1. Introduction
Tropical precipitation plays an important role in the global freshwater balance and is sensitive to large-scale disturbances in the atmosphere and the oceans. In addition, accurate observation of tropical precipitation is crucial for monitoring anomalies such as El Niño–Southern Oscillation (ENSO), which has global impacts (New et al. 2001).
Monsoon systems are another important climatic feature that can be monitored with global precipitation datasets. Such systems transport moisture and precipitation from the oceans toward the continents (Dobler and Ahrens 2010). Because of the lack of ground-based observations over the ocean and other isolated parts of the world, satellite-based and reanalysis datasets are an important data source. The high spatiotemporal variability of tropical precipitation complicates the evaluation of precipitation products in the analyzed region.
The usage and diversity of satellite-based precipitation datasets has grown in recent years (Kidd and Huffman 2011; Kidd and Levizzani 2011). The availability of long time series of data together with good data homogeneity is essential for the analysis of climate and its variability. The lifetimes of satellites are limited, which make changes in the observing system within a data time series unavoidable. Moreover, the number of satellite sensors used to generate time series of data is not fixed, influencing temporal coverage. Since satellite data are also widely assimilated into atmospheric reanalysis, it is not safe to assume that both satellite-based and reanalysis data will be homogeneous a priori. Precipitation data from satellites face problems owing to limited temporal sampling, especially when low-Earth-orbiting satellite platforms are used, as sites close to the equator are scanned only once or twice per day. Precipitation datasets are usually based on measurements by passive microwave sensors, since they are able to detect precipitation-sized particles in a relatively direct manner. Emission by liquid drops and scattering on ice particles are signals in the microwave spectrum that were proved to be convertible into precipitation amounts (Wilheit et al. 1977; Ferraro et al. 1996; Ferraro 1997). One prominent, widely used series of microwave sensors is the Special Sensor Microwave Imager (SSM/I) instruments flying on board the Defense Meteorological Satellite Program satellite series (http://www.ngdc.noaa.gov/dmsp/). The precipitation datasets evaluated in this study make use of SSM/I data, either directly or through data assimilation, as in the case of reanalysis products.
Independent reference data used to evaluate satellite estimates and numerical models are rare, especially over the ocean, where the major portion of global rainfall takes place. Whereas ship measurements may provide data only along the ship route, buoy rain gauge measurements deliver time series from fixed positions over the ocean. Tropical buoy rain gauge data have been used to validate satellite-derived precipitation datasets over tropical and subtropical oceans (Bowman et al. 2003; Bowman 2005). Bowman et al. (2009) compared long-term averages of buoy and satellite data during the time period 1997–2006 and found biases of up to 25%. An evaluation of the Hamburg Ocean Atmosphere Parameter and Fluxes from Satellite Data (HOAPS) shows that its mean precipitation amounts in the western tropical Pacific Ocean are lower than those in the current interim reanalysis of the European Centre for Medium-Range Weather Forecasts (ERA-Interim) and higher than those in the Global Precipitation Climatology Project (GPCP) dataset (Andersson et al. 2011). A comparison of HOAPS precipitation data with instantaneous ship measurements has proven that the HOAPS algorithm can detect small-scale convective rainfall reasonably well (Klepp et al. 2003). The ability of HOAPS to monitor monthly precipitation will be analyzed in this study.
Another oceanic precipitation study focuses on intercomparisons between satellite-based and reanalysis-based precipitation datasets. This results in relatively good agreement over the ocean concerning variability patterns but substantial disagreement in intensities (Shin et al. 2011). Tian and Peters-Lidard (2010) analyzed various multisatellite-based precipitation datasets in a 2-yr period and showed that relative uncertainties can be considered small over tropical oceans. A study of seasonal and interannual variations found that the western South Pacific is a region that exhibits large differences among all analyzed data, including data from the GPCP and HOAPS. Precipitation products using a composite of various data sources perform best (Béranger et al. 2006). Adler et al. (2012) estimated the climatological bias of the GPCP to be from −10% to −15% in the tropical western Pacific. Data from the GPCP have been validated with the Comprehensive Pacific Rainfall Database (PACRAIN) atoll station data, which are also used as reference data in this study. The result of this validation was a negative bias of −16% during a time period from 1979 to 2001 (Adler et al. 2003). One year of PACRAIN atoll station data have also been used within the Third Precipitation Intercomparison Project. Results revealed that the temporal coverage of polar-orbiting satellites seems to be a limiting factor for the correlation and that atmospheric models are less accurate than satellite products (Adler et al. 2001).
In precipitation validation studies, the PACRAIN database has received relatively less attention because the influence of the atoll landmass on oceanic precipitation was seen with skepticism. A recent study on rain on small tropical islands indicates that the influence of atoll-sized islands on precipitation is negligibly small (Sobel et al. 2011), motivating our use of PACRAIN as a reliable reference.
The objective of the present study is to evaluate the ability of satellite-based and reanalysis precipitation datasets to observe regional variability and anomalies on a monthly time scale. The evaluation takes place in a climatic zone that is characterized by heavy convection-induced precipitation events, contributing disproportionately to global precipitation amounts. Hence, tropical precipitation plays an important role in the global freshwater balance. Further, precipitation in the tropical Pacific is sensitive to atmospheric and oceanic disturbances (e.g., from ENSO), which have large-scale impacts on precipitation amounts and distribution patterns. The high temporal and spatial variability of tropical precipitation poses a major challenge for precipitation datasets.
In this study, satellite-based precipitation data from GPCP and HOAPS and reanalysis-derived precipitation data from ERA-Interim and from the National Aeronautics and Space Administration (NASA) Modern-Era Retrospective Analysis for Research and Applications (MERRA) are evaluated. As reference data, in situ rain gauge measurements on atoll stations extracted from the PACRAIN database (Greene et al. 2008) are used. The validation covers a variety of data sources ranging from those adopting the single-sensor approach (HOAPS) to those based on the multisatellite–multisensor approach (GPCP) to reanalysis products. The latter assimilates various satellite data and generates precipitation data that are based on numerical model parameterizations. The time period of validation is between 1989 and 2005, according to the overlapping period of the analyzed data.
2. Data
a. GPCP V2.2
The GPCP is part of the World Climate Research Programme and of the Global Energy and Water Cycle Experiment (GEWEX). It provides global estimates of monthly precipitation using a multisatellite approach; that is, it combines the data from various satellites into a final merged precipitation product. The GPCP data used in this study are from version 2.2 of the monthly precipitation analysis, with a spatial resolution of 2.5° (latitude) × 2.5° (longitude). Over-ocean GPCP data incorporate precipitation estimates based on different geostationary infrared (IR) sensors operated by the United States, Europe, and Japan; however, for filling gaps, passive Earth-orbiting microwave sensors (e.g., SSM/I) and sounders [e.g., Television and Infrared Observation Satellite (TIROS) Operational Vertical Sounder (TOVS)] are also used. The temporal coverage of GPCP data is improved by integrating multiple satellite data sources rather than single-sensor-based datasets. A detailed description of the data basis included in the GPCP dataset and a description of the merging and combining methods used can be found in Adler et al. (2003), Huffman et al. (2009), and in the dataset documentation available online (
b. HOAPS
HOAPS delivers precipitation and evaporation data over global, ice-free oceans (Andersson et al. 2011). The approach of HOAPS differs from that of the GPCP, because solely intercalibrated SSM/I data are used as a basis for deriving the HOAPS parameters. The algorithm used to estimate precipitation is based on a neural network, which was trained with SSM/I brightness temperatures and dedicated European Centre for Medium-Range Weather Forecasts (ECMWF) precipitation data (Andersson et al. 2010). The data used in this study are the HOAPS-G, version 3, precipitation data, which are delivered on a monthly time scale with a spatial resolution of 0.5° (latitude) × 0.5° (longitude) (obtained online from http://www.hoaps.org).
c. ERA-Interim reanalysis by ECMWF
The ERA-Interim reanalysis is the latest global atmospheric reanalysis operated by ECMWF (obtained at http://www.ecmwf.int). ERA-Interim makes use of an extensive data assimilation effort. It uses a wide variety of available observations, including satellite data, to obtain a global state of the atmosphere (Dee et al. 2011). The analyzed monthly precipitation is calculated based on 12–24-h forecasts within the ERA-Interim model system. Assimilated satellite data, including SSM/I, act as an important input for water vapor profiles and, thus, implicitly influence precipitation forecasts. The native grid resolution of ERA-Interim is T255, which corresponds to about 0.75° in latitude and longitude.
d. MERRA
MERRA is a recently launched NASA reanalysis of the atmosphere for the era of satellite observations (Rienecker et al. 2011). A state-of-the-art data assimilation system was developed and applied by the Global Modeling and Assimilation Office to synthesize various observations. MERRA covers the time period from 1979 onward and has a special focus on the analysis of the hydrological cycle. MERRA is based on the Goddard Earth Observing System (GEOS-5) general circulation model and offers a native spatial resolution of ½° (latitude) × ⅔° (longitude). The data assimilation system also integrates rain-rate estimates from passive microwave measurements. The MERRA variable used is the diagnostics of total precipitation, time averaged at surface level. The MERRA monthly mean total precipitation was downloaded via the MDISC Data Subsetter (http://disc.gsfc.nasa.gov). The product used is the monthly incremental analysis update (IAU) 2D land surface diagnostics.
e. PACRAIN
PACRAIN was developed under a research grant from the National Oceanic and Atmospheric Administration. It collects ground-based rainfall data in a data-poor region of the world—the tropical Pacific (see http://pacrain.evac.ou.edu/ for database information and data access). The input data to PACRAIN consist of daily and monthly rain gauge data from inland, coastal, island, and atoll stations, and the database is updated monthly. The data undergo a rigorous quality control process. Detailed information about the PACRAIN database is given in Greene et al. (2008). We used a subsample of the atoll stations in this study (cf. Fig. 1). The PACRAIN atoll station data used are not included in any of the evaluated satellite or reanalysis datasets, which makes it an independent validation dataset. Atoll stations are fairly representative of open-ocean conditions owing to their small size and flat orography (Sobel et al. 2011). The atolls used are of a size assumed not to significantly influence relevant satellite precipitation algorithms.
PACRAIN atoll stations and their average precipitation (mm day−1; color bar) during 1989–2005, and grid boxes (gray) showing the gridded PACRAIN data.
Citation: Journal of Applied Meteorology and Climatology 52, 3; 10.1175/JAMC-D-12-049.1
3. Methods
Daily PACRAIN atoll station data have been used to calculate mean precipitation for each month, if at least 25 days per month were available. Finally, 34 atoll stations offering data time series as complete as possible have been chosen manually (see Fig. 1). The evaluation in this study takes place on a monthly time scale. PACRAIN monthly means are compared with the monthly means of the satellite and reanalysis datasets as given by the data provider. Because the time series of PACRAIN stations vary in terms of their completeness, this ensures common evaluation time periods. It should be noted that the more inhomogeneous a variable is in space and time, the larger the deviations between the gridded data and the point data will be (Ahrens and Beck 2008), since gridded data represent grid-box area-averaged data and the real small-scale spatial distribution is unknown. To counteract this issue of data comparability, all data are analyzed in terms of monthly precipitation averages. For monthly PACRAIN atoll station data, Morrissey (1991) found correlation lengths of hundreds of kilometers. Additionally, spatial averaging is applied by gridding PACRAIN atoll stations, where the station density is sufficient.
The evaluation is applied in four steps. First, satellite and reanalysis data are compared to the respective PACRAIN atoll stations located inside the grid box. Thereby satellite and reanalysis datasets are kept in their native spatial resolutions (cf. Table 1). This evaluation gives information about the ability of the dataset to reproduce monthly precipitation amounts on local scales. Overall, 4682 months distributed over 34 PACRAIN stations and spanning the time period from 1989 to 2005 were used for this part of the precipitation evaluation. The average monthly precipitation of the PACRAIN stations used is 6.6 mm day−1. In the second step, HOAPS, ERA-Interim, and MERRA data are interpolated onto a common regular 2.5° latitude–longitude grid with first-order conservative remapping (Jones 1999) applied with the help of a Climate Data Operator (https://code.zmaw.de/projects/cdo) to analyze the effects of the different spatial resolutions of the dataset.
Native spatial resolutions of analyzed datasets.
In the third step, spatial interpolation of atoll stations to grid boxes of 2.5° latitude–longitude resolution is applied to account for the quite inhomogeneous spatial distribution of precipitation in the tropics. Therefore, an inverse-distance-weighted interpolation is used when at least two stations are available close (1.5°) to the grid-box center. Nine grid boxes fulfilled this condition and were used to perform an evaluation on individual common 2.5° latitude–longitude grid boxes (cf. Fig. 1). Overall, 1154 observations are used in the last part of the evaluation. Hence, there is a reduced data basis for this validation step.
Finally, to analyze the evaluation of the data, various statistical measures and scatterplots are used. Thereby, focus is placed on Pearson’s correlation coefficient, referred to as correlation hereafter, and on the median of absolute deviations (MAD) between datasets and the PACRAIN reference. The measure of absolute deviations is highly relevant, especially when evaluating climate data and its anomalies. Data offering strong variability, such as precipitation, cannot be evaluated by using the bias alone, which might deliver misleading results as errors can cancel out. Hence, whenever the bias is used, scientists should at least deliver another measure—the average absolute deviation—as an add-on. The interpretation of the MAD for the user is simple since deviations are treated linearly. Confidence intervals are added to the measures of correlation of absolute values and anomalies. These intervals help us to interpret the results as they support statements of correlations being significantly different or not. Additionally, to measure the scatter of deviations, the interquartile range of MAD is used.
To evaluate possible systematic under- or overestimations in different ranges of precipitation amounts, we used a piecewise linear regression model to visualize existing relationships. This segmented linear regression model is able to estimate slopes and multiple breakpoints (Muggeo 2008). The applied method enables a universal regression that is not dependent on specific fitting functions (e.g., linear function). Hence, it performs well for nonlinear distributions whereby it would lead to a linear regression if the scattering would exhibit a linear pattern of behavior (see Fig. 3, top right). The applied method is therefore more sophisticated and preferable compared to simple linear regression when nonlinear behavior is or might be apparent. The regression model was set to fit the regression using up to three breakpoints between the piecewise linear regressions. Because of the large scatter of data, the regression model is fit after some smoothing of the scatterplots by using a least squares method (Cleveland 1981).
4. Evaluation of precipitation data
a. Evaluation of datasets in native resolution with PACRAIN stations
As described in section 3, datasets are kept at their native resolutions (see Table 1) during this part of the evaluation process. Dataset grid boxes are compared with the respective PACRAIN atoll stations located within the grid box. In general, the evaluation is performed on monthly data.
Results of this evaluation are shown in Fig. 2. In general, it is evident that all analyzed precipitation datasets exhibit a large scatter with reference to the PACRAIN atoll stations. In Table 2, additional statistical measures are shown. The mean correlation ranges between r = 0.70 and r = 0.77. MAD values range from 1.75 to about 2 mm day−1, which stand for relative MAD values of about 25%–30% (cf. Table 2). The correlation between HOAPS and PACRAIN is lowest with r = 0.7, while the segmented regression line is closest to the one-to-one line over the full range of precipitation amounts (see Fig. 2, top right). Only a small systematic underestimation of high monthly precipitation amounts can be found in the HOAPS data, whereas both low and medium precipitation amounts almost match the one-to-one line. GPCP offers the highest correlation (0.77) and the smallest monthly MAD (1.75 mm day−1) with respect to PACRAIN station data, whereas higher precipitation amounts are somewhat more strongly underestimated compared to HOAPS (see Fig. 2, top left). The correlations for the reanalysis datasets of ERA-Interim and MERRA rank in between GPCP and HOAPS, but similar systematics are exhibited: small and medium precipitation amounts are overestimated, whereas high amounts are underestimated (see regression lines in Fig. 2, bottom left and bottom right). The large scatter of absolute monthly precipitation with reference to PACRAIN station data is reflected in the large interquartile ranges (IQRs) in the range of 2.6–3.2 mm day−1. HOAPS thereby has the highest IQR and GPCP the lowest (cf. Table 2). The satellite products of GPCP and HOAPS show negative biases of about 10%, whereas ERA-Interim has a positive bias and MERRA has almost no bias to PACRAIN.
Scatterplots of monthly absolute precipitation of (top left) GPCP, (top right) HOAPS, (bottom left) ERA-Interim, and (bottom right) MERRA against PACRAIN stations in 1989–2005, with a segmented regression line.
Citation: Journal of Applied Meteorology and Climatology 52, 3; 10.1175/JAMC-D-12-049.1
Evaluation of GPCP, HOAPS, ERA-Interim, and MERRA precipitation data at native resolution with single PACRAIN stations during 1989–2005. Measures are correlation (cor) and 90% confidence interval (CI), MAD and interquartile range (IQR) (mm day−1), and median bias (mm day−1). Boldface shows the best match to the PACRAIN reference.
Correlations of monthly precipitation anomalies are generally smaller (see Table 2, bottom) relative to absolute values (see Table 2, top). This is because the seasonal cycle of precipitation positively affects correlations of absolute amounts. In contrast, the seasonal cycle of precipitation is not included in time series of monthly anomalies of the corresponding long-term monthly average. Again, GPCP delivers the highest correlation (0.71) and the smallest MAD (1.55 mm day−1) for precipitation anomalies (see Table 2, bottom). As in the case of absolute values, HOAPS has the lowest correlation (0.62) and largest MAD (1.87 mm day−1) for precipitation anomalies in native data resolution. Some interpretations and discussion of these results are presented in section 5.
b. Evaluation of datasets with PACRAIN data on a common grid
After the evaluation of precipitation datasets in their native resolution in section 4a, datasets are now evaluated on a common 2.5° latitude–longitude grid with individual PACRAIN stations and with gridded PACRAIN data.
The results of the dataset evaluation of absolute values and monthly anomalies on the common grid with single PACRAIN stations are presented in Table 3. The effects of spatial averaging on the evaluation result are varied. Whereas the reanalysis datasets of ERA-Interim and MERRA perform similarly, the correlation of HOAPS improves from 0.70 to 0.75, resulting in a correlation not significantly lower than that of GPCP (0.77) (cf. evaluation of absolute amounts in Tables 2 and 3). For MAD, values range between 1.8 and 2 mm day−1. The improvement of the HOAPS correlation and MAD might be explained by the cancellation of small-scale deviations as a result of spatial averaging. These small-scale deviations are likely a result of the relatively low temporal coverage of HOAPS, owing to the SSM/I-only approach. Especially in the tropics, which exhibit large spatiotemporal variability of rainfall, temporal coverage is crucial to monitoring precipitation. To evaluate this hypothesis, a separate analysis of the HOAPS time series is done in the native spatial resolution, during successive time periods in which one, two, and three SSM/I sensors are used for data generation.
The results presented in Table 4 confirm that correlation increases when temporal coverage is increased. MAD values do decrease slightly. During the period when the temporal coverage of HOAPS is best (three SSM/I sensors used), the HOAPS correlation is similar to the correlations of the reanalysis precipitation data (cf. Tables 2 and 4). To eliminate the effects of analyzing HOAPS during different time periods, HOAPS monthly precipitation amounts generated by the combination of one, two, and three sensors are additionally analyzed in a common time period. The effect of temporal coverage on the evaluation results, as seen in Table 2, is confirmed by evaluating HOAPS while piecewise increasing the number of SSM/I sensors used for data generation during the time period that goes from 2000 through 2005 (cf. Table 5).
Evaluation of HOAPS absolute monthly precipitation with 34 PACRAIN stations during three successive periods when HOAPS is based on one, two, and three SSM/I sensors. Measures are as in Table 2.
As in Table 4, but with precipitation data based on one, two, and three satellites during the common time period from 2000 through 2005 with PACRAIN stations.
The biases between the precipitation products and the PACRAIN reference are almost independent of the dataset resolution. Hence, biases on the coarser grid are in the same range as those in the native resolution (cf. Tables 2 and 3).
The precipitation data from GPCP, HOAPS, ERA-Interim, and MERRA are also compared with gridded PACRAIN data. Datasets are compared on individual common 2.5° latitude–longitude grid boxes. Therefore, individual PACRAIN stations are spatially interpolated to grid boxes, where station density is sufficient. Hence, only a reduced PACRAIN database, as seen in Fig. 1 (gray boxes), is used in this part of the evaluation. The average monthly precipitation of the reduced PACRAIN data is 6.9 mm day−1.
Results show that correlations of absolute precipitation amounts of GPCP, HOAPS, and ERA-Interim with the PACRAIN reference slightly increase when the data are evaluated on common grid boxes (cf. Tables 6 and 3), relative to the evaluation with individual stations. The correlation of GPCP and PACRAIN grid boxes reaches 0.81, which is the highest correlation among all analyzed datasets. Correlations of HOAPS and ERA-Interim with PACRAIN are similar (0.77 and 0.76, respectively), while MERRA has a correlation with PACRAIN of 0.71. MAD values of the analyzed precipitation products and PACRAIN gridded data range from 1.6 to 1.9 mm day−1, corresponding to relative MAD values of about 23%–28%, with GPCP having the lowest MAD of 1.57 mm day−1. For the median bias, results are diverse but similar to the results obtained by using single PACRAIN stations. GPCP has a negative bias of about 12%, which is in accordance with the findings of Adler et al. (2012) in the tropical Pacific. ERA-Interim has a positive bias of about 9% and HOAPS has a negative bias of 9%, but MERRA has no bias with respect to the PACRAIN atoll station data. Hence, relative biases are not influenced by the reduction of the reference database. As in the case of the evaluation of reanalysis products in native resolution, in this evaluation too, the reanalysis datasets of ERA-Interim and MERRA both show underestimations of high precipitation amounts, which the satellite datasets obtained by HOAPS and GPCP can avoid (cf. Fig. 3 top and bottom). Again, an overestimation of low and medium precipitation amounts using the reanalysis data, especially ERA-Interim, is found.
Evaluation of GPCP, HOAPS, ERA-Interim, and MERRA precipitation data on a 2.5° lat–lon grid with PACRAIN gridded data during 1989–2005. Measures are as in Table 2.
As in Fig. 2, but against PACRAIN gridded data, on a 2.5° lat–lon grid.
Citation: Journal of Applied Meteorology and Climatology 52, 3; 10.1175/JAMC-D-12-049.1
For the evaluation of anomalies on the common grid boxes, GPCP offers the highest anomaly correlation (0.77) to the gridded reference data, which is an important measure for climate datasets. HOAPS, ERA-Interim, and MERRA give lower correlations of anomalies within the range of 0.70–0.72 (see Table 6). MAD values of monthly anomalies range from 1.35 to 1.6 mm day−1, with IQR values of about 2 mm day−1.
5. Discussion and summary
Tropical precipitation is characterized by high spatiotemporal variability and originates primarily from convective events, which can be very short lived. In addition, the tropical climate is characterized by systematic diurnal cycles of precipitation (Sato et al. 2009). The evaluation of four precipitation datasets during the period of 1989–2005 showed that GPCP performs best for the correlation of absolute values and of anomalies, evaluated either with single PACRAIN stations or with gridded PACRAIN data. This can be attributed to the fact that GPCP uses the synergy effect of relatively direct derivations of precipitation by microwave sensors and the high temporal coverage of geostationary infrared observations, calibrated with the microwave measurements. This increases the temporal sampling and seems to compensate for the disadvantage of a larger spatial resolution relative to the HOAPS dataset, emphasizing the importance of high temporal coverage in observing tropical precipitation and demonstrating the power of the chosen approach relative to the SSM/I-only approach applied to generate HOAPS. In native spatial resolution, the reanalysis datasets of ERA-Interim and MERRA outperform the HOAPS precipitation data in terms of correlation, when compared with the individual PACRAIN stations.
However, GPCP exhibits the largest negative bias in the analyzed domain (−12%) of the used datasets, which is a disadvantage when studying tropical precipitation. Because of the high variability of tropical rainfall, median absolute deviations between the datasets and PACRAIN are generally quite large, with values of 20%–30%. MAD values of HOAPS, ERA-Interim, and MERRA, which offer higher spatial resolutions than GPCP, are in the same range (25%–30%). Hence, the usage of the reanalysis products to analyze precipitation in high spatial resolution would be favored in conjunction with higher correlations at small scales. Further, it reflects improvements in reanalysis datasets in recent years (e.g., more sophisticated data assimilation schemes, inclusion of data from multiple satellites, and improved parameterizations of precipitation and relevant processes). In light of the additional atmospheric variables that reanalysis datasets can provide, reanalysis products can be a useful tool for studying the climate system.
The HOAPS correlation is significantly improved if the evaluation is done on the coarser 2.5° latitude–longitude grid with PACRAIN stations. The results of the evaluation on the common coarser grid revealed that the HOAPS monthly precipitation data suffer from a combination of a high spatial resolution and a relatively rare temporal sampling. This is why the reanalysis products of ERA-Interim and MERRA offer higher correlations with PACRAIN than does HOAPS at high resolution. HOAPS, ERA-Interim, and MERRA MAD values are within the same range of about 25%, while deviations in GPCP are somewhat smaller. Biases of the analyzed datasets are within the same range, with the strongest underestimations being those of the GPCP data (−12%) and the strongest overestimations being those of the ERA-Interim data (+9%). In addition, results show that the bias is not influenced by the spatial resolution of the datasets in the analyzed domain.
For the regression lines shown in the scatterplots, HOAPS is closest to the one-to-one line of all analyzed data at any data resolution (see Figs. 2 and 3), indicating that there are no systematic under- or overestimations of precipitation on a monthly time scale. In contrast, there exists a systematic pattern of behavior in both reanalysis products with reference to the PACRAIN atoll station data: high precipitation amounts are underestimated, whereas small and medium amounts are overestimated, which is especially true for ERA-Interim. This is a disadvantage when using the data for climate studies and might be due to the difficulties numerical prediction models have in simulating precipitation. This in turn might hint at drawbacks in the parameterizations of precipitation in the models. However, it is notable that despite this disadvantage, the differences in the error measures of the satellite datasets and reanalysis products are relatively small, which is likely a result of the extensive assimilation of satellite radiances into the reanalysis.
The separate analyses of the HOAPS time series in native spatial resolution when one, two, and three SSM/I sensors are available for data generation show that temporal coverage is a crucial factor when monitoring tropical precipitation. The correlation significantly increases if temporal coverage is increased. This in turn might result in a better resolution of diurnal cycles of precipitation. In the period when three SSM/I sensors are available for data generation, the correlation of HOAPS is similar to the correlation of the reanalysis products at high resolution. In addition, we analyzed the coherence between the temporal coverage and correlation during a common time period for HOAPS data. It is shown that correlation increases with increasing temporal coverage (cf. Table 5). This relation has to be regarded as being especially true for tropical precipitation.
Even though GPCP, ERA-Interim, and MERRA also make use of the SSM/I data, which vary in temporal sampling, no major jumps in the correlation could be found in comparisons with PACRAIN. The reason for this might be that GPCP uses only one SSM/I sensor at a given time and that reanalysis precipitation is strongly dependent on the model microphysics and parameterizations, in addition to the assimilation of SSM/I and other satellite data.
Because the HOAPS precipitation algorithm performs well on an instantaneous basis (Klepp et al. 2003), it could be improved when temporal coverage is increased. HOAPS is a single-source dataset, which is its strength and also its weakness. High stability and homogeneity owing to the well-calibrated SSM/I data come along with a relatively low degree of spatiotemporal coverage. This became clear especially at high spatial resolution. Nevertheless, there is a need for such datasets (e.g., for model-independent studies or just for comparison).
The results gained from the evaluation of datasets with gridded PACRAIN data should not be regarded in direct relation to the dataset evaluation with single PACRAIN stations, because in the gridding procedure, only a reduced number of stations have been used. Nevertheless, the precipitation regime is similar. The results presented in Fig. 3 and Table 6 show that GPCP, HOAPS, and ERA-Interim are in better agreement with the PACRAIN grid boxes than is MERRA in terms of correlation of absolute values. Concerning the correlation of monthly anomalies, GPCP has the highest value (0.77), while the other datasets are within the same, lower range (0.70–0.72). Hence, it can be concluded that GPCP best represents the variability at 2.5° latitude–longitude resolution. Deviations in the monthly anomalies range from 1.35 mm day−1 (GPCP) to 1.60 mm day−1 (HOAPS), with IQR values on the order of 2 mm day−1.
The relatively large deviations of 20%–30% between analyzed precipitation datasets and the PACRAIN reference are also a consequence of the climate regime in the analyzed domain, which is dominated by convective, highly variable rainfall. The uncertainties in the monthly precipitation datasets are significantly higher than those for other essential climate variables derived by satellites (e.g., cloud albedo or radiation). In recent years the spatiotemporal resolution of satellite-based precipitation datasets could be increased, for example, with help of the Tropical Rainfall Measuring Mission satellite (http://trmm.gsfc.nasa.gov), but time series of high-resolution precipitation datasets are still short. Despite the use of combined satellite precipitation products, the use of global precipitation datasets for regional climate monitoring and analysis is rather limited owing to the relatively large uncertainties at small scales. New satellite missions, such as the Global Precipitation Measurement Mission (http://pmm.nasa.gov/GPM), are planned, promising improved spatial and temporal data resolutions and a higher degree of data accuracy. The validation performed here demonstrates that there is a need for improved precipitation datasets. High spatial resolution in line with increased temporal coverage is one key to attaining these improvements. This study shows that the quality of precipitation datasets can be improved by increasing their spatiotemporal resolution.
Acknowledgments
The research leading to these results has received funding from the European Union, Seventh Framework Programme (FP7/2007-2013), under Grant Agreement 242093. The authors acknowledge dataset provisions by PACRAIN, GPCP, HOAPS, ERA-Interim, and MERRA. We thank Michael Klatt for some extra information about PACRAIN and Karsten Fennig and Axel Andersson for useful information about HOAPS and data support. Author BA acknowledges funding from the Hessian initiative for the development of scientific and economic excellence (LOEWE) at the Biodiversity and Climate Research Centre (BiK-F), Frankfurt am Main, Germany.
REFERENCES
Adler, R. F., C. Kidd, G. Petty, M. Morissey, and H. M. Goodman, 2001: Intercomparison of global precipitation products: The Third Precipitation Intercomparison Project (PIP-3). Bull. Amer. Meteor. Soc., 82, 1377–1396.
Adler, R. F., and Coauthors, 2003: The version-2 Global Precipitation Climatology Project (GPCP) monthly precipitation analysis (1979–present). J. Hydrometeor., 4, 1147–1167.
Adler, R. F., G. Gu, and G. J. Huffman, 2012: Estimating climatological bias errors for the Global Precipitation Climatology Project (GPCP). J. Appl. Meteor. Climatol., 51, 48–99.
Ahrens, B., and A. Beck, 2008: On upscaling of rain-gauge data for evaluating numerical weather forecasts. Meteor. Atmos. Phys., 99, 155–167.
Andersson, A., K. Fennig, C. Klepp, S. Bakan, H. Gral, and J. Schulz, 2010: The Hamburg Ocean Atmosphere Parameters and Fluxes from Satellite Data—HOAPS-3. Earth Syst. Sci. Data, 2, 215–234.
Andersson, A., C. Klepp, S. Bakan, H. Grassl, and J. Schulz, 2011: Evaluation of HOAPS-3 ocean surface freshwater flux components. J. Appl. Meteor. Climatol., 50, 379–398.
Béranger, K., B. Barnier, S. Gulev, and M. Crépon, 2006: Comparing 20 years of precipitation estimates from different sources over the World Ocean. Ocean Dyn., 56, 104–138.
Bowman, K. P., 2005: Comparison of TRMM precipitation retrievals with rain gauge data from ocean buoys. J. Climate, 18, 178–190.
Bowman, K. P., A. B. Philips, and G. R. North, 2003: Comparison of TRMM rainfall retrievals with rain gauge date from the TOA/TRITON buoy array. Geophys. Res. Lett., 30, 1757, doi:10.1029/2003GL017552.
Bowman, K. P., C. R. Homeyer, and D. G. Stone, 2009: A comparison of oceanic precipitation estimates in the tropics and subtropics. J. Appl. Meteor. Climatol., 48, 1335–1344.
Cleveland, W. S., 1981: LOWESS: A program for smoothing scatterplots by robust locally weighted regression. Amer. Stat., 35, 54.
Dee, D. P., and Coauthors, 2011: The ERA-Interim reanalysis: Configuration and performance of the data assimilation system. Quart. J. Roy. Meteor. Soc., 137, 553–597.
Dobler, A., and B. Ahrens, 2010: Analysis of the Indian summer monsoon system in the regional climate model COSMO-CLM. J. Geophys. Res., 115, D16101, doi:10.1029/2009JD013497.
Ferraro, R. R., 1997: Special Sensor Microwave Imager derived global rainfall estimates for climatological applications. J. Geophys. Res., 102, 16 715–16 735.
Ferraro, R. R., F. Weng, N. C. Grody, and A. Basist, 1996: An eight-year (1987–1994) time series of rainfall, clouds, water vapor, snow cover, and sea ice derived from SSM/I measurements. Bull. Amer. Meteor. Soc., 77, 891–905.
Greene, J. S., M. Klatt, M. Morrisey, and S. Postawko, 2008: The Comprehensive Pacific Rainfall Database. J. Atmos. Oceanic Technol., 25, 71–82.
Huffman, G. J., R. F. Adler, D. T. Bolvin, and G. Gu, 2009: Improving the global precipitation record: GPCP version 2.1. Geophys. Res. Lett., 36, LI17808, doi:10.1029/2009GL040000.
Jones, P. W., 1999: First- and second-order conservative remapping schemes for spherical coordinates. Mon. Wea. Rev., 127, 2204–2210.
Kidd, C., and G. Huffman, 2011: Review global precipitation measurement. Meteor. Appl., 18, 334–353.
Kidd, C., and V. Levizzani, 2011: Status of satellite precipitation retrievals. Hydrol. Earth Syst. Sci., 15, 1109–1116.
Klepp, C.-P., S. Bakan, and H. Grassl, 2003: Improvements of satellite-derived cyclonic rainfall over the North Atlantic. J. Climate, 16, 657–669.
Morrissey, M. L., 1991: Using sparse raingages to test satellite-based rainfall algorithms. J. Geophys. Res., 96, 18 561–18 571.
Muggeo, V. M. R., 2008: Segmented: An R package to fit regression models with broken-line relationships. R News,8 (1), 20–25.
New, M., M. Todd, M. Hulme, and P. Jones, 2001: Precipitation measurements and trends in the twentieth century. Int. J. Climatol., 21, 1899–1922.
Rienecker, M. M., and Coauthors, 2011: MERRA: NASA’s Modern-Era Retrospective Analysis for Research and Applications. J. Climate, 24, 3624–3648.
Sato, T., H. Miura, M. Satoh, Y. N. Takayabu, and Y. Wang, 2009: Diurnal cycle of precipitation in the tropics simulated in a global cloud-resolving model. J. Climate, 22, 4809–4826.
Shin, D.-B., J.-H. Kim, and H. j. Park, 2011: Agreement between monthly precipitation estimates from TRMM satellite, NCEP reanalysis, and merged gauge-satellite analysis. J. Geophys. Res., 116, D16105, doi:10.1029/2010JD015483.
Sobel, A. H., C. D. Burleyson, and S. E. Yuter, 2011: Rain on small tropical islands. J. Geophys. Res., 116, D08102, doi:10.1029/2010JD014695.
Tian, Y., and C. D. Peters-Lidard, 2010: A global map of uncertainties in satellite-based precipitation measurements. Geophys. Res. Lett., 37, L244407, doi:10.1029/2010GL046008.
Wilheit, T. T., A. T. C. Chang, M. S. V. Rao, E. B. Rodgers, and J. S. Theon, 1977: A satellite technique for quantitative mapping rainfall rates over the oceans. J. Appl. Meteor., 16, 551–560.
