (e.g., Wilheit 1986; Spencer et al. 1989; Petty 1994a, b; Smith et al. 1994a, b,c; Kummerow et al. 1996; Olson et al. 1996). However, over land where the scattering-based algorithms are usually applied, rain rates are still largely empirically retrieved and have significant uncertainties. Not only is the scattering of microwave radiation highly dependent on the poorly known details of the ice-size distribution, but also the precipitation estimate is primarily related to the presence of frozen hydrometeors aloft, which have a less direct relationship with surface rain rates. Further, highly variable land surface—including different vegetation cover, snow and ice cover, and soil wetness, all of which may change significantly with time—makes it difficult to distinguish rainfall pixels from complicated surface backgrounds (e.g., Wilheit 1986; Kummerow et al. 2001; Ferraro et al. 2005). Rain retrievals from microwave sounder data may have uncertainties over land similar to those from radiometer data. On the other hand, applying algorithms based on scattering only to high-frequency microwave sounder data over tropical and subtropical oceans usually leads to biases in detecting rain events that lack ice phase processes (e.g., Joyce et al. 2004; Ferraro et al. 2005; Huffman et al. 2007). Quantitatively evaluating these satellite rainfall retrievals against the “true” rainfall distribution over both ocean and land is therefore an important and challenging task.

A major obstacle in evaluating satellite rainfall products is that there are no perfect observations to serve as the “truth.” Despite well-known issues with ground-based measurements (e.g., calibration and representativeness errors), validation studies often use rain gauge and surface radar data as the reference to directly evaluate satellite rainfall retrieval algorithms (e.g., Allam et al. 1993; Smith et al. 1998; Ebert and Manton 1998; Kummerow et al. 1998, 2001; Olson et al. 2006). Although many valuable insights have been gained into the retrieval error statistics by using surface measurements to evaluate satellite rainfall estimations, a number of issues remain when using ground measurements as the reference:

• (i) Evaluations of satellite rainfall retrievals against ground-based observations are typically performed over a few sites over open oceans and some well-instrumented land areas (e.g., Smith et al. 1998; Yang et al. 2006). Because the rainfall error statistics can be highly regime-dependent, there have been concerns that the error statistics derived from limited oceanic sites and land areas may not be applicable to other parts of the world (e.g., Kummerow et al. 1998; Nesbitt et al. 2004; Berg et al. 2006).

• (ii) PMW sensors and their rainfall retrieval algorithms have different sensitivities on the rain/no rain detection, and rain retrieval errors may vary as a function of rain intensities. Because of the small number of raining samples resulting from limited satellite overpasses over ground validation sites, the traditional scatterplot and a single correlation number are not able to quantitatively describe the nonlinear properties of rainfall retrieval error statistics at different rain intensities (e.g., Conner and Petty 1998).

• (iii) Monthly-mean or seasonal-mean comparisons are usually performed to evaluate rainfall retrievals on board different satellites at regional and global scales. Most satellites are nongeostationary, and if retrievals are not coincidently compared with a common reference, retrieval errors, random errors, and diurnal undersampling errors would all be merged in the computed statistics (e.g., Wilheit 1986; Hong et al. 1997; Lin et al. 2002), which makes it difficult to feed back the retrieval problem to the algorithm developers.

The successful operation of Tropical Rainfall Measuring Mission (TRMM; Simpson et al. 1988, 1996) has provided new and extensive space-borne observations of the hydrologic cycle in the tropics and subtropics. Of the two rain-sensing instruments flown on TRMM, the precipitation radar (PR; Meneghini and Kozu 1990; Kummerow et al. 1998, 2000; Iguchi et al. 2000) provides the first spaceborne active microwave rainfall retrieval, with a horizontal footprint of 5 km over a swath of 247 km (after the orbit boost in August 2001). Not only have PR data been undergoing rigorous internal and external calibrations, they have also been evaluated favorably against surface observations over different oceanic and land validation sites (e.g., Schumacher and Houze 2000; Liao et al. 2001). Considering the PR’s stable and active microwave sensing features (especially its theoretical superiority to overland PMW technique), the PR data together with conventional ground-based measurements can provide independent yet mutually supplemental references for assessing other rainfall retrievals.

To guide the planning of future satellite precipitation-measuring missions, we have compared instantaneous rainfall estimates provided by several widely available operational rainfall products in the tropics and subtropics. To address some of the concerns regarding rainfall evaluation discussed in the previous section, both TRMM PR data and merged surface radar and gauge data over the continental United States are used as references to quantitatively evaluate the error statistics of surface rainfall retrievals derived from three operational sounders—the Advanced Microwave Sounding Unit-B (AMSU-B) instruments on NOAA-15, -16, and -17 satellites—and five imagers: the TRMM Microwave Imager (TMI; Kummerow et al. 1998), the Special Sensor Microwave Imager (SSM/I) instrument on the Defense Meteorological Satellite Program (DMSP) F-13, -14, and -15 satellites, and the Advanced Microwave Scanning Radiometer for the Earth Observing System (AMSR-E) instrument on the Aqua satellite.

In this study, comparisons of PMW retrievals against PR data are based on “coincident” observations, defined as instantaneous retrievals (spatially averaged to 0.25° in latitude and 0.25° in longitude) within a 10-min interval collected over a 20-month period from January 2005 to August 2006. Further coincident comparisons against surface radar and gauge data are made at 1-h intervals for a 3-month period from June to August 2005 over the continental United States south of 35°N. The goal is to quantitatively assess surface rain retrievals from cross-track scanning microwave humidity sounders relative to those from conically scanning microwave imagers. Statistics of departures of the coincident retrievals from reference measurements as given by the TRMM PR and ground radar and gauges are computed as a function of rain intensity over land and ocean.

Section 2 introduces satellite rainfall retrievals, explains how they are analyzed, and discusses the definition of coincidence and coincident samples for different sensors. Section 3 examines the rain–no rain detection in different PMW retrievals. Sections 4 and 5 present results from collocated comparisons against TRMM PR over land and ocean, focusing on quantifying retrieval error statistics for different PMW rainfall estimates. Section 6 further examines the satellite rainfall retrievals, including those from TRMM PR against ground measurements over the continental United States. Section 7 presents the final conclusions of the study and proposes some future work for further quantitative evaluations of satellite rainfall estimations.

2. Data and analysis methods

3. Rain detection

Successful simulations of rain frequency and rain intensity at high temporal and spatial resolutions represent key metrics for testing the influence of model physical parameterizations on the hydrological cycle. As an important resource of observational dataset, satellite rain retrievals, however, may have significantly large differences in rain incidence and probability density function of raining events because of sensors’ and algorithms’ technical detection capabilities and the employment of various rain–no rain screening techniques (Grody 1991; Ebert et al. 1996; Smith et al. 1998; Ferraro et al. 1998). Such problems in rain pixel identification become especially severe over highly variable surface backgrounds and/or when examining light-rain events; this will be an important issue to be improved in future satellite missions.

In this study, instead of examining the discrepancies of rain detection capabilities at the pixel level, we evaluate the area-averaged rain rate at a 0.25° × 0.25° resolution so that different retrievals can be meaningfully compared at the same resolution. Similar to those calculations performed by Kummerow et al. (2001) and Chen and Staelin (2003) using ground-based measurements, all the collocated rain retrievals (including both zero and nonzero rain rates) are stratified into four groups: (i) both PR and PMW sensor rain rates = 0, (ii) both PR and PMW sensor rain rates > 0, (iii) PR rain rate > 0 but the sensor rain rate = 0, and (iv) PR rain rate = 0 but the sensor rain rate > 0.

Figure 2 shows the percentage of incidence of these four groups relative to the total coincident samples over the ocean between 35°S and 35°N (the TRMM ocean domain, hereinafter referred to as TRMM Ocean). As expected, the situation when retrievals from both PR and individual PMW sensors indicate zero rain rates dominates the coincident samples (ranging from 82% to 85%), with TMI and AMSR-E at the low end, SSM/I in the middle, and AMSU-B at the high end. The imager and sounder data indicate large differences in the situation when retrievals from both PR and PMW sensors indicate above-zero rain rates: there are about 8% of coincident samples in which TMI and AMSR-E indicate rain incidence when PR does. SSM/I retrievals, although employing the same GPROF 6 algorithm, have only about 5% of coincident samples showing nonzero rain simultaneously. Retrievals from AMSU-B, however, indicate the lowest frequency of occurrence (slightly above 2%).

Large discrepancies can also be noticed in the situation when PR can detect rain but individual PMW sensors cannot. Among different PMW retrievals, more than 10% of AMSU-B total coincident samples cannot detect any rain over ocean when PR sees rain at the same time. The frequency of occurrence for the sounder data is about 2%–3% higher than those for SSM/I and is almost twice as much as those for TMI and AMSR-E. Joyce et al. (2004), in developing a methodology to produce a high-resolution global precipitation product, compared SSM/I and AMSU-B retrievals against TMI rainfall estimates. They attributed the small difference between SSM/I and TMI to the different retrieval footprint resolutions resulting from satellite altitudes; the large difference between AMSU-B and TMI is mainly associated with AMSU-B’s high-frequency-only channels that cannot detect liquid-phase hydrometeors over ocean. Huffman et al. (2007), in calibrating different PMW retrievals using TMI data, also noticed that the AMSU-B fractional occurrence of precipitation in the subtropical oceans where light-rain events frequently occur is notably deficient relative to TMI. Considering that the situation in which PR indicates no rain but individual PMW sensors suggest nonzero rain occurs infrequently especially for SSM/I and AMSU-B retrievals, most of the above discrepancies among PMW rainfall estimates support earlier findings that over tropical and subtropical oceans, the sounder data, not being able to detect emission signatures of liquid-phase hydrometeors, have less sensible detection capabilities than the imager data.

By dividing the collocated PR rain estimates into different rain intensity bins, we can further examine the corresponding PMW detection capability as a function of PR rain intensities. Displayed in Fig. 3 are coincident sample numbers and percentages of PMW rain incidence for different PR rain intensity bins. Except for TMI, which is always collocated with PR, the binned coincident sample numbers for other satellites are similar and range between 100 and 10 000 for rain rates <30 mm h⁻¹, large enough for meaningful statistical computations. Above 30 mm h⁻¹, sample numbers for each rain intensity bin range from 10 to 100 and caution should be taken when examining the statistical results.

All the PMW sensors indicate improved detection capabilities as PR rain intensity becomes larger. AMSR-E and TMI overall compare better to PR than SSM/I and AMSU-B do, and percentages of rain incidence increase from 50% at 0.1 mm h⁻¹ to 93% at 1.0 mm h⁻¹ for AMSR-E and from 35% at 0.1 mm h⁻¹ to 90% at 1.0 mm h⁻¹ for TMI. Neither SSM/I nor AMSU-B compares well to PR at light-rain rate ranges, but SSM/I generally performs better than AMSU-B, with the percentage of rain incidence around 55% at 1.0 mm h⁻¹. On the other hand, not being able to detect liquid-phase hydrometeors over ocean, AMSU-B tends to miss most of the light-rain events, with the percentage of rain incidence around 7% at 0.1 mm h⁻¹ and 25% at 1.0 mm h⁻¹. One interesting thing to note is that when the PR rain intensity is above 10 mm h⁻¹, all imager data can detect rain incidence as PR does, but there are still 6%–10% of AMSU-B coincident samples that fail to detect any rain even though PR suggests such high rain intensities. This discrepancy could probably result from heavy precipitating events that do not involve ice microphysical processes.

Although it is arguable which retrievals are better over ocean, the retrievals from the imager data (especially TMI and AMSR-E) or the PR estimates, PR data are generally considered to outperform PMW retrievals over land. The difficulties in distinguishing small radiance contrasts between rainfall and the land surface based on the scattering-only signatures have made PMW rainfall estimations over land a very challenging work. Very limited cloud simulation samples over land, together with highly variable land surface (e.g., wetness, vegetation coverage, seasonal changes) which involves many screening techniques, further complicate the problem (e.g., Grody 1991; Adler et al. 1994; Ferraro et al. 1998; Kummerow et al. 2001). Over the TRMM land area between 35°N and 35°S (hereinafter referred to as TRMM Land; Fig. 4), the situation in which both PR and PMW retrievals indicate zero rain rates again dominates the total coincident samples (about 85%–90%). The incidence when both PR and PMW retrieval indicate above-zero rain rates is about 3%–5%, with retrievals from AMSU-B indicating a slightly higher frequency of occurrence. The most prominent feature is that the PMW retrievals from both the imager and sounder data tend to have similar incidence for the situation in which PR suggests rain but individual PMW sensors do not. AMSU-B data even indicate a slightly higher frequency of occurrence (1%–2% higher) relative to the imager data when PMW sensors indicate rain but PR does not.

Figure 5 further shows the coincident sample numbers and percentages of PMW rain incidence for different PR rain intensity bins over the TRMM Land. Again, the binned coincident sample number for each satellite is large enough for meaningful statistical analysis. All the passive microwave retrievals generally indicate similar detection capabilities over land. But there are subtle differences: below 1.0 mm h⁻¹, the sounder data overall agree better with the PR data than the imager data do in terms of rain incidence, whereas above 1.0 mm h⁻¹, the sounder data appear to perform equally well with TMI and AMSR-E and are in general better than SSM/I.

Because the rain retrieval algorithms have significantly different features over land and ocean, in the follow sections we evaluate their frequency distribution functions and rain error statistics over TRMM Ocean and TRMM Land separately.

4. Coincident comparisons with PR over TRMM Ocean

Many earlier studies have evaluated PMW rainfall retrievals from individual sensors on board single satellites over tropical and subtropical oceans. Here we compare several instantaneous PMW rainfall retrievals from imagers and sounders on board different satellites with PR data over ocean. Figure 6 shows the conventional scatterplot between collocated rain retrievals from TRMM PR and individual PMW sensors over TRMM Ocean for August 2005. Similar to what has been noticed in earlier studies, the coincident rain samples are dominated by samples with rain intensities below 10 mm h⁻¹. Although some underestimations can be noted at very high intensities, TMI retrievals appear to agree pretty well with PR at all high, intermediate, and low rain intensities. This is consistent with in the findings of Yang et al. (2006), namely, that different GPROF versions of TMI rainfall retrievals at instantaneous 0.5° resolution have high correlations with coincident PR estimates over tropical and subtropical oceans. Rain estimations from AMSR-E, SSM/I, and AMSU-B also correlate nicely at high and intermediate rain intensities. However, close to and below a rain intensity of 1 mm h⁻¹ (which is about the cutoff rate at these retrievals’ pixel level), both SSM/I and AMSU-B retrievals appear to indicate slightly higher intensities than their coincident PR data and tend to set their rain rates around the cutoff rates for many situations in which PR rain rates are below 1 mm h⁻¹. Considering what is shown in Fig. 3 (i.e., that SSM/I and AMSU-B tend to have much lower detection capabilities than TMI and AMSR-E), we hypothesize that such discrepancies in rain incidence and scatterplot are primarily associated with the difference in sensitivity and detection capability among PMW sensors in light-rain events. SSM/I and AMSU-B tend either to retrieve no-rain situations or to bias toward higher rain rates (the beamfilling problem), whereas TMI and AMSR-E have better sensitivities in light-rain cases because of the inclusion of information from better resolutions, more emission channels, and careful calibrations with TRMM PR data.

To further illustrate the impact of detection capabilities as a function of PR rain intensities among various PMW retrievals, probability density functions (PDFs) of collocated raining samples over the TRMM Ocean are shown in Fig. 7. Consistent with earlier studies (e.g., Kummerow et al. 2001; Joyce et al. 2004; Nesbitt et al. 2004), PR data indicate that a large portion of raining samples consists of light-rain events. The frequency of rain occurrence from both TMI and AMSR-E agree pretty well with PR data above 1 mm h⁻¹, with some small underestimations below 1 mm h⁻¹. AMSR-E estimates appear to be even slightly better than TMI. The three SSM/I datasets tend to underestimate PR more than TMI and AMSR-E do in term of the frequency of light-rain events and to overestimate PR in the range between 1.5 and 6.0 mm h⁻¹. Rain retrievals from AMSU-B appear to need some more improvement over the tropical and subtropical ocean. Most of the light-rain events are not sufficiently detected because AMSU-B is less capable of detecting the prevalent precipitation from warm-topped clouds (Petty 1999; Joyce et al. 2004; Huffman et al. 2007). Similar to what shown in the SSM/I plots, rain events at intermediate rain intensities (ranging between 2.0 and 10.0 mm h⁻¹, higher that those shown in SSM/I) are overly sampled by AMSU-B. These underestimations of the frequency of light-rain events and overestimations of the frequency of intermediate-rain events in SSM/I and AMSU-B retrievals, also noted in Joyce et al. (2004), are probably manifestations of the misrepresentation of light-rain events by these PMW sensors: they could detect too many no-rain situations and/or sometimes overestimate the light-rain intensity, thus aliasing errors into intermediate rain rate ranges.

Because statistical measures could be easily weighted toward raining samples that occur the most frequently (Conner and Petty 1998), namely, the light-rain events, rain error statistics have long been recognized as a strong function of rain intensity. A number of studies have examined single PMW rainfall retrievals against PR or ground validation data at different rain intensities (e.g., Chen and Staelin 2003; Olson et al. 2006); however, the rain error statistics among various PMW retrievals on board different satellites have not been thoroughly compared. In this study, by grouping a tremendous amount of coincident PR rain retrievals into rain intensity bins, systematic biases, root-mean-square (RMS) errors, and standard deviation (STD) errors can be evaluated for various PMW retrievals against the same reference framework. Such error statistics are also very useful for assimilating different satellite rainfall retrievals into numerical weather prediction models. For each individual PR rain intensity bin, the coincident retrievals from PMW sensors can be any valid value, including both zero and nonzero rain rates.

Figure 8 shows the normalized systematic biases and RMS errors in percentage for collocated PR rain intensities (at a 0.25° × 0.25° resolution) from January 2005 to August 2006 over ocean areas sampled by TRMM. Areas with rain intensity <1.0 mm h⁻¹, representing the largest minimum detectable rain rate for these PMW sensors, are lightly shaded. Rain retrievals from both imager and sounder data perform very well between 0.3 and 10 mm h⁻¹ and the largest systematic bias is smaller than 20%. There is a negative bias in high-rain intensity ranges (above 10 mm h⁻¹) for all PMW retrievals. The amplitude of the negative bias tends to become larger as the rain intensity increases, and the sounder data appear to have larger negative biases than the imager data. Below 0.3 mm h⁻¹, TMI and AMSR-E retrievals have smaller biases than retrievals from SSM/I and AMSU-B. There are large positive biases below 0.3 mm h⁻¹ for SSM/I and AMSU-B retrievals, and the normalized bias increases with reducing rain intensities. These positive biases, as discussed earlier, are possibly associated with the frequent misrepresentation of light rain as intermediate-rain pixels in SSM/I and AMSU-B algorithms.

All the PMW rainfall retrievals indicate that their normalized RMS errors tend to have an inverse relationship with the TRMM PR rain rate, so that relative errors become much larger in light-rain events and smaller in heavy-rain events. These features have been noticed in many earlier studies over both oceanic and land areas (e.g., Bell and Reid 1993; Huffman 1997; Olson et al. 2006). For heavy-rain events (above 10 mm h⁻¹), all the algorithms perform pretty well over the TRMM Ocean. The largest relative RMS errors, at a resolution of 0.25° × 0.25°, range between 50% and 70% for the imager and sounder data. Below 5.0 mm h⁻¹, however, there are significant differences in the relative RMS error among these PMW retrievals. Both TMI and AMSR-E retrievals continue to outperform SSM/I retrievals from three DMSP satellites and AMSU-B retrievals from three NOAA satellites, with the RMS error close to 100%–120% at a rain intensity of 1.0 mm h⁻¹. AMSU-B retrievals, although from different NOAA sun-synchronous satellites, agree with one another and have the largest RMS error relative to AMSR-E, TMI, and SSM/I. The relative RMS error is around 230% at a rain intensity of 1.0 mm h⁻¹.

Taking their minimum cutoff rain rates into account, all the PMW retrievals have similarly small systematic biases over the TRMM Ocean in comparison with PR data. For instantaneous rain rates above 10 mm h⁻¹, all the PMW retrievals also have similar RMS errors, indicating that both imager and sounder data have similar capabilities to capture heavy-rain events. For instantaneous rain rates below 5 mm h⁻¹, the imager data clearly have smaller RMS errors than the sounder data over tropical and subtropical oceans, especially in light-rain events. Part of the reason may be associated with the fact that AMSU-B rain estimates retrieve precipitation based on the relationship between surface precipitation and ice water content and thus miss the frequent and important warm-rain processes in the tropics and subtropics (Petty 1999; Joyce et al. 2004; Huffman et al. 2007).

5. Coincident comparisons with TRMM PR over land

It is well known that rain retrievals by PMW sensors are less reliable over land than those over ocean because of the empirical relationship between cloud ice content and surface rain rate, as well as the difficulties in distinguishing rain pixels from nonuniform and complicated backgrounds. Improving PMW rainfall estimation accuracy over land presents an important challenge to any future satellite precipitation measuring mission. On the other hand, because PR does not need to take varying land surface into account, spaceborne PR estimations are both theoretically and practically superior to PMW retrievals over land. Displayed in Fig. 9 are scatterplots between collocated rain retrievals from TRMM PR and individual PMW sensors over the TRMM Land for August 2005. All the PMW retrievals show good agreement with TRMM PR above 1.0 mm h⁻¹, suggesting that both PR and PMW instantaneous rain retrievals can capture heavy and intermediate rain events over land reasonably well. For PR rain rates below 1.0 mm h⁻¹ there is a tendency for PMW retrievals to gradually lose the linear relationship with PR data, and the PMW rainfall estimates could be anywhere between 0.1 and 4.0 mm h⁻¹. Such discrepancies again reflect limitations in detecting light-rain events by current PMW retrievals over land.

Figure 10 shows PDFs of collocated raining samples over the TRMM Land for a period of 20 months. For rain intensities below 0.5 mm h⁻¹, all the imager and sounder data indicate similar underestimations in the incidence of light-rain events. Above 0.5 mm h⁻¹, TMI and AMSR-E retrievals tend to perform the best and nearly match with TRMM PR data. SSM/I and AMSU-B retrievals, however, indicate some overestimations in rain incidence between 1.0 and 10.0 mm h⁻¹, with AMSU-B showing the largest overestimations. These overestimations, as shown in plots for TRMM Ocean, could possibly result from aliasing errors by mistakenly identifying light-rain samples as intermediate-rain samples.

The normalized systematic biases over the TRMM Land (in Fig. 11), similar to those over the TRMM Ocean, continue to show increasingly negative biases for heavy-rain samples and increasingly positive biases for light-rain samples. Ferraro et al. (2005), in comparing AMSU rainfall retrievals versus a radar–rain gauge composite rainfall product for a landfalling tropical cyclone, also noticed that AMSU retrievals tend to overestimate the light rainfall and underestimate the heavier rainfall. For heavy-rain samples, the amplitudes of negative biases over land and ocean are similar, with TMI and AMSR-E even indicating smaller biases over land relative to their counterparts over ocean. Among all the land algorithms, the AMSR-E retrieval overall has the smallest systematic bias, and TMI the second smallest. Both SSM/I and AMSU-B indicate a relatively large systematic bias below 2.0 mm h⁻¹, in dramatic contrast to the generally small systematic bias shown in the ocean algorithms.

The normalized RMS error plot shows that rain estimates from the sounder and imager data have comparable RMS errors over land. Although TMI and AMSR-E still indicate slightly smaller RMS errors below 1.0 mm h⁻¹ than AMSU-B does, AMSU-B generally performs better than SSM/I and the overall difference between the sounder and imager data is much smaller over land than over ocean for light-rain events. Relative to SSM/I, AMSR-E, and TMI, AMSU-B data even show slightly smaller RMS errors for rain intensities above 2.0 mm h⁻¹. The average relative RMS error is between 200% and 250% when PR rain intensity is close to 1.0 mm h⁻¹ and between 50% and 80% when PR rain intensity is around 10 mm h⁻¹. Therefore, even though AMSU-B data are somewhat worse than the imager data over ocean, especially for rain intensities <5 mm h⁻¹, they are in general comparable in quality to those derived from conically scanning radiometers over land for instantaneous rain rates. If this is true, then cross-track microwave sounders with high-frequency channels on operational satellites could be included in future precipitation measurement constellations to achieve better sampling and coverage over land.

6. Comparisons with ground measurements over the continental United States

Although PR data are extensively used as a reference in this study to evaluate PMW imager and sounder data, they do have uncertainties and should not be treated as the truth. To further confirm the results obtained from coincident comparisons with TRMM PR over land, we have also performed similar coincident comparisons over the continental United States, using surface radar and rain gauge data as another independent reference to check the consistency of the result. Another purpose of this exercise is to demonstrate whether or not the theoretically superior TRMM PR data are indeed better than PMW rain retrievals over land. The ground validation data used in this study are the merged surface radar and rain gauge product from the National Centers for Environmental Prediction (NCEP) National Hourly Multisensor Precipitation Analysis Stage IV (Lin and Mitchell 2005). This dataset collects hourly radar rainfall estimates from about 140 Weather Surveillance Radar-1988 Doppler (WSR-88D) operational radars over the continental United States, merging with about 3000 hourly gauge reports. The stage IV data are preliminarily quality controlled and calibrated. The original precipitation product is at the 4-km resolution, and the data are horizontally averaged onto the same 0.25° × 0.25° grids as the satellite rain retrievals. Because the ground measurements are only available at 1-h intervals, the definition of coincidence has to be relaxed to 1 h in this section, but the temporal variance of instantaneous rain estimates versus hourly means will be examined. The coincident comparisons are made for a 3-month period (June–August 2005) over the continental United States south of 35°N.

Figure 12 compares the normalized systematic bias of collocated rain rates from TRMM PR and PMW sensors against surface radar and gauge data. All the satellite rainfall retrievals again indicate increasingly positive biases above 10 mm h⁻¹ and increasingly negative biases <5 mm h⁻¹. PR data indeed have the smallest systematic bias at light-rain and intermediate-rain intensities. Above 10 mm h⁻¹, however, PR data have systematic biases similar to those of the PMW retrievals. Rain retrievals from the sounder data have the largest systematic bias at light-rain intensities, consistent with the result using PR data as the reference. Above 3 mm h⁻¹, the sounder, imager, and PR data tend to have similar systematic biases.

Illustrated in Fig. 13 are the normalized STD errors of collocated rain rates from TRMM PR and PMW sensors against the merged surface radar and gauge product at 1-h intervals. The two black dashed lines represent the temporal variance of instantaneous rain rates compared to hourly mean rain rates based on surface radar data collected from two TRMM ground validation sites (Melbourne, Florida, and Houston, Texas, courtesy of David Wolf and Tom Bell). These idealized curves provide an estimation of the largest random STD error due to mismatches in observation times within an hourly window for perfect retrievals. Compared with all the PMW retrievals over land, PR rainfall estimations not only have the smallest systematic bias but also have the smallest STD error <5.0 mm h⁻¹, confirming the PR’s theoretical superiority relative to PMW retrievals in accurately retrieving rainfall over land, especially at light- and intermediate-rain events. Above 5.0 mm h⁻¹, both PR and PMW retrievals tend to gradually approach the idealized curves as rain intensity increases, suggesting that PR and PMW instantaneous rain retrievals can provide similarly accurate rainfall estimations for heavy-rain samples. Comparisons of STD errors from the sounder and imager data again indicate that AMSU-B data indeed have smaller STD errors than SSM/I, AMSR-E, and TMI retrievals for instantaneous rain rates between 1.0 and 10.0 mm h⁻¹. Therefore, using a completely different and independent reference, we demonstrate that the AMSU-B data are comparable in quality to those derived from conically scanning radiometers over land. In addition, TRMM PR data are indeed found to be superior to PMW rainfall estimations over land for rain intensities below roughly 5 mm h⁻¹. For heavy-rain events, PR data appear to perform equally well to the rainfall retrievals from the sounder and imager data.

7. Summary and conclusions

This study examines instantaneous rainfall estimates provided by the current generation of retrieval algorithms for PMW instruments using retrievals from the TRMM PR and merged surface radar/gauge measurements over the continental United States as references. The PMW sensors included in the comparison are three operational sounders (AMSU-B instruments on NOAA-15, -16, and -17 satellites) and five imagers (TRMM TMI; SSM/I instruments on DMSP F-13, -14, and -15 satellites; and AMSR-E on Aqua). The purpose is to quantitatively assess surface rain retrievals from cross-track scanning microwave humidity sounders relative to those from conically scanning microwave imagers. Coincident comparisons are conducted against TRMM PR by averaging instantaneous rainfall retrievals onto a 0.25° × 0.25° grid at 10-min intervals over a 20-month period from January 2005 to August 2006. Rain detection capabilities and retrieval error statistics are computed as a function of rain intensity over land and ocean. Additional coincident comparisons are also made against ground radar and gauge data over the continental United States at 1-h intervals over a 3-month period from June 2005 to August 2005.

Comparisons of systematic errors and RMS errors indicate that AMSU-B sounder rain estimates are comparable in quality to those from conically scanning radiometers for instantaneous rain rates between 1.0 and 10.0 mm h⁻¹. This result holds true for comparisons using either TRMM PR estimates over tropical land areas or merged ground radar–gauge measurements over the continental United States as the reference.

Over tropical and subtropical land, the sounder and imager data have similar detection capabilities, with the sounder data showing slightly better detection capabilities than the imager data. Although SSM/I and AMSU-B display a relatively larger systematic bias than TMI and AMSR-E <2.0 mm h⁻¹, all the PMW retrievals show similar incidence of light-rain events.

Over the tropical and subtropical ocean, the radiometer data generally have better detection capabilities than the sounder data, especially for light-rain events. Among the radiometer data, SSM/I has slightly worse detection capabilities than TMI and AMSR-E, largely due to the SSM/I’s coarser retrieval footprint resolution resulting from satellite altitudes and fewer microwave low-frequency channels. The AMSU-B data, which are less capable of detecting emission signatures of liquid phase hydrometeors due to the availability of only high-frequency channels, tend to miss a lot of warm-topped precipitation events that are found to be prevalent in the tropics and subtropics (e.g., Petty 1995, 1999; Johnson et al. 1999; Lau and Wu 2003; Huffman et al. 2007). Some of these lower detection capabilities are manifested by underestimations of the frequency of light-rain events and overestimations of the frequency of intermediate-rain events, which are likely to result from aliasing errors. Significant efforts are being made to merge low-frequency emission signals on the AMSU-A instruments with the high-frequency scattering signals on the AMSU-B instruments so that information about liquid-phase hydrometeors can be better detected over tropical and subtropical oceans (Vila et al. 2007).

All the PMW retrievals have similarly small systematic biases over the tropical and subtropical ocean relative to TRMM PR data, and the largest relative systematic bias is smaller than 20%. The standard deviation errors are comparable between imager and sounder retrievals for rain intensities >5 mm h⁻¹, below which the imagers are noticeably better than the sounders.

The results of this study suggest that cross-track scanning microwave humidity sounders on operational satellites can be used to augment conically scanning microwave radiometers to provide better temporal sampling over land in planning future satellite missions, such as the Global Precipitation Measurement (GPM) mission (A. Y. Hou et al. 2007, unpublished manuscript), to provide the next generation of global precipitation measurement from space.