1. Introduction

Despite the inherent weakness of the physical relationship between cloud-top temperature and the underlying rain rate, time-sequenced infrared (IR) imagery from space provides an excellent depiction of the movement of clouds and weather systems. Microwave measurements, by contrast, have more direct and physical connections, resulting in more accurate rain measurements by satellites. However, their less frequent time sampling limits the use of products derived from microwave-only datasets for daily or shorter time scales, particularly those associated with rapidly developing storms. Following the launch of the Tropical Rainfall Measuring Mission (TRMM), increased attention has been paid to developing blended-satellite, high-resolution precipitation products (HRPP) that capitalize on rapid-update visible/infrared (VIS/IR) observations available from geostationary-based imagers, while augmenting less frequent but more accurate microwave rainfall retrievals (Ebert et al. 2007).

More recently, new merged data with higher spatial and temporal resolutions (0.25° and 3-h scales, respectively) have become available [e.g., Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN; Sorooshian et al. 2000), Climate Prediction Center Morphing Method (CMORPH; Joyce et al. 2004), TRMM Multisatellite Precipitation Analysis (TMPA; Huffman et al. 2007), and the National Research Laboratory blended precipitation dataset (NRL-blended; Turk and Miller 2005)], referred to here as HRPPs. Although HRPP data are more often used in a wide range, it is essential to validate those products using common in situ measurements such as rain gauge data or radar observations across various climate regimes before applying them. In response to this need, the International Precipitation Working Group (IPWG; see http://www.isac.cnr.it/~ipwg/) began a validation program that led to a comprehensive global evaluation of HRPP named the Program to Evaluate High Resolution Precipitation Product (PEHRPP; Turk et al. 2008), along the lines of previous validation efforts such as the Precipitation Intercomparison Projects and the Algorithm Intercomparison Projects of the Global Precipitation Climatology Project (GPCP; Arkin and Xie 1994; Ebert et al. 1996). The aim of PEHRPP is to characterize errors in various HRPPs on different spatial and temporal scales in different climate regimes to help developers improve their accuracy and help users be aware of the limits of the products they use.

As part of the PEHRPP efforts, we intend to assess the performance of four HRPP products (PERSIANN, CMORPH, TMPA, and NRL-blended) during the summer over the Korean Peninsula, where summer monsoons and typhoons are two predominant rain-producing weather phenomena. This analysis uses rain gauge data from the dense, 1-min updating Korean Meteorological Administration (KMA) Automated Weather Station (AWS) network (see Fig. 1 for the location of about 520 gauges across South Korea). We consider KMA AWS gauge data dense enough to yield spatially homogeneous coverage for meaningful comparisons with satellite-based rain rates, and we attempt to validate the HRPPs. The AWS data are first collocated to HRPP data by time and location. We then construct a two-dimensional matrix of correlation, mean bias, and RMSE, which vary with rain averaging spatial and time scales. Validation is also conducted by examining how the diurnal cycle can be resolved, using here spherical harmonics analysis.

In this paper, precipitation, rainfall, and rain rate are used interchangeably, although the rain rate given in millimeters per hour is most appropriate. Descriptions of the data and associated algorithms are provided in section 2. Section 3 provides results of the geographic comparison, and section 4 deals with statistical analysis. Comparisons of diurnal variations are provided in section 5 and conclusions are drawn in section 6.

2. Datasets used

This analysis uses four HRPPs, the TMPA gauge-adjusted version (3B42), CMORPH, PERSIANN, and NRL-blended precipitation products, which offer a global rainfall map every 3 h with a 0.25° resolution. In all of the algorithms, microwave-based precipitation measurements are augmented by geostationary-based IR brightness measurements. A good summary of those algorithms and their inputs is found in Sapiano and Arkin (2009), and here we provide a brief description of the algorithms used.

In TMPA version-6 3B42 (here referred to as TMPA), window channel IR brightness temperatures (IR TBs) observed by geosynchronous weather satellites and the passive microwave (PMW) precipitation data from the TRMM Microwave Imager (TMI), Special Sensor Microwave Imager (SSM/I), Advanced Microwave Sounding Unit (AMSU), and Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E) measurements are used (Huffman et al. 2007). When PMW and IR estimates are combined by calibrating the IR measurements to yield overall consistent PMW–IR products, the TMPA algorithm takes the PMW estimate preferentially, and other empty grids are filled with IR estimates. Then, the combined rainfall estimates are rescaled against the GPCP monthly surface rain gauge data over the land (Huffman et al. 1995).

The CMORPH combines various PMW rain estimates (such as TMI, SSM/I, AMSU, and AMSR-E), but calibrates against TMI values. However, because of insufficient global coverage by microwave measurements at a 30-min time scale only, vast areas may have gaps where PMW estimates are not available. To fill these gaps, microwave-based rainfall values are interpolated with time according to the propagation of cloud systems obtained from geostationary IR-based motion vectors (Joyce et al. 2004).

The PERSIANN algorithm is based on a neural network approach (Hsu et al. 1997; Sorooshian et al. 2000) to determine rainfall from geostationary IR measurements. Rainfall is estimated by training the IR TB to the collocated PMW rain estimates. In this study, PERSIANN HRPP uses the TMI rain estimate to train the IR TB–rain-rate relationship over the west Pacific region (Sorooshian et al. 2000).

The NRL-blended algorithm involves locally based collocations of all microwave and IR pixels when PMW sensors cover any geostationary satellite observation area (Turk and Miller 2005). The acquired PMW estimates from TRMM and SSM/I are paired with IR TBs, yielding a probability matching of rainfall histograms, which can then be used to retrieve rain from IR TBs. When new PMW estimates are available, the probability matching is continuously updated.

The PEHRPP provides 3-h HRPP of TMPA, CMORPH, PERSIANN, and NRL-blended in a 0.25° × 0.25° grid format (available online at http://essic.umd.edu/~msapiano/PEHRPP/). The HRPP data were collected across South Korea for four summers [June, July, and August (JJA)] from 2003 to 2006. The summer period was selected because a large proportion of annual precipitation over South Korea occurs during the summer, in association with monsoon- and typhoon-related rain events.

Also included is the TMI rain rate, which was used as a reference for four HRPPs in this study. The TMI rain rates were estimated from TMI 1B11 brightness temperatures (downloaded from http://mirador.gsfc.nasa.gov/) by applying the Goddard profiling algorithm (GPROF), version 6 (Olson et al. 2006). The retrieved rain rates are averaged over the four summer seasons from 2003 to 2006 for each 0.25° × 0.25° grid.

Along with HRPP data, surface rain gauge measurements were obtained for the same four summer periods (2003–06) from the Korean AWS rain gauge network, operated by the KMA. The network consists of about 520 tipping-bucket rain gauges that update every minute with a measuring unit of 0.5 mm h⁻¹, with a spacing of about 15 km. Figure 1 depicts the geographic distribution of AWS gauge stations over South Korea. The rain rate (mm h⁻¹) is derived from accumulated AWS rainfall measurements by adopting the TRMM/Gauge Data Software Package algorithm (Wang et al. 2008), which converts the cumulative rainfall to rainfall intensity using a cubic spline function.

Gridding the gauge data as in HRPPs, an arithmetic average of gauge data was taken instead of the objective analysis since gauges are quite uniformly spaced and most of the 0.25° × 0.25° grids contain more than two gauges. However, different weights are applied, depending on the number of gauges available within the target grid. If the number of gauges (N) within a given 0.25° × 0.25° grid is equal to or greater than 3 (i.e., N ≥ 3), an average of gauge measurements with equal weights was taken for the rain rate for the grid. If the number of gauges within the target grid is less than 3 (N < 3), a search is done to find more gauges located closest to the center of the grid (i.e., three, four, and five more gauges located in N = 2, 1, 0, respectively). Once gauges are chosen, different weights are applied to obtain the grid value. While a unit weight is applied for the gauges located within the 0.25° × 0.25° grid, weights of 0.3, 0.5, and 0.5 are applied to the three, four, and five gauges located outside the target grid for N = 2, 1, 0, respectively. Then, the rain rate for each grid is obtained by normalizing the sum of the weighted rain rate by the total weight.

To compare HRPP data with rain gauge estimates, the same time-averaging window between HRPP and AWS is necessary. Because the 3-h rain rate of HRPP at a given time t represents a 3-h window from t − 1.5 to t + 1.5 h for TMPA and from t to t + 3 h for others (CMORPH, PERSIANN, and NRL-blended), AWS data are processed in time to be compatible with each HRPP.

3. Geographic rain distribution

Before statistically analyzing the quality of HRPP products, we examine how the summer mean rainfall climatologies of four HRPPs and TMI rain rate compare with AWS distributions. Figure 2 presents the mean rain rates (mm h⁻¹) averaged over four summer seasons (2003–06) for each 0.25° × 0.25° grid point.

In the AWS rain gauge measurements, a heavy rainfall area showing a rain rate up to 5 mm h⁻¹ extends from the central south to the northeastern part of South Korea. The east and west coast areas show relatively low rainfall. Isolated heavy rainfall is noted on Jeju Island located in the bottom of the analysis domain. Such distributions are well reproduced by all HRPPs except PERSIANN, which shows a blurred and smooth rainfall pattern. When compared with the AWS, a broader heavy rainfall distribution is noted in TMPA with an overestimated magnitude over most of South Korea. The other three HRPPs (CMORPH, PERSIANN, and NRL-blended) show comparatively weak rain distributions throughout South Korea. Considering that TMPA PMW rain estimates are rescaled against the GPCP monthly rain gauge data over land, the closer agreement between AWS and TMPA may be due to the inclusion of gauge information in TMPA.

The summer mean distribution of PMW-based TMI rain products is also provided in Fig. 2f. This is because HRPP algorithms are primarily tuned against the TMI retrievals for constructing 3-hourly data. Relative to the AWS rain distributions, the TMI product shows significant underestimates over the peninsula. The GPROF algorithm used for TMI products takes the Bayesian approach supported by six different cloud simulations (Olson et al. 2006). However, considering that those six simulations focus mainly on tropics and ocean cases, the midlatitude cases currently in question may not be well represented by the GRPOF database, resulting in biased results, and raising a question about the use of the GPROF database as an initial input to the PMW algorithm over the Korean Peninsula. Indeed, it is not surprising that we observe underestimates by CMORPH, PERSIANN, and NRL-blended because of the tuning against underestimated TMI data over the analysis domain.

4. Statistical behavior

For the grid-to-grid comparison between the AWS and HRPP products, scatter diagrams for the hourly rain rate obtained from AWS gauge data and four HRPPs are given in Fig. 3, with associated statistics provided in Table 1. In Fig. 3, the TMPA shows a widely scattered pattern, with a correlation coefficient (R) of 0.47 and a root-mean-square error of 3.69 mm h⁻¹. The RMSE for TMPA appears to be larger than the other three products. The large RMSE for TMPA seems to be due to the fact that TMPA often strongly overestimates rain rates, especially for light or moderate rain events. This is also clear in the averages map of Fig. 2, in which TMPA is generally higher than the AWS measurement and much higher than the other three HRPP products.

CMORPH shows the highest correlation (R = 0.58) among the four HRPPs and a relatively small RMSE of 1.40 mm h⁻¹. PERSIANN and NRL-blended look quite similar to each other, with correlation coefficients of 0.38 for both and not very discernable RMSE and mean bias. PERSIANN rain rates seem to have an upper limit of less than 20 mm h⁻¹, resulting in some considerable underestimates, particularly for heavy rain events. Considering the smaller dynamic range of the PERSIANN rain rates, it is obvious that PERSIANN is less accurate at retrieving rainfall for heavy rain events over South Korea. The smallest RMSE, around 0.87 mm h⁻¹, also appears to be due to a smaller dynamic range.

In summary, the climatological distributions of rain rate and statistical behaviors suggest that there is a general difference between TMPA and the other three HRPPs. Although all HRPPs underestimate rain rates for heavy rain events over South Korea, the underestimation is substantial, particularly for CMORPH, PERSIANN, and NRL-blended. This underestimate is not surprising because most HRPP products are calibrated against microwave measurements, which show significant underestimates over South Korea, as in the TMI estimate (Fig. 2f). It is clear that the gauge-adjusted TMPA generally overestimates rain rates for light and moderate rain cases. Normalized cumulative frequency distributions of AWS gauge data and four HRPPs are shown in Fig. 4. About 96% of TMPA and 98% of AWS rainfall data fall within 10 mm h⁻¹, while less than 0.5% of CMORPH, PERSIANN, and NRL-blended rainfall is heavier than 10 mm h⁻¹. The overestimation by TMPA is indeed due to the adjustment of the multisatellite product to the rain gauge measurement. A more detailed explanation of the possible overestimation of TMPA and its consequence in rainfall monitoring is discussed later in this section. By contrast, underestimations of CMORPH, PERSIANN, and NRL-blended are unavoidable because the GPROF-based rain retrieval algorithm for the microwave measurement over land produces the significant underestimate, as shown in Fig. 2f, at least over the Korean Peninsula.

The scatterplots and associated statistics are based on the use of rain gauge data as true rain events. Thus, the statistics obtained from scatterplots may not represent the performance of satellite algorithms if there are false alarm events by satellite estimates. Because of this, we examine how well HRPP products detect rain. Following the terminology adopted by Ebert et al. (2007), each grid box can be classified as a hit (H: satellite method correctly detects the observed rain), miss (M: satellite fails to detect the observed rain), false alarm (F: detected rain but not observed), and null (no rain detected and no rain observed). The false-alarm ratio (FAR) is defined as FAR = F/(H + F), and the probability of detection (POD) is defined as POD = H/(H + M). To avoid obscure rain events such as intermittent events, a threshold of 1 mm day⁻¹ is applied to discriminate between rain and no rain.

Obtained error statistics for the summers of 2003–06 are given in Table 2. Results indicate that CMORPH has the highest POD (about 68%), whereas TMPA and NRL-blended have the lowest (about 45%). The lowest FAR is found for the TMPA data, probably due to the smaller POD. In other words, TMPA detects less and fails less. This characteristic is also found in the NRL-blended data. We calculate a relatively smaller FAR of around 34% for CMORPH. Among the four products, PERSIANN shows the highest rate of false alarms, about 53% (Fig. 4).

We further examined how the POD varies with the rain intensity (Fig. 5). In this calculation, gauge-estimated rain rates are binned with a 1 mm h⁻¹ interval and the POD is calculated for each rain bin. As expected, lowest PODs are found for rain rates less than 1 mm h⁻¹. The lowest 29% of POD is found for TMPA, while 52% is found for CMORPH when the rain rates are lower than 1 mm h⁻¹. The NRL-blended and PERSIANN data show 32% and 44% of POD, respectively. Of the four algorithms, CMORPH demonstrates a superior detection capability, in particular when the rain rates are weak or moderate, less than 10 mm h⁻¹. We note that PERSIANN shows the second highest capability for detecting rain but the highest FAR, suggesting that PERSIANN simply detects too much rain area. This may explain the lowest correlation coefficient noted in Table 1.

Better agreement can be reached between surface in situ measurements and satellite-estimated rainfall if a longer time, a wider area, or both are used in the rain average. Thus, we examine the statistical behaviors of the four HRPPs over different spatial and temporal scales. This analysis was done in a two-dimensional space and time domain by changing the grid size from 0.25° to 1.5° for the areal average, with an increment of 0.25°, and varying the time-average window from 3 h to 4 days.

Figures 6 –8 show the correlation, mean bias, and RMSE from the analysis for varying spatial and temporal scales in a two-dimensional frame, in which the x and y axes represent the spatial scale and averaging time period, respectively. As expected, all three statistical parameters improve as either the averaging period is increased or the grid size becomes larger, indicating that random errors of gauge measurements and satellite-derived products cancel each other out. The contours for correlation do not flow smoothly at the longer time scale because of the relatively small number of data points available from the full 12-month period. Likewise, the small size of South Korea (maximum 2.5°) produces a small number of data points when the data are averaged over the coarsest spatial scales.

The correlation coefficients for all four HRPPs begin to fall off quickly once the time average drops below a certain period (depending on the individual algorithm) and/or the spatial scale falls under a certain scale (also depending on the individual algorithm). This suggests that certain averaging scales are needed to acquire reasonable statistics (or accuracy). For example, to obtain a correlation coefficient of 0.8, the TMPA data at a 1.0° grid scale should be averaged at least over a 9-h time scale, and at least in a 1.5-day averaging period for 0.5° gridded data. For the CMORPH, at least 7- and 18-h time averages are appropriate for 1.0° and 0.5° grid data, respectively, to obtain the same correlation coefficient. A better correlation for CMORPH is consistent with results from the grid-to-grid comparison in Fig. 3. Relatively weaker correlation coefficients are found for both PERSIANN and NRL-blended. For example, PERSIANN HRPP needs a 2.5-day time average at a 1.0° grid scale. NRL-blended in particular shows a much more irregular contour pattern, partly due to missing NRL-blended data points during the analysis period.

The mean bias pattern of TMPA is quite different from that of the other three products (CMORPH, PERSIANN, and NRL-blended). Contour lines for CMORPH, PERSIANN, and NRL-blended within a range shorter than 1 day and smaller than 1° are dense, indicating a rapid increase of mean biases within this range, whereas TMPA shows much smaller and more homogeneous biases. The smaller bias shown in TMPA may be due to the uniqueness of its gauge correction applied over land. The other three products show a general decrease of mean bias with increasing averaging temporal and spatial scales.

A similar type of decrease in performance is also evident in the RMSE (Fig. 8). TMPA shows the largest RMSE when the spatial and time scales fall under 1° and 1 day, respectively. The other three products are not very different from each other; PERSIANN shows the lowest RMSE, but this may not be very meaningful because of the low dynamic range. Considering that satellite estimates are extended from a few snapshots of rain events during satellite overpasses, the gauge correction applied to TMPA may be more prone to a larger RMSE because the integration of these snapshots is never the same as continuous records of rain events. By contrast, similar patterns in CMORPH, PERSIANN, and NRL-blended may be due to the fact that they all use the same type of inputs, based on PMW instantaneous rain estimates.

Since all products, including the real-time TMPA (3B42RT; before the gauge adjustment), are tuned against PMW results, rain-rate estimates by microwave measurements are crucial to determine the HRPP quality. However, as shown in Fig. 2f, microwave-based rain rates (TMI) based on the 85-GHz scattering algorithm (McCollum and Ferraro 2003) show significant underestimates relative to the AWS rain rate. Thus, the correction of real-time TMPA products to gauge values seems to result in the larger difference between TMPA and the other datasets.

Nevertheless, we examine how the TMPA correction influences the data quality. Figure 9 gives scatterplots showing the difference before (3B42RT) and after the adjustment (3B42) and their associated distributions of normalized frequency for July 2005–06. Correlation coefficients for both products are similar (0.51 versus 0.53); however, the monthly rain gauge adjustment induced a much smaller mean bias (from −0.55 to 0.03) and higher RMSE errors (from 2.31 to 3.35) (Table 3). The geographic distribution of the summer mean TMPA shows rainfall amounts comparable to gauge measurements (Figs. 2a,b), probably due to the adjustment of TMPA 3B42RT monthly products to monthly gauge measurements. However, the adjustment seems to have more impact on light and moderate rain events because gauge-measured heavy rainfall events are not well depicted by TMPA 3B42RT, which shows a significant underestimate, whereas heavy rainfall events by TMPA are generally shown as moderate by AWS. When the obtained ratio between the monthly mean 3B42RT rain rate and GPCP surface rain gauge data is equally applied to the 3-hourly rainfall estimate (Huffman et al. 2007), weak–moderate rain events become far heavier rain events, as shown in Fig. 9b. By contrast, the adjustment of heavy rain events is minimal. It appears that the monthly gauge correction improves the bias at the expense of increased RMSE, with little improved correlation. This comparison further suggests that the monthly gauge adjustment works best if there is a strong linear relationship between the satellite and the gauge before the adjustment (or when the correlation is high).

The satellite-derived rainfall products at finer scales are ideal and valuable in that the rainfall behaviors associated with mesoscale convective systems can be resolved. The statistical characteristics shown in Figs. 6 –8 suggest that despite the usefulness of high-resolution rain products, users should pay attention to both their benefits and their limitations. Satellite products still require improvement. Averaging in the coarser time and space domains may improve the agreement between satellite estimates with gauge measurements, as demonstrated in this section; however, this coarser resolution limits studies of detailed features associated with rain events. This study demonstrates that satellite data currently involve a trade-off between resolution and quality.

5. Comparison of diurnal variations

The 3-h HRPPs offer an opportunity to examine time variations of rainfall quantities, such as a time series of rain fluctuations and diurnal variations. These kinds of data were not readily available from other coarse-resolution datasets because of the lower coverage of subdaily time-scale signals. Comparing the diurnal cycle of rainfall is more intriguing because it is poorly understood in the atmospheric sciences and poorly represented by numerical models (e.g., Betts and Jakob 2002; Nesbitt and Zipser 2003). Here, we first examine the summer mean diurnal variations over South Korea, and then the decomposed daily cycle at the individual grid points.

Figure 10 shows the diurnal variations of the summer-mean rain rates averaged over South Korea. The AWS rain gauge measurements indicate that a rainfall peak appears in the early morning around 0600 LT, and then a minimum appears around 2100 LT. All of the HRPPs seem to capture the peak with a time lag, giving rise to a peak between 0600 and 0900 LT, although NRL-blended shows a blurred peak. All HRPP products also show a time lag at the minimum location. There seems to be a 1.5-h time lag between TMPA and the other three HRPP products, but this is likely due to the different time windows used for the 3-hourly average. TMPA used a time window 1.5 h off from the other three HRPP products. Of the four products, TMPA shows the largest amplitude of diurnal variation, probably due to the gauge adjustment, and CMORPH shows a surprisingly similar phase to AWS. PERSIANN and NRL-blended show much weaker diurnal variations.

We further analyze the diurnal cycle of each rain product using harmonic analysis at each grid point (Fig. 11). The arrow magnitude and direction in Fig. 11 represent the amplitude and the peak time in the form of a local 24-h clock dial, respectively. The clockwise rotation of an arrow from 0° to 360° means the change of the peak time from 0 to 24 h, such that up-, right-, down-, and left-pointing arrows denote maxima at 0000 (midnight), 0600 (dawn), 1200 (noon), and 1800 LT (sunset), respectively.

In the AWS, relatively large diurnal variations of rainfall are observed over the southern coastal area, with a peak around 0900 LT, somewhat later than the typical open ocean (Yang and Smith 2006), and much earlier than the convective land type (Chung et al. 2007). In fact, in the middle of the peninsula along 35° and 37°N, rainfall peaks are generally between 0600 and 0900 LT, but with a weak signal.

In TMPA, peaks showing diurnal variations generally agree well with AWS. Unlike the well-matched peak times, amplitudes are significantly larger than AWS. Over most of the peninsula, weaker but similar patterns are found in CMORPH, including the 0900 LT peak over the southern coastal area. The smaller magnitudes of the diurnal cycle of CMORPH are again likely due to underestimates of the rain rate by the PMW algorithm over the peninsula. To improve the HRPP algorithms without adjusting to the gauge values as in TMPA, more accurate measurements by PMW algorithms appear to be necessary.

6. Conclusions and discussion

As a part of PEHRPP activities, we examine the qualities of four independently produced HRPPs (TMPA, CMORPH, PERSIANN, and NRL-blended) over South Korea. We use 1-min surface rain measurements taken by gauges from the Automated Weather Station (AWS) network over South Korea during four summer months (JJA) from 2003 to 2006. The AWS dataset is unique because it includes 520 rain gauges with about a 15-km spacing. Quality assessment of the four HRPPs was done by generating statistics: first with a 0.25° grid and 3-h time scale and then at varying spatial and temporal scales. Diurnal variations of HRPPs were also examined for how well HRPPs reproduce the diurnal variations over the Korean Peninsula.

We show that general patterns in the monthly means and statistics of CMORPH, PERSIANN, and NRL-blended are alike in many respects. This is because CMORPH, PERSIANN, and NRL-blended use PMW-estimated rainfall as a reference and the PMW method seems to substantially underestimate the rain rate over the Korean Peninsula, as shown in the TMI rain rate compared with AWS. For better estimates by these methods, more accurate PMW measurements (in particular via an improved land algorithm) seem to be necessary.

Unlike those three products, the gauge-adjusted TMPA 3B42 seems to have less bias and shows a similar pattern to the climatology. However, the adjustment results in increased RMSE. We suggest that TMPA works best when correlation between preadjusted values and gauge measurements is high because adjustments can be made homogeneously throughout the rainfall range. PERSIANN performs poorly, suggesting that a globally or regionally trained algorithm may not work properly in some local areas; the Korean Peninsula may be such a case. This further suggests that a locally based training algorithm may be necessary to gain better agreement. CMORPH seems to reproduce the rain distribution expected from PMW-based rain estimates and shows the highest correlation and probability of rain detection, compared to the gauge measurements.

We here summarize how the results obtained over the Korean Peninsula compare with those obtained for PEHRPP sites around the world. Satellite methods seem to better estimate convective rain during the summer. Ebert et al. (2007) assessed various HRPP data over Australia and showed that satellite rain estimates are relatively better in both the tropics and midlatitude regimes, whereas significant underestimates occur during the winter because of the difficulty in detecting nonconvective rain. Similar features were also found in the United States, with better agreement during the summer and difficulty in detecting winter precipitation (Gottschalck et al. 2005; Ebert et al. 2007; Tian et al. 2007; Sapiano and Arkin 2009). Difficulties in detecting light rain are also reported over the United Kingdom (Ebert et al. 2007). Over the tropical Pacific Ocean, however, the rainfall appears to be underestimated by satellites, compared to moored buoy rain gauge data (Sapiano and Arkin 2009). Over the Japanese islands, which are near the current analysis area, HRPPs appear to perform poorly for both light rain and heavy rain events during the warm season, resulting in underestimates by satellite-only methods (Kubota et al. 2009). Although the current study examines only the summer rainfall, our results are consistent with those from Japan, suggesting that the quality of satellite-determined rainfall is highly dependent on the rainfall regime and/or the season. Considering that all of the IR PMW algorithms except the gauge-combined TMPA method discussed in this study tune the IR algorithms to the PMW estimate, the significant underestimates found in the three satellite-only methods (CMORPH, PERSIANN, and NRL-blended) appear to be mainly due to the deficiencies of PMW rain retrievals, particularly over land. Nevertheless, the diurnal variations are fairly well resolved by satellite methods (Tian et al. 2007; Sapiano and Arkin 2009; Kubota et al. 2009), as also demonstrated in this study.