1. Introduction

Remotely sensed measurements from meteorological satellite instruments play an extremely important role in providing valuable information on many key parameters of the global-scale hydrological cycle (water vapor, precipitation, snow cover, etc.). These satellite measurements supplement ground-based observations, especially in areas where in situ measurements are limited. Global precipitation is one of the most challenging parameters to retrieve, yet one of the most important of the hydrological cycle. The development of rainfall estimates from passive microwave satellite measurements, specifically those from the Defense Meteorological Satellite Program (DMSP) series, Special Sensor Microwave Imager (SSM/I) have been one of the most important sources of data because (i) the length of the dataset (e.g., SSM/I has been in operation since June 1987 to present); (ii) the operating frequency range (from 19 to 85 GHz) has the unique ability to penetrate through cirrus clouds and sense the emitted and scattered radiation by raindrops and precipitation-sized ice particles, respectively; and (iii) the conical-scan viewing geometry allows for maintaining a fixed viewing angle and a constant footprint size along the scan for each frequency (Poe et al. 2001).

The primary algorithm used in this particular study is an 85-GHz scattering-based algorithm over land, while a combined 85-GHz scattering and 19/37-GHz emission is used over ocean [see Ferraro (1997) appendix A1 for details]. The land portion of this algorithm is one of the components of the blended Global Precipitation Climatology Project (GPCP) rainfall climatology (Adler et al. 2003) and continues to be used as the operational product generated at the U.S. Navy’s Fleet Numerical Meteorology and Oceanography Center (FNMOC). Because a portion of the National Oceanic and Atmospheric Administration’s (NOAA) SSM/I contribution to the GPCP component is performed in real time, only basic quality control (QC) procedures have been performed to remove unrealistic values in the input data (i.e., antenna temperature). The first purpose of this study was that, although the rain algorithm used at FNMOC and for GPCP are the same, some additional parameterizations and new screening processes (e.g., sea/ice detection) were added for improving monthly products delivered to GPCP (Ferraro 1997). With more than 20 years of SSM/I data now available, enhanced QC procedures can be applied to improve the products (Semunegus and Bates 2008; Semunegus et al. 2010). Thus, the primary goals of this paper are to perform a statistical-based QC procedure on the input data (⅓° daily antenna temperature files) to remove spurious values not detected in the original database and to reprocess the rainfall product using the current version of the algorithm for the period 1992–2007. The election of this period is based on the fact that during the period of June 1990–December 1991, the 85-GHz channels aboard the SSM/I F-8 failed, so it is not possible to use the proposed algorithm. A second reason is because from January 1992 to October 2008, more than one operational SSM/I instrument was flying on DMSP satellites, improving the temporal sampling dramatically and reducing errors. This allows for creating dual-satellite estimates for the whole period to improve the accuracy of monthly rainfall.

The secondary purpose of this study is to assess the discrepancies associated with the SSM/I-derived monthly rainfall products through comparisons with various gauge-based and other satellite-derived rainfall estimates.

2. Data and methodology

SSM/I temperature data record (TDR) files from the DMSP SSM/I from the Comprehensive Large Array-Data Stewardship System (CLASS) were used. CLASS is a Web-based data archive and distribution system for NOAA’s environmental data that contain calibrated and earth-located antenna temperature (AT) data prior to the application of irreversible antenna pattern correction. More information about SSM/I TDR files used in this study can be found on the NOAA National Climatic Data Center Web site (available online at http://www.ncdc.noaa.gov/oa/rsad/ssmi/ssmi.html). These files have been mapped to ⅓° latitude × ⅓° longitude linear daily grids for the ascending and descending nodes, respectively. This method was established in the early stages of GPCP and must be maintained for continuity of the GPCP, version 2, (V2) global rainfall products; however, it is recognized that a superior product could be developed using the highest resolution data. For this study, we have used all available data from the F-10, F-11, F-13, F-14, and F-15 satellites. Table 1 summarizes the archive of SSM/I gridded data assembled for this application. In a second step, F-11 and F-13 (descending orbit at approximately 0600 LST) were combined into a single dataset. This allows for continuity in the time series (1992–2007) using nearly the same satellite observing times (early morning). The same procedure was followed with the F-10, F-14, and F-15 satellites (descending orbit varying from 0800 to 1000 LST) to create a late-morning time series for the same period (1992–2007). This dual-satellite dataset offers an excellent opportunity to achieve a better temporal sampling, as well as sampling at a different time of the day, critical to capture a portion of the diurnal variability of precipitation.

a. The statistical-based quality control procedure

The statistical QC scheme is based on the detection of outliers for each grid box and for every channel on the remapped antenna temperature files. This procedure was performed based on the mean (μ) and standard deviation (σ) in each grid box for the period 1992–2007. In this case, 16 yr of all available daily grids for all satellites and all nodes (ascending and descending) were used to create a unique spatial distribution field for both variables (mean and standard deviation). Figure 1 shows the μ and σ for all of the vertical (V) polarization channels. The 85-GHz V channel for January exhibits a relatively larger variability over land than over ocean (Fig. 1g). This fact is related to surface temperature variability over both surface types. The presence of ice and snow tends to depress the 85-GHz V channel, whereas over the Southern Hemisphere summer, the highest values are observed over Australia. For the standard deviation (Fig. 1h), the larger variability is observed in those regions with possible snow and ice cover. In particular, the accumulation–melting process produces a large variability in 85-GHz V channel while a secondary relative maximum is observed in those regions where convection is present. The scattering process produces a large depression in 85-GHz V, so the combination of relatively clear-sky days (high AT) and deep convection (low AT) leads to a high variability of this channel.

The process to remove unrealistic values is summarized in Fig. 2. In the first step, the standardized temperature bias ΔT* for each channel at a given location is performed:

View Expanded

where T_k is the observed temperature for a given location, μ_k is the mean, and σ_k is the standard deviation for that grid box. In this case, the subscript k represents each channel. This value is related to the distance (in terms of standard deviation units) between the observed temperature and the mean value. The outlier points are defined as those measurements with |ΔT*_k| > 10 for a given grid box and channel. If for a given location at least one channel achieves the limits defined before, then this data is flagged. After applying this process, it was found that not all spurious values are removed from the database. In those regions with large variability, 10 times the standard deviation leads to very high or very low physically inconsistent boundaries that cannot be used to remove suspicious values. In this case, a maximum and a minimum temperature threshold is set to 325 and 70 K, respectively, and all data beyond those limits are also flagged. This issue has been discussed in previous papers (see Ferraro et al. 1998), and the current maximum and minimum thresholds appear very close to those cited in the literature. In the original version of the rainfall retrieval algorithm, those channels beyond the maximum and minimum temperature thresholds are also flagged as missing data but just only those locations with missing data for all channels were excluded from the statistics. This issue represents a major difference compared between the original and current approaches: if one pixel is flagged as missing data for a given channel and this channel is used in the rainfall retrieval algorithm, then the erroneous value is retained in the original approach while that value is excluded from statistics in the current one.

From a visual inspection of several images, it can be seen that failures are often aligned along the scan line (a conical scan sensor, in this case). If one channel is out of range for a given region, the same failure is frequently observed in more than one channel. In a second step, the ΔT* = [ΔT*₁, ΔT*₂, … , ΔT*₇] is considered as a vector for the entire measurement at a given location. If ΔT*_k > 6 for k ≥ 4, then the entire vector for that location is also excluded from statistics. The same procedure is carried out for those data with ΔT*_k < −6. This constraint, considering the entire vector as outlier, appears reasonable from a physical point of view: considering extreme events, it is possible to achieve an extreme value for one channel (e.g., 85-GHz V is very low for large convective storms because the presence of ice) but the chances for getting a combination of at least four channels with very large or very low values (all more than or all less than 6 times the standard deviation) is very low. To gain a better understanding on how the QC procedure works, SSM/I F-13 descending orbits on 2 August 2005 over the eastern United States were selected as a case study. Figure 3 shows all the individual channels and the result of the QC procedure applied on 85-GHz V. While some pixels were removed because the sensor failed in all channels (i.e., Canadian northeastern coast), in other cases, the failure was found in one or more channel but not in all channels. In those cases, the algorithm has successfully detected most of wrong values, and those pixels were removed or flagged according to the procedure described in Fig. 2. In this case, all flagged and removed values are shown in white.

Figure 4 shows just the flagged data for SSM/I F-13 during August 2006. It is easy to recognize the arc-shaped features of the conical scanning sensors, which contain spurious values in the original database. This figure represents an “accumulation” of target data along the time; in other words, for one particular file, if one pixel is detected as outlier (as defined earlier), then that pixel is flagged in this figure. Figure 4 includes those data erroneously included in the statistic from the original database and removed after the new QC process. The total number of data affected with the QC process for this particular month represents approximately 4% of total area; however, considering each daily file (ascending or descending). This number represents less than 0.1% of the data for a given grid. The numbers are fairly constant over the time, but they get larger starting in 2004. Despite the fact that the number of data is relatively small, the effect can be substantial on the monthly rainfall accumulation. The comparison of monthly rainfall rates with the original dataset will be discussed in section 3a.

3. Results

Monthly rainfall estimates were generated using the algorithm previously described for the period January 1992–December 2007 using the SSM/I on board the DMSP F-10, F-11, F-13, F-14, and F-15 satellites. The SSM/I time series used in this study is 16 yr in length, which represents an increase of more than 10 yr from the previous study completed by Ferraro (1997) using the same algorithm. As in the aforementioned study, F-11 and F-13 (descending orbit at approximately 0600 LST) were combined into a single dataset to create an early-morning time series, while the same procedure was applied with F-10, F-14, and F-15 satellites (descending orbit varying from 0800 to 1000 LST) to create a late-morning time series. The transition time for every constellation (early- and midmorning time series) is May 2005 for F-11/F-13, May 1997 for F-10/F-14, and January 2000 for F-14/F-15. A dual-satellite dataset was created averaging both time series using the relative frequency (defined as the ratio between the number of valid samples and the number of possible samples) as a weighting factor. Although 16 yr of data is not long enough to define a traditional climatology, Fig. 5 shows the seasonal changes of rainfall for the whole period. Note the changes in the mid- and high-latitude storm tracks over the ocean, the intensification of and the northward shift of the ITCZ in June–August, and the monsoonal seasons of India and the southwestern United States.

In this section, comparisons between the current rainfall retrievals and the original on a monthly global basis will be performed to assess the effect of the QC changes. The comparison of SSM/I rainfall estimates with well-established rainfall databases, such as GPCP and the Global Precipitation Climatology Centre (GPCC) among others, will provide to the scientific community a framework of the accuracies and deficiencies of the current SSM/I rainfall algorithm.

4. Discussion and summary

This study presented global monthly rainfall estimates derived from 16 yr (1992–2007) of SSM/I measurements. A new statistical quality-control scheme based on the detection of outliers for each grid box and every channel on the remapped antenna temperature files was performed, and approximately 4% of the surface is flagged as unrealistic input data for each month. After removing all spurious data, the mean bias between the original and the reprocessed dataset is approximately 3 mm month⁻¹. This bias is somewhat larger over land (Fig. 7) than over ocean (Fig. 9). Although this bias is relatively small on a global scale, that amount could be significant in long trend analysis. The Intergovernmental Panel on Climate Change (IPCC) working group found trends ranging from −7 to +2 mm decade⁻¹ for the 1951–2005 period and slightly larger for the last two to three decades (Solomon et al. 2007). On the other hand, on regional-scale studies a large amount of rainfall (more than 60 mm month⁻¹ or 2 mm day⁻¹) could be erroneously placed because of the existence of spurious data in the original AT database.

Comparisons of the monthly rainfall were made with the best available rainfall estimates. GPCC gauge dataset was used as a primary source of comparison over land, while GPCP is used over the ocean. Some conclusions that can be made include the following:

• A good agreement between the annual zonal mean for GPCC and the dual-satellite estimates is found for the period 1992–97 in the region between 30°N and 30°S. Beyond this region, the difference becomes larger because of limitations of the scattering approach used in this study. This agreement is also reflected in the time series of the annual running mean for the tropical region (30°N–30°S).

• Similar results are observed over ocean where the agreement between SSM/I estimates and GPCP can be observed in the annual zonal mean and the time series. In this case, it can be observed that a GPCP exhibits larger values than the SSM/I rainfall product (∼4 mm month⁻¹), while GPROF is closer to the SSM/I output.