1. Introduction
Long-term, homogeneous climate data records of precipitation are needed to both establish baselines against which future change can be detected (Xie and Arkin 1997; Adler et al. 2003, 2009), as well as for validation of climate models (Tapiador et al. 2017). In hydrology, long-term time series are also essential to establish recurrence frameworks (Knapp et al. 2011). With a changing climate, and the possibility that rainfall statistics may be changing, this becomes particularly important.
Current observational precipitation products may be categorized as climatological products that aim to construct a homogeneous time series at the expense of using all available data [i.e., Climate Data Records (CDRs)]. The gauge-based Global Precipitation Climatology Centre (GPCC; Schneider et al. 2016, 2017) has climatological products spanning over 70 years, but it is limited to land and uses only the subset of rain gauges with a continuous record for the entire time series (e.g., Beck et al. 2005). This reduces the available global gauge network from approximately 100 000 gauges to a mere 10 000 gauges for the 1951–2020 climatology (Schneider et al. 2008). Likewise, satellite products such as Global Precipitation Climatology Project (GPCP; Adler et al. 2003) represent a climatology in that a single passive microwave sensor is used to intercalibrate available IR data for the complete time series. This differs from products such as CMORPH (Joyce et al. 2004), GSMaP (Okamoto et al. 2005), IMERG (Huffman et al. 2020), and others that aim to produce precipitation with the highest space–time resolution available but at the expense of consistency in the long-term time series.
A number of efforts have been undertaken to establish the accuracy of the products listed above by comparing them with each other as well as with the global model reanalysis. It was through one of these efforts (Adler et al. 2012) that GPCP estimated uncertainties of approximately 10% in the global mean precipitation. However, a requisite step in such product intercomparisons, including in the cited work, is the need to eliminate any products that do not converge with the ensemble mean to within a specified value. Masunaga et al. (2019) reported on such an intercomparison of widely used satellite merged products. While they found that mean values were relatively consistent among products, the extremes were very different. In that study, the gauge data had larger disagreements in the extremes than did the different satellite products. This shows that even when means agree, it may not be for the same reasons, and rain gauges have their own set of issues, related largely to the need to interpolate across large distances.
Uncertainties are of course central to any climate time series as all of the products have limitations. Yet, no estimates of uncertainty currently exist beyond the product intercomparisons. This is likely due to the difficulty in quantifying uncertainties from first principles. In this study, the authors use the GPROF algorithm (Kummerow et al. 2015) to try to build an error model for a time series consisting of a number of different satellites—each with different channels, spatial resolution, equator crossing times, and potentially unresolved calibration issues. In principle, the GPROF algorithm can incorporate a diverse range of sensors in a fully parametric way that is essential to translate error characteristics from one sensor to another. The GPM program (Skofronick-Jackson et al. 2017; Hou et al. 2014), has already devoted significant resources to correct for calibration differences between spaceborne radiometers as part of the XCAL intercalibration effort (Berg et al. 2016).
As with any uncertainty, discussion immediately turns to random and systematic errors. Unfortunately, these prove inadequate for the current study. Random errors, which will be labeled as algorithm errors in the subsequent discussion, reduce as the number of satellite-retrieved pixels (i.e., observations) increase. As true random errors, they decrease at
Section 2 describes the sensors, algorithms, and time series that constitute a CDR before describing our treatment of the random, sampling, convective fraction, diurnal cycle, and information content errors. Section 3 lists the results, followed by discussion and conclusions in section 4.
2. Data collection and methods
To investigate satellite PMW precipitation CDR uncertainties, this study uses satellite and reanalysis datasets corresponding to the 34-yr period from 1987 to 2020. The four data products considered are 1) PMW observations from the intercalibrated level-1C brightness temperature (Tb) dataset (Berg et al. 2018), 2) precipitation estimates from NASA’s GPM PMW precipitation retrieval (GPROF; Kummerow et al. 2015), 3) precipitation estimates from GPM DPR-combined product (GPM_2BCMB; Olson 2017), and 4) the most recent ECMWF reanalysis (ERA5; Copernicus Climate Change Service 2017). The four products are interconnected and highly correlated across temporal/spatial scales (Watters et al. 2021). PMW satellite data are limited to conical scanning microwave imagers only (see the full list of acronyms in the appendix).
a. Data collection
1) Intercalibrated level-1C data
The level-1C data containing intercalibrated microwave brightness temperatures relies on observations from 14 conical-scanning microwave imagers launched between 1987 and 2014 (Berg et al. 2018). Development of this data record involved quality control, view angle and geolocation errors, emissive reflector issues, solar and lunar intrusions into the warm load, antenna pattern spillover effects, and intercalibration of the following PMW imagers: GMI, TMI, SSM/I (F8, F10, F11, F13, F14, and F15), SSMIS (F16, F17, F18, and F19), AMSR-E, and AMSR2. The constellation members timeline is shown in Fig. 1, with sensor-specific channels listed in Table 1.
Timeline of available window channel microwave imagers. The study uses a satellite climate precipitation record that was created by using a total of 14 PMW sensors covering the pre-TRMM, TRMM, and GPM eras (1987–2020).
Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-22-0179.1
PMW instruments, frequencies (GHz), and their polarizations (h for horizontal; v for vertical) used to generate the long-term precipitation CDR.
Rigorous quality control has resulted in removal and/or flagging of some instruments observations (details available in Berg 2017). SSM/I F13 is used as a reference for consistency over the SSM/I-era, while GPM GMI observations are considered to be an absolute calibration reference for the entire level-1C (i.e., Tb) dataset. The level-1C product used in this study is available from the NASA Precipitation Processing System (https://pmm.nasa.gov/data-access/downloads/gpm).
2) PMW precipitation dataset: GPROF algorithm
Developed in the mid-1990s at Goddard Space Flight Center, the GPROF precipitation algorithm (Kummerow and Giglio 1994) has been serving as the operational PMW precipitation retrieval at NASA Precipitation Processing System (PPS) for over three decades. Its fully parametric scheme ensures consistency across a constellation of cross-track (e.g., Kidd et al. 2021) and conical scanning sensors, including the above-mentioned radiometers. An operational version of the algorithm [version 5 (V05); operational at NASA PPS during the 2017–21 time frame] is used in this study for production of precipitation estimates over the period of available brightness temperature (i.e., level 1C) data record. While the evolution of the retrieval is documented in Kummerow et al. (2015), the most important algorithm properties relevant to this study are outlined below.
3) Calibration and reference datasets
Launched in February 2014, the GPM mission Core Observatory includes a 13-channel conical-scanning GMI and a dual-frequency precipitation radar. The imager is designed to 1) serve as a calibration standard for the entire GPM PMW constellation, and 2) in synergy with DPR, offer a link between the passive- and active-microwave signatures of the atmospheric column under all weather conditions. The DPR itself, being the most advanced satellite precipitation radar to date, provides invaluable information on vertical profile of global precipitation systems. Data collected over a 5-yr period (i.e., March 2014–February 2019) are used in the present study. The two products of particular interest are DPR-combined level-2 precipitation estimates (GPM_2BCMB; Olson 2017) and GMI level-1C PMW brightness temperatures (Berg 2016).
4) ERA5
ECMWF reanalysis (ERA) data are used at its native spatial and temporal resolution (0.281 25°, hourly) in this study to provide 1) information on environmental conditions and 2) the necessary elements for constructing idealized satellite sampling at various scales. Parameters of particular interest include 2-m temperature (2mT), sea surface temperature (SST), and precipitation rate.
b. Methodology: Defining and assigning uncertainties
1) Uncertainty of instantaneous rainfall estimates: Random error uncertainty
Random uncertainty, hereinafter referred to as Bayesian uncertainty, is estimated by calculating gridded-product standard deviations using instantaneous rate errors and DPR-combined product as a reference. For this purpose, the GPROF algorithm is tasked against the same data that compose its a priori database. This ensures that the bias, or nonrepresentativeness, is not a potential source of error, and that all the errors are truly random. The full sequence of steps consists of the following: 1) retrieve precipitation against the a priori database, 2) create monthly global grids of the output and reference data, 3) compute the differences across all grid boxes, sampled by surface type and rainfall rate. To preserve precipitation retrieval output dependance on surface type, GPROF surface type maps are resampled to match the output resolution (e.g., 5° × 5°) keeping only four main surface types: ocean, land, snow, and mixed (a mix of land and ocean). To address uncertainty’s dependency on the rainfall rate itself, five balanced (i.e., equally populated) rain-rate bins are identified: (0, 0.3, 0.9, 1.7, 3.6, and infinity) mm day−1. Last, the result provides Bayesian uncertainty as a function of surface type, rain rate, and month. The method is repeated for each PMW sensor in the data.
2) Sampling-induced uncertainty
Sampling uncertainty (i.e., observing frequency uncertainty) is defined as a deviation from the true value caused by limited sampling of precipitating systems within a grid box during a time interval of interest. This uncertainty is driven by the satellite orbit properties and the total number of observations (i.e., samples) available during the sampling period. To estimate sampling uncertainty, modeled precipitation fields are employed in simulating a wide range of different sampling scenarios over a domain of choice (e.g., 5° monthly global grid). The simulations consider an ideal scenario, where a region of interest is sampled continuously, as a reference that is then compared with other subsampled cases to estimate uncertainty as a function of month and sampling frequency.
To ensure intra-annual variability does not play an important role, a 5-yr GPM-era (2014–18) period of ERA5 precipitation fields, at 1-h 0.281 25° resolution, is used to calculate the sampling uncertainty. The high-resolution hourly data are first gridded into 5° × 5° hourly global grids, for four different surface types (ocean, land, snow, and mixed), and split into monthly time series with up to 740 individual hourly samples (24 h for 30 days). These time series are then used to derive 1) adjustment factor for any bias caused by a satellite drift, and 2) uncertainty caused by suboptimal sampling. In the first step, monthly means corresponding to different sampling local times are recorded to be applied to any satellite products that suffer from significant drift in equatorial crossing time (e.g., DMSP sensor series). In the second step, the time series are randomly sampled to simulate subsampled monthly records. The process is repeated until each sampling frequency (from 1 to 740) of the monthly intervals is simulated 100 times. Next, the simulated time series are used to calculate monthly means at the 5° × 5° grid. The resulting means are assessed against the reference, providing a distribution (sample size 100) of the differences for each sampling frequency. Last, using 3-month centered data, standard deviations of these distributions define the sampling uncertainty of monthly precipitation product.
3) Diurnal cycle uncertainty
After removing the mean diurnal cycle bias, however, a residual uncertainty associated with variability in the diurnal cycle remains. This is because the diurnal cycle varies both regionally as well as seasonally or over time. These variations are affected by a large number of factors. While our approach assumes that to first order most of the diurnal cycle variability relates to surface type and precipitation regime (e.g., the West Coasts’ stratiform systems, MJO, Amazon/Congo convection), this is an assumption that cannot capture the full range of variations in the diurnal cycle due to limited data, errors in the ERA5 reanalysis, etc. In our analysis, the solution of Eq. (4) is used to create diurnal cycle data at each 5° grid box over the 5-yr period from the ERA5 data. With each grid box being characterized by 60 (12 months × 5 years) diurnal cycle sinusoid curves, a k-means clustering method is used to identify grid boxes with similar diurnal precipitation variability regimes. After iterating through the results for a range between 2 and 10 clusters, five clusters were identified as those producing distinctive diurnal patterns (shown in Fig. 2 for August). These five clusters were named based on the local time and amplitude of the anomaly’s peak: high-amplitude early peak (HAEP), low-amplitude late-peak (LALP), high-amplitude late-peak (HALP), low-amplitude early peak (LAEP), and medium-amplitude late-peak (MALP) diurnal types. Figure 3 offers an insight to the clusters’ spatial distribution implying their links to general precipitation regimes.
Monthly averages of precipitation diurnal anomaly (for August) characterized by five distinct clusters: high-amplitude early peak (red), low-amplitude late peak (yellow), high-amplitude late peak (green), low-amplitude early peak (blue), and medium-amplitude late peak (magenta) diurnal types. Data source: ERA5 hourly product, precipitation fields regridded into 5° monthly fields for a period of 5 years (2014–19).
Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-22-0179.1
Map of the five clusters shown in Fig. 2 (for August).
Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-22-0179.1
With each 5° grid box being characterized by one of the five clusters, diurnal cycle uncertainty is defined as the standard deviation of the precipitation ratio anomaly for a given cluster. This uncertainty is calculated as a function of a local hour and cluster using all global 3-month centered diurnal cycle sinusoids.
4) Information content uncertainty
In remote sensing theory, the available information content defines how strongly (or loosely) the inverse problem can be constrained. PMW observations used for generating precipitation estimates in this study come from sensors with different channels and capabilities resulting in varying information content related to precipitation. In particular, the footprint sizes and the number of observed frequencies and polarizations (i.e., channels) a sensor uses to collect the radiometric signature of the underlying atmospheric column greatly affects the accuracy of the retrieval. Increased resolution and a greater number of observed frequencies typically deliver significantly higher information content, resulting in a more accurate precipitation estimate. To estimate the information content contribution to the uncertainties of the precipitation data record, GMI is once again considered as a reference standard and used to simulate observations from the other sensors, including their channel configuration and sampling geometry. These simulated observations are then used to retrieve precipitation, which is assessed against the precipitation estimates obtained using the reference standard (i.e., GMI). Since the simulated Tb for each sensor is based on the GMI sampling, and thus covers the same spatiotemporal domain as GMI over the 5-yr of GPM-era (2014–19), the resulting differences between the simulated products and the GMI-based estimates should come exclusively from the differences in the information content between GMI and the simulated sensor (e.g., SSM/I). The GMI level-1 product (2014–19) is used to simulate a 7-channel SSM/I, a 9-channel AMSR/TMI, and an 11-channel SSMIS (see Table 1) synthetic level-1 data. As the GMI orbit covers only 66°N–66°S, all simulated datasets in this analysis correspond to the same latitudinal region, except for the TMI, which is restricted to the TRMM domain (40°N–40°S). The FOV resampling process uses the Backus and Gilbert (1970) approach to (de)convolve GMI Tb to the resolution and sampling of the corresponding sensor. For each simulated FOV, sensor-specific (de)convolution coefficients are applied to a 11 × 11 pixels patch of GMI-observed Tb (Fig. 4, left panel). Applying an optimized gain function to account for sensors’ channel specifications, including the sampling geometry of the sensor, Tbs are computed for the sensor to be considered using the closest available GMI frequency and polarization. The scan geometry of GMI is thus converted into a pseudoswath of a simulated sensor that is slightly narrower than the original GMI ground track. An example of GMI-SSM/I resampled geometry is given in Fig. 4 (right panel).
Spatial distribution of GMI native and SSM/I-simulated scanning pattern. Each SSM/I FOV is derived using a centered 11 × 11 pixel patch of GMI observations. The nonoverlapping portion of the GMI swath was used only for (de)convolving purposes and was excluded from GPROF runs comparisons when determining information-content variability.
Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-22-0179.1
The pseudoswath Tbs are delivered at the same channel frequencies as those of GMI but reflect their sensor’s spatial resolution and sampling. Therefore, prior to running the retrieval on the simulated-sensor pseudo Tbs, the GPROF a priori database must be adjusted for any systematic differences between the two sets of Tb (pseudo vs GMI-observed ones). To estimate the differences due to the information content, monthly 5° averaged precipitation estimates for each simulated sensor are compared with those from GMI. The differences are recorded as a function of year, month, and location (i.e., grid box). For each month in the year, using the 3-month centered windows (e.g., DJF for January), the information content uncertainty is defined as standard deviation of the corresponding differences.
5) Convective fraction uncertainty
Another important contribution to the overall uncertainty comes from the precipitation retrieval itself. While the GPROF algorithm [section 2a(2)] is a state-of-the-art enterprise retrieval, it is important to consider the impact of errors in the retrieval and their propagation into spatially and temporally averaged estimates (Elsaesser and Kummerow 2015). For the applications considered in this study, many of the retrieval errors make a negligible contribution to the time-averaged large-scale precipitation products. However, the ability of the retrieval to distinguish between radiometrically similar scenes characterized by different precipitation rates is one that requires attention even when considering 5° monthly products. Due to the equal treatment of distinct hydrometeor profiles in the GPROF algorithm with similar Tbs, an error introduced to the instantaneous precipitation rates becomes a function of precipitation type (Petković et al. 2017, 2019).
Given that there is no direct information on precipitation type, a proxy for convective fraction over a grid box is identified based on surface temperature (Fig. 5). This proxy is thus used to estimate the effect of the variability in precipitation systems morphology on the total uncertainty of the retrieval’s output. Using estimates from the combined-DPR product over the same 5-yr GPM-era period, the relationship between GPROF precipitation biases and the convective ratio proxy (i.e., 2-m temperature from ERA5) is established for each month of the year (Fig. 6) for four different surface types and five rainfall rate bins [as defined in section 2b(1)]. Uncertainty induced by varying the convective fraction is estimated as a standard deviation of precipitation bias over the 3-month centered periods for each month of the year.
Convective fraction as a function of SST. The relationship uses monthly 5° gridded products of DPR global ocean observations and collocated ERA5 SST. Each 5° grid box has at least 50 000 DPR samples.
Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-22-0179.1
GPROF bias as a function of SST–proxy for convective fraction, showing the relationship of monthly 5° gridded product generated from GPROF output for global oceans and collocated ERA5 SSTs. Box-and-whisker bars denote the 10th, 25th, 50th, 75th, and 90th percentile of the bias values (by count), for every 5 K of SST.
Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-22-0179.1
6) Calibration uncertainty
Although the level-1C Tbs are intercalibrated for consistency between sensors [section 2a(1)], residual calibration errors likely remain. A number of issues can impact the calibration and thus the quality of the observed Tb. These include, but are not limited to, permanent or intermittent loss of a channel, an increase in the channel noise or noise equivalent differential temperature (NEDT), changes in the orbit and the local observing time, deviations in the satellite orientation, and changes in antenna and feedhorn properties. To provide an estimate on how these effects may translate to the level-2 products used in this study, a set of synthetic experiments is performed. The GPROF retrieval is run multiple times using 1 month of SSMIS simulated observations. With each run, a varying amount of noise and/or bias is introduced to the Bayesian error covariance S [see Eq. (1)] and the level-1 input Tb, respectively. The retrieval outputs are then used to provide global mean monthly precipitation rates as function of the Tb bias and sensors sensitivity, providing an upper limit on possible calibration uncertainty contributions.
c. Methodology: Combining the uncertainties
With the five main contributors to the uncertainty defined as described above, and without accounting for any effects arising from calibration uncertainty causes [section 2b(6)], the total uncertainty of the CDR product is calculated using Eq. (3) but accounting for the number of observations within each grid box (i.e., sample size). The total uncertainty is computed for each 5° grid box in units of millimeters per hour based on the month of the year, surface type, rain rate, diurnal cycle mode, 2-m temperature, and the number of samples during the month. Table 2 lists the individual components of the uncertainty along with the dimensions and the properties used for assigning the uncertainty information to the CDR.
Uncertainty contributors and their dimensions. TPW and 2-m temperature bins are of 1-mm and 5-K widths, respectively.
3. Results (per uncertainty contributor and combined)
The uncertainties for each of the contributors listed in Table 2 are presented individually before the merged result is shown.
a. Uncertainty from random retrieval errors
The pixel-level random uncertainty contribution to any monthly scale product, resulting from the Bayesian estimation described above, is expected to be small. Table 3 lists mean pixel-level Bayesian uncertainties over 5° monthly grids for SSMIS sensor during August as a function of surface type and rain rate. They are taken directly from the retrieval output that reports these errors based on the brightness temperature fit of various rain profiles and the observed Tb vector. It is notable that uncertainties stay within 0.4 mm h−1 for rain rates characteristic to 5° monthly grids (e.g., up to 1 mm h−1). The independent nature of pixel-level random uncertainties and a typically large sample size of observations available to 5° monthly products will yield [see Eq. (3)] a total uncertainty from random retrieval errors to be by far the smallest among all contributors considered in this study. Bayesian uncertainty for other sensors (not shown here) are found to be similar to those of the SSMIS.
Mean Bayesian (i.e., random) pixel-level uncertainty for August as a function of the surface type and rainfall rate; 5° monthly scale.
b. Uncertainty induced by varying sampling frequency
Variability in the sampling frequency of precipitating systems across the globe is driven by the orbit inclination of the individual satellites and the total number of orbiting sensors at a given time. The timeline of sampling frequency over the 30-yr period is presented at the end of this section (section 3g). Figure 7 shows the sampling-induced uncertainty as a function of the number of hourly samples during August. While an increase in sampling frequency leads to the expected reduction in uncertainty, differences in the amplitude of the sampling uncertainty for different surface types depicts the variability in precipitation rate distribution from hourly to monthly scales (e.g., afternoon convection over ocean vs random—system-specific—events over land, and coastal mixed surface type).
Sampling uncertainty for August as a function of the surface type and sampling frequency; 5° gridded product.
Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-22-0179.1
c. Uncertainty induced by variations of diurnal cycle
For large-scale climate applications, it is important to account for the fact that sun-synchronous sensors do not capture diurnal cycle of precipitation. In this study, monthly precipitation rate estimate is corrected based on the ratio of the mean daily precipitation rate to the mean estimated rate at the local time(s) specific to a given satellites overpass schedule. While this mean correction removes the overall diurnal bias, the variability of the diurnal cycle regionally and over monthly scales both during and prior to the GPM-era remains and can propagate into the spatially and temporally averaged level-3 product. An example of the residual diurnal uncertainty for August is shown in Fig. 8. Not surprisingly, when comparing the regimes (i.e., surface types/clusters), the amplitude of diurnal cycle uncertainty is strongly correlated to the total amount of precipitation within a given regime. The example shown in Fig. 8 may be related to an MJO event in the Indian Ocean that has different diurnal characteristics than the background precipitation.
Diurnal cycle variability–induced uncertainty for August as a function of surface type; 5° gridded product.
Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-22-0179.1
d. Uncertainty induced by variations in sensor information content
As described previously, the information content uncertainty is estimated using GMI as an absolute reference. Figure 9 compares the information content uncertainty for each of the constellation sensors relative to GMI. Improvements in spatial resolution and the addition of new channels for more current sensors leads to a decrease in the information content uncertainty for the more recent observations. With similar channels and spatial resolution to GMI, estimates from TMI and AMSR2 exhibit the smallest amount of information content uncertainties relative to GMI. On the other hand, the limited channel availability and low spatial resolution of the SSM/I sensors, made even worse by the loss of the 85-GHz channels on DMSP F8, results in the highest information content errors. Because sensors with lower information content will tend to more closely follow the retrieval’s database mean, regional biases are not only possible but highly likely as the precipitation differs from its statistical average.
Information content uncertainty for TMI, SSM/I, SSMIS, and AMSR2 as a function of month.
Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-22-0179.1
e. Uncertainty induced by variations of convective fraction of precipitating systems
Another consequence of using limited information content to retrieve precipitation rate, are uncertainties related to the cloud system morphology. Unable to distinguish between precipitation types, retrieval uses a same Tb-to-rain-rate relationship for all storm profiles with similar Tbs, even when those storms have significantly different surface precipitation rates. Consequently, seasonal, subseasonal, and regional biases are introduced to instantaneous level-2 precipitation estimates and then propagated to the spatially and temporally averaged level-3 products. Using GPM GMI and DPR observations during the GPM era, the contribution to the total uncertainty due to variations in storm morphology is estimated to be between 0.1 and 0.4 mm day−1. Table 4 shows the convective fraction uncertainty as a function of month. The large variation in the amplitude of the uncertainty estimates shown here is related to the prevalence of strong convection that is typically associated with high precipitation rate regimes in the global monthly means.
Uncertainty (mm day−1) of long-term 5° gridded product induced by variations of convective fraction, for combined surface types in August, as a function of rain rate and temperature.
f. Estimating uncertainty induced by nonoptimal calibration
As explained above, increased sensor noise and/or errors in the calibration can also lead to increased uncertainties in the precipitation estimates. To provide an estimate on the scale of such effects, Table 5 shows what one could expect, in terms of the uncertainty introduced by all unaccounted calibration effects to the brightness temperature level-1C data used in this study.
Changes in the mean daily global precipitation rate of the SSMIS sensor as a consequence of added noise and bias to the level-1 product (i.e., GPROF input). Values correspond to 1-month global means (June 2017). Low channels: less than or equal to 37 GHz; high channels: greater than or equal to 91 GHz.
g. The overall uncertainty of long-term precipitation record
The overall uncertainty is estimated by combining the contributors listed in Table 2. The uncorrelated nature of their origin makes Eq. (3) suitable for calculating the total uncertainty of the precipitation data record. Figure 10 presents global distribution of individual and combined uncertainties for August 2017 at 5° resolution. With five available conical-scanning radiometers during this month (SSMIS F17, SSMIS F18, SSMIS F19, AMSR2, and GMI) the overall uncertainty is dominantly driven by convective fraction variability, followed by contributions from sampling, diurnal, information content, and random errors. The same hierarchy of the contributors’ importance is seen throughout the entire data record.
Monthly mean precipitation data record uncertainties for August 2017 at 5° grid. Available sensors: GMI, SSMIS F17, SSMIS F18, SSMIS F19, and AMSR2.
Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-22-0179.1
To depict the evolution of the uncertainty during the three decades of passive microwave observations, and the effect the introduction of additional sensors has to the precipitation product quality, Fig. 11 presents a timeline of several relevant parameters. Notably, the higher sampling yields lower uncertainties, while a decline in sensor performance, such as that of SSM/I in late 1987, strongly decreases the reliability of the precipitation data record. The loss of the 85-GHz channel on the SSM/I instrument in the 1980s, in combination with low sampling, led to the highest uncertainty in the record. However, the change in the precipitation rate anomaly related to this loss of SSM/I channels significantly exceeds the estimated uncertainty. As the data record evolves, the introduction of TMI in the 1990s, followed by AMSR-E, AMSR2, and GMI in the 2002–14 period, mitigates the uncertainty sources related to relatively low information content of SSM/I and SSMIS instrument series.
Timeline of the 34-yr global monthly mean precipitation data record (1987–2020) at 5° grid: (top) Precipitation monthly anomalies (solid line), their uncertainty (red shading), and mean number of samples per grid box (blue). (middle) Seven-month running-mean global monthly precipitation anomalies, the corresponding uncertainty (blue shading), and the number of passive microwave instruments contributing to the data record (green). (bottom) Individual contributors and total uncertainty of the CDR precipitation product.
Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-22-0179.1
Low but notable trends in global precipitation rate appear to be well outside the uncertainty bars (the blue shade around the mean anomaly shown in the middle panel of Fig. 11). However, as discussed in the next section, this is not a sufficient condition to qualify the data record as a global precipitation trend reference.
4. Conclusions and discussions
The results presented here consider the main contributors to the uncertainty in a long-term global precipitation record obtained from passive microwave conical-scanning satellite sensors from the GPROF algorithms. While the choices made in identifying and estimating these uncertainties are relatively easy to explain and deliver, combining them into a single uncertainty proved challenging. The challenge is twofold: The first has to do with sensor calibration. Lasting 2 to 5 times their design-defined lifetime, satellite radiometers, just as any other spaceborne sensors, are exposed to the environment that inevitably affects their performance. Relying on prelaunch calibration and testing, with no physical access to the instrument, leads to potential errors that are not simple to quantify. Simulations of added noise and bias (see section 3f) suggest that realistic effects of those changes exceed the identified trends in large-scale global precipitation rate. These time-dependent calibration-induced effects are likely responsible for sensors’ occasional but large departures in the mean global precipitation rate. As documented in Fig. 12, the sensors from the SSM/I (F8/F10) and SSMIS (F16/F18) series are most prone to this problem. The second issue is related to the assumptions in the algorithm. The assumptions in the GPROF algorithm that were known to impact convective and stratiform precipitation estimates led to the use of convective organization as a predictor of uncertainty during periods where no radars are available to quantify the degree of convective organization. However, it is not only possible, but even likely, that the algorithm produces biased results, although perhaps with smaller magnitudes, in response to other large-scale cloud and precipitation properties. We contend that significant work is still needed to not only validate precipitation products, but to learn how to predict validation results based on large-scale environmental conditions that led to the precipitation.
Global mean precipitation anomalies (70°S–70°N), with uncertainties by satellite. The time series is smoothed using a 7-month running-mean filter. Shaded color denotes the uncertainty of each satellite product.
Citation: Journal of Applied Meteorology and Climatology 62, 8; 10.1175/JAMC-D-22-0179.1
Work remains to be done on both calibration methods and predicting algorithm uncertainties. One important result of this study, however, is that the uncertainties are not simply random or systematic, but have a strong dependence on time and space scales that couple to the source of error itself. The full analysis, requiring an examination of which errors persist and which errors become small as a result of natural variability, must be undertaken for each space and time resolution under consideration. It is not simply enough to assign an uncertainty to a satellite pixel. Instead, a full elaboration of errors requires each satellite pixel to provide the source of the uncertainty, as well as the time and space correlation of those uncertainties. For now, it may be necessary to provide these uncertainties in a postprocessing step that considers the particular time–space resolution of the product—be it instantaneous, 5° monthly, or global monthly time series.
Acknowledgments.
This work was supported by NASA Grant NNX08AT04A from the MEaSUREs program. Authors Petković and Kummerow thank Ralph Ferraro for hosting Petković, and they acknowledge the support provided through the University of Maryland/ESSIC and the Cooperative Institute for Satellite Earth System Studies–CISESS (NOAA Grant NA19NES4320002).
Data availability statement.
Data used in this study can be freely accessed either through the regular public repositories, as cited in the text (e.g., ERA5: https://doi.org/10.24381/cds.bd0915c6; GPM-DPR product: https://storm.pps.eosdis.nasa.gov/storm/), or at Colorado State University ftp, per user request (e.g., level-1C intercalibrated brightness temperature product and level-3 multisensor long-term precipitation record).
APPENDIX
List of Acronyms
| AMSR-E |
Advanced Microwave Scanning Radiometer for Earth Observing System |
| AMSR2 |
Advanced Microwave Scanning Radiometer 2 |
| CDR |
Climate Data Records |
| CMORPH |
Climate Prediction Center morphing technique |
| DMSP |
Defense Meteorological Satellite Program |
| DPR |
Dual-frequency precipitation radar |
| ECMWF |
European Centre for Medium-Range Weather Forecasts |
| ERA5 |
Fifth major global reanalysis produced by ECMWF |
| FCDR |
Fundamental Climate Data Record |
| FOV |
Field of view |
| GMI |
GPM Microwave Imager |
| GPCC |
Global Precipitation Climatology Centre |
| GPCP |
Global Precipitation Climatology Project |
| GPM |
Global Precipitation Measurement |
| GPROF |
Goddard profiling algorithm |
| GSMaP |
Global Satellite Mapping of Precipitation |
| HAEP |
High-amplitude early peak |
| HALP |
High-amplitude late peak |
| IMERG |
Integrated Multi-Satellite Retrievals for GPM |
| IR |
Infrared |
| LAEP |
Low-amplitude early peak |
| LALP |
Low-amplitude late peak |
| MALP |
Medium-amplitude late peak |
| MHS |
Microwave humidity sounder |
| MJO |
Madden–Julian oscillation |
| NASA |
National Aeronautics and Space Administration |
| NOAA |
National Oceanic and Atmospheric Administration |
| PMW |
Passive microwave |
| PPS |
Precipitation Processing System |
| SSM/I |
Special Sensor Microwave Imager |
| SSMIS |
Special Sensor Microwave Imager/Sounder |
| SST |
Sea surface temperature |
| Tb |
Brightness temperature |
| TMI |
TRMM Microwave Imager |
| TPW |
Total precipitable water |
| TRMM |
Tropical Rainfall Measuring Mission |
| XCAL |
Intersatellite calibration |
| 2mT |
2-m temperature |
REFERENCES
Adler, R. F., and Coauthors, 2003: The Version-2 Global Precipitation Climatology Project (GPCP) monthly precipitation analysis (1979–present). J. Hydrometeor., 4, 1147–1167, https://doi.org/10.1175/1525-7541(2003)004<1147:TVGPCP>2.0.CO;2.
Adler, R. F., J. J. Wang, G. Gu, and G. J. Huffman, 2009: A ten-year tropical rainfall climatology based on a composite of TRMM products. J. Meteor. Soc. Japan, 87A, 281–293, https://doi.org/10.2151/jmsj.87A.281.
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, 84–99, https://doi.org/10.1175/JAMC-D-11-052.1.
Aires, F., C. Prigent, F. Bernardo, C. Jiménez, R. Saunders, and P. Brunel, 2011: A Tool to Estimate Land–Surface Emissivities at Microwave frequencies (TELSEM) for use in numerical weather prediction. Quart. J. Roy. Meteor. Soc., 137, 690–699, https://doi.org/10.1002/qj.803.
Backus, G., and F. Gilbert, 1970: Uniqueness in the inversion of inaccurate gross Earth data. Philos. Trans. Roy. Soc., A266, 123–192, https://doi.org/10.1098/rsta.1970.0005.
Beck, C., J. Grieser, and B. Rudolf, 2005: A new monthly precipitation climatology for the global land areas for the period 1951 to 2000. Klimastatusbericht, 2004, 181–190.
Berg, W., 2016: GPM GMI_R common calibrated brightness temperatures collocated L1C 1.5 hours 13 km V05. Goddard Earth Sciences Data and Information Services Center, accessed 1 March 2019, https://doi.org/10.5067/GPM/GMI/R/1C/05.
Berg, W., 2017: Towards developing a long-term high-quality intercalibrated TRMM/GPM radiometer dataset. 2017 IEEE Int. Geoscience and Remote Sensing Symp. (IGARSS), Fort Worth, TX, IEEE, 248–250, https://doi.org/10.1109/IGARSS.2017.8126941.
Berg, W., T. L’Ecuyer, and C. D. Kummerow, 2006: Rainfall climate regimes: The relationship of regional TRMM rainfall biases to the environment. J. Appl. Meteor. Climatol., 45, 434–454, https://doi.org/10.1175/JAM2331.1.
Berg, W., and Coauthors, 2016: Intercalibration of the GPM microwave radiometer constellation. J. Atmos. Oceanic Technol., 33, 2639–2654, https://doi.org/10.1175/JTECH-D-16-0100.1.
Berg, W., R. Kroodsma, C. D. Kummerow, and D. S. McKague, 2018: Fundamental climate data records of microwave brightness temperatures. Remote Sens., 10, 1306, https://doi.org/10.3390/rs10081306.
Copernicus Climate Change Service, 2017: ERA5: Fifth generation of ECMWF atmospheric reanalyses of the global climate. Copernicus Climate Change Service Climate Data Store, accessed 10 January 2019, https://cds.climate.copernicus.eu/cdsapp#!/home.
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, https://doi.org/10.1002/qj.828.
Elsaesser, G. S., and C. D. Kummerow, 2015: The sensitivity of rainfall estimation to error assumptions in a Bayesian passive microwave retrieval algorithm. J. Appl. Meteor. Climatol., 54, 408–422, https://doi.org/10.1175/JAMC-D-14-0105.1.
Hou, A. Y., and Coauthors, 2014: The Global Precipitation Measurement mission. Bull. Amer. Meteor. Soc., 95, 701–722, https://doi.org/10.1175/BAMS-D-13-00164.1.
Huffman, G. J., and Coauthors, 2020: Integrated Multi-satellite Retrievals for the Global Precipitation Measurement (GPM) Mission (IMERG). Satellite Precipitation Measurement, V. Levizzani et al., Eds., Advances in Global Change Research, Vol. 67, Springer, 343–353, https://doi.org/10.1007/978-3-030-24568-9_19.
Joyce, R. J., J. E. Janowiak, P. A. Arkin, and P. Xie, 2004: CMORPH: A method that produces global precipitation estimates from passive microwave and infrared data at high spatial and temporal resolution. J. Hydrometeor., 5, 487–503, https://doi.org/10.1175/1525-7541(2004)005<0487:CAMTPG>2.0.CO;2.
Kidd, C., T. Matsui, and S. Ringerud, 2021: Precipitation retrievals from passive microwave cross-track sensors: The Precipitation Retrieval and Profiling Scheme. Remote Sens., 13, 947, https://doi.org/10.3390/rs13050947.
Knapp, K. R., and Coauthors, 2011: Globally gridded satellite observations for climate studies. Bull. Amer. Meteor. Soc., 92, 893–907, https://doi.org/10.1175/2011BAMS3039.1.
Kummerow, C. D., and L. Giglio, 1994: A passive microwave technique for estimating rainfall and vertical structure information from space. Part I: Algorithm description. J. Appl. Meteor., 33, 3–18, https://doi.org/10.1175/1520-0450(1994)033<0003:APMTFE>2.0.CO;2.
Kummerow, C. D., S. Ringerud, J. Crook, D. Randel, and W. Berg, 2011: An observationally generated a priori database for microwave rainfall retrievals. J. Atmos. Oceanic Technol., 28, 113–130, https://doi.org/10.1175/2010JTECHA1468.1.
Kummerow, C. D., D. L. Randel, M. Kulie, N.-Y. Wang, R. Ferraro, S. Joseph Munchak, and V. Petkovic, 2015: The evolution of the Goddard profiling algorithm to a fully parametric scheme. J. Atmos. Oceanic Technol., 32, 2265–2280, https://doi.org/10.1175/JTECH-D-15-0039.1.
Masunaga, H., M. Schröder, F. A. Furuzawa, C. Kummerow, E. Rustemeier, and U. Schneider, 2019: Inter-product biases in global precipitation extremes. Environ. Res. Lett., 14, 125016, https://doi.org/10.1088/1748-9326/ab5da9.
Okamoto, K., T. Iguchi, N. Takahashi, K. Iwanami, and T. Ushio, 2005: The Global Satellite Mapping of Precipitation (GSMaP) project. IGARSS’05: Proc. 2005 IEEE Int. Geoscience and Remote Sensing Symp., Seoul, South Korea, IEEE, 3414–3416, https://doi.org/10.1109/IGARSS.2005.1526575.
Olson, W., 2017: GPM DPR and GMI combined precipitation L2B 1.5 hours 5 km V05. Goddard Earth Sciences Data and Information Services Center, accessed 1 March 2019, https://doi.org/10.5067/GPM/DPRGMI/CMB/2B/05.
Petković, V., and C. D. Kummerow, 2017: Understanding the sources of satellite passive microwave rainfall retrieval systematic errors over land. J. Appl. Meteor. Climatol., 56, 597–614, https://doi.org/10.1175/JAMC-D-16-0174.1.
Petković, V., M. Orescanin, P. Kirstetter, C. Kummerow, and R. Ferraro, 2019: Enhancing PMW satellite precipitation estimation: Detecting convective class. J. Atmos. Oceanic Technol., 36, 2349–2363, https://doi.org/10.1175/JTECH-D-19-0008.1.
Romanov, P., G. Gutman, and I. Csiszar, 2000: Automated monitoring of snow cover over North America with multispectral satellite data. J. Appl. Meteor., 39, 1866–1880, https://doi.org/10.1175/1520-0450(2000)039<1866:AMOSCO>2.0.CO;2.
Schneider, U., T. Fuchs, A. Meyer-Christoffer, and B. Rudolf, 2008: Global precipitation analysis products of the GPCC. GPCC DWD Publ. 112, 17 pp., https://opendata.dwd.de/climate_environment/GPCC/PDF/GPCC_intro_products_lastversion.pdf.
Schneider, U., M. Ziese, A. Meyer-Christoffer, P. Finger, E. Rustemeier, and A. Becker, 2016: The new portfolio of global precipitation data products of the Global Precipitation Climatology Centre suitable to assess and quantify the global water cycle and resources. IAHS Publ., 374, 29–34, https://doi.org/10.5194/piahs-374-29-2016.
Schneider, U., P. Finger, A. Meyer-Christoffer, E. Rustemeier, M. Ziese, and A. Becker, 2017: Evaluating the hydrological cycle over land using the newly-corrected precipitation climatology from the Global Precipitation Climatology Centre (GPCC). Atmosphere, 8, 52, https://doi.org/10.3390/atmos8030052.
Skofronick-Jackson, G., and Coauthors, 2017: The Global Precipitation Measurement (GPM) mission for science and society. Bull. Amer. Meteor. Soc., 98, 1679–1695, https://doi.org/10.1175/BAMS-D-15-00306.1.
Tapiador, F. J., and Coauthors, 2017: Global precipitation measurements for validating climate models. Atmos. Res., 197, 1–20, https://doi.org/10.1016/j.atmosres.2017.06.021.
Watters, D., A. Battaglia, and R. P. Allan, 2021: The diurnal cycle of precipitation according to multiple decades of global satellite observations, three CMIP6 models, and the ECMWF reanalysis. J. Climate, 34, 5063–5080, https://doi.org/10.1175/JCLI-D-20-0966.1.
Xie, P., and P. A. Arkin, 1997: Global precipitation: A 17-year monthly analysis based on gauge observations, satellite estimates, and numerical model outputs. Bull. Amer. Meteor. Soc., 78, 2539–2558, https://doi.org/10.1175/1520-0477(1997)078<2539:GPAYMA>2.0.CO;2.
