An updated version of the Goddard Profiling Algorithm (GPROF 2014) with a new overland scheme was released with the launch of the Global Precipitation Mission (GPM) core satellite in February 2014. The algorithm is designed to provide consistent precipitation estimates over both ocean and land across diverse satellite platforms. This study tests the performance of the new retrieval, focusing specifically on an extreme rainfall event. Two contrasting 72-h precipitation events over the same area are used to compare the retrieved products against ground measurements. The first event is characterized by persistent and intense precipitation of an unusually strong and widespread system, which caused historical flooding of the central Balkan region of southeastern Europe in May 2014. The second event serves as a baseline case for a more typical midlatitude regime. Rainfall rates and 3-day accumulations given by five conically scanning radiometers (GMI; AMSR2; and SSMIS F16, F17, and F18) in the GPM constellation are compared against ground radar data from the Operational Program for Exchange of Weather Radar Information (OPERA) network and in situ measurements. Satellite products show good agreement with ground radars; the retrieval closely reproduces spatial and temporal characteristics of both events. Strong biases related to precipitation regimes are found when satellite and radar measurements are compared to ground gauges. While the GPM constellation performs well during the nonextreme event, showing ~10% negative bias, it underestimates gauge accumulations of the Balkan flood event by 60%. Analyses show that the biases are caused by the differences between the expected and observed ice-scattering signals, suggesting that better understanding of the environment and its impact on rain profiles is the key for successful retrievals in extreme events.
1. Background and introduction
Observations of precipitation are an important focus of water resource management. According to the Fifth Assessment Report of IPCC Working Group I, observations and model-projected future changes both indicate increases in extreme precipitation associated with climate change. This is supported by analysis of observed annual maximum 1-day precipitation that indicates a significant increase in extreme precipitation globally, with a median increase of approximately 7% per 1°C of global mean surface temperature increase (Westra et al. 2013). Expectations are that higher moisture content in the atmosphere leads to stronger extreme precipitation as extreme precipitation typically scales with total column moisture. These projections, together with consideration of direct (destruction, floods, etc.) and indirect (contamination, diseases, damaged infrastructure) effects of extreme precipitation, make their detection a priority in hydrometeorological observations.
Today, the WMO as well as national agencies utilize all available resources in an effort to provide the best possible estimates of rain and snow accumulations. (See Table 1 for a full list of acronyms used throughout the paper.) Satellite products play an integral role in this scheme, particularly in areas that are not well instrumented. Relying largely on passive microwave measurements, significant challenges exist because of poor temporal sampling and the inability of land retrievals to correctly address these extreme events over the areas where they are of the greatest interest. The launch of the GPM core satellite (Hou et al. 2014) offers new potential in precipitation observations. Better reference precipitation from GPM’s dual-frequency radar, together with increased temporal sampling provided by GPM constellation satellites offer great potential for capturing extreme events. New 3-hourly observations are expected to contribute toward improving the existing rainfall accumulations and accompanying flood warning systems. To meet these expectations, highly accurate rainfall retrievals are needed with sufficient temporal sampling over extended regions. A significant challenge remains to test whether the sometimes limited information content of the passive microwave radiometers can properly retrieve rainfall rates associated with a broad spectrum of atmospheric conditions.
To better understand the challenges of retrieving precipitation over land from passive microwave algorithms, aspects of the physical basis and architecture of the retrieval is reviewed. The GPM operational passive microwave rainfall retrieval (GPROF 2014; Kummerow et al. 2015) was released after the launch of the GPM core satellite in February 2014. It is a Bayesian retrieval that utilizes ground radar observations to relate observed brightness temperatures Tb to surface rainfall rates. To accomplish this, an a priori database was created of coupled NMQ (Zhang et al. 2011) precipitation rates and corresponding Tb (for each instrument in the GPM constellation). The retrieval uses Bayes’ theorem to relate the Tb observation vector y to the rainfall profile x if an a priori distribution of x is known:
where Pr(x) is the probability of observing a certain rain profile x and Pr(y | x) is the probability of observing the y vector for a given profile x. This can also be cast in the framework of an “expected value” problem as shown in Eq. (2), where the distances between the observed and database Tb vectors are used to assign weight to individual database entries. This approach allows physical and consistent retrievals across different sensors, providing a statistically expected rainfall rate x:
where xi represents all database profiles; y is the observation vector; H(xi) is the simulated observation vector corresponding to profile xi, with H being the observation operator; and are the observation and model error covariance matrices, respectively; and A is a scalar constant serving as the normalization factor (Kummerow et al. 2001). While this is usually advantageous for the purpose of generating robust rainfall estimates, it may not be optimal for extreme rain events that are not well represented in the a priori Pr(x) database. The problem of correctly retrieving extremes from a Bayesian methodology therefore depends on the extent to which the scheme can identify the correct extreme cases in the database and to which the scheme does not overly average database entries toward the mean of the distribution. Both of these factors depend on the information content of the Tb signal itself.
While the radiative properties of water are well understood, the signal is dramatically different over land and ocean. Over the oceans, a low-emissivity surface provides significant contrast between the radiatively cold background and warm precipitation signatures. Unfortunately, this is not the case over the land where high emissivity of the surface and its large variability mask atmospheric emission signatures and make precipitation nearly indistinguishable from the background. To overcome this problem, passive microwave retrievals over land focus on ice-scattering signals, which are less well related to precipitation but more easily detected over a warm surface background. This limited information content weakens the linkage between satellite measurements and the a priori database and exposes the algorithm to excessive averaging. The key for a successful rainfall rate estimate over land is therefore a good understanding of the relationship between the amount of ice in the cloud and surface rainfall rate. Currently, the a priori database is constructed using one year of ground radar observations from the U.S. NEXRAD network matched to Tb from satellite overpasses. The database is stratified by TPW and land surface temperature obtained from global reanalysis data. This provides a broad set of profiles for retrieving a wide range of rainfall rates, including those above 50 mm h−1, but it implicitly imposes a mean relationship between ice aloft and surface precipitation that is representative of the continental United States. Strong evidence exists, however, that the ratio of ice to rain is regionally dependent and sensitive to environmental forcing, such as atmospheric stability, which is often far from average in extreme precipitation events.
The primary goal of this study is to explore GPM’s current passive microwave retrieval (GPROF 2014, versions 1–4) performance in an extreme precipitation event and to provide a deeper understanding of its potential in the case of extreme events. As a second goal, the study seeks to quantitatively evaluate the differences between satellite retrieval and ground measurements in extreme precipitation conditions to gain a better understanding of their relationship.
A case of the Balkan floods of 2014 (described in following section) is seen as a perfect example of an event with microphysics that does not correspond to typical conditions although its rain rates are well within the database range. This makes it ideal for performing a comparison between retrieved satellite accumulations and ground references to provide information on GPROF skills under nontypical atmospheric conditions. Available gauge and ground radar networks over this region (OPERA; see section 3) allow for understanding of how each dataset interprets the flood event.
2. Events description
This study chose two 3-day rainfall events over the Balkan region. An extreme event and a more typical Balkan rain event were selected.
a. Flood event (14–16 May 2014)
The extreme precipitation event (hereafter called “Balkan flood” or “flood event”) focuses on a 3-day period that started at 0000 UTC 14 May 2014; spread over a significant area (see Fig. 1) in the central Balkan region; and affected the countries of Serbia, Bosnia and Herzegovina (BiH), and Croatia. During this event, historical readings at both rain and river gauges were recorded throughout the region that was devastated by floods and mudslides. A number of fatalities and more than 2 billion U.S. dollars in damage1 were directly caused by this flood event. The area itself contains two catchments: 1) the Adriatic, which is the southernmost 15% of the box shown in Fig. 1 that did not flood; and 2) the Black Sea, which is the upper 85% of the box with two major rivers (Danube and Sava) and a dozen regional-size basins where the majority of flooding occurred. The study area (42°–47°N, 15°–21°E) is defined by the flood region and its Black Sea upstream catchment. It ranges in elevation from 100 m in the predominantly flat plains to the north to 1700 m in steep elevated terrain to the south of 45°N latitude.
1) The synoptic environment and climatology
Two weeks prior the Balkan flood event, the region experienced localized thunderstorms (1–3 May) and the passage of a cyclone from the western Mediterranean (5 May) that brought light-to-moderate rain to the region. A warm air mass and higher temperatures followed on 7–12 May, which contributed to an increase in melting of the remaining snow accumulation in the mountains. This consequently resulted in increased river flows and moderately saturated soil. The period from 12 to 14 May was characterized by clear and warm weather during which a large area, including most of central and southeastern Europe, experienced a drop in the geopotential height. The trough formation at the surface was followed by a surface cutoff low on 15 May over the central Balkans, which was accompanied by an extremely strong and wide 500-mb level depression whose center was positioned over the central and western Balkans. The slow moving low had its center close to the flood region, causing intense precipitation and a decrease in temperature in the period from 14 to 18 May. A strong positive west–east gradient of specific humidity at 850 mb was present throughout the event, while the 500-mb values remained uniform and relatively low. The majority of precipitation occurred between 14 and 17 May with rain at lower elevation and snow generally above 1200 m. While a mean monthly temperature over the area was only 1°C below the climatological mean, the event mean temperature was 7°–8°C lower.
2) Snow, rain, and river gauge readings
The majority of rain and river gauges in the area of interest recorded historical levels. Climatological monthly means were exceeded by 130%–400% at all rain gauges in central and western Serbia and central and north BiH while 3-day accumulations exceeded the long-term mean May monthly values. Five climatological rain gauges in Serbia reported historical levels for 24-h accumulations. The greatest 3-day accumulation at any station was 209 mm, while most of the area received approximately one-fourth of its average annual precipitation in only 72 h. According to Gumbel’s method (Gumbel 1958; Wolf 1966) of extreme value distribution, the 3-day rainfall accumulation had a return period ranging between 100 and 1000 years throughout most of the flood region. Major rivers (i.e., the Sava, Drina, and Bosna) reached 200%–300% of their monthly climatological marks. On 15 May the maximum temperature at elevations above 1500 m remained below 0°C, with some stations exceeding 60 cm of new snow accumulations (e.g., Kopaonik Mountain, 61 cm) during the 24-h time interval.
b. Nonflood event (1–3 May 2014)
A 3-day period (hereafter called the nonflood event) starting at 0000 UTC 1 May 2014 is used to study more average conditions over the same 5° × 6° area defined in the case of the flood event. The event was chosen to be 1) a 72-h period of frequent precipitation, 2) within proximity (a couple weeks) to the flood event, and 3) characterized by typical precipitation regimes for the season. This provides a strong contrast in rain intensity, environment, and system structures (microphysics and thermodynamics) between the two events.
The synoptic environment and climatology
On 1 May, at 500 mb, the formation of a trough over central Europe occurred as a consequence of an easterly propagating wave that simultaneously formed a strong ridge aligned with the eastern Atlantic coast. As the trough deepened, a cutoff low occurred over the Apennine Peninsula on 2 May. This was followed by the formation of a surface low and cyclonic circulation over most of the Balkan region on 3 May. Under the influence of a cold air mass advected from the north-northwest and moisture coming from the southwest, the region experienced scattered precipitation within the first 48 h, followed by rain produced by a more organized midlatitude system on 3 May. In contrast to the flood event, no particularly strong gradients in specific humidity were observed. Daily rain accumulations and temperature values were close to their climatological values, with neither moisture nor the persistence of the low being atypical for this time of the year.
The study utilizes remote and in situ rainfall measurements from gauge and radar networks as well as the GPM constellation of satellite radiometers. Each dataset is available in near–real time and, with the exception of OPERA products, is used in wide spectrum of applications.
Within EUMETNET, OPERA and its radar data center have been in operation producing network-wide radar mosaics from volumetric data since 2011 (Huuskonen et al. 2014). The radar network spreads over most of Europe and exploits more than 130 weather radars of different types and frequencies distributed in 21 European countries.
This study utilizes OPERA’s near-surface rainfall rate and maximum reflectivity products, both provided on 2 km × 2 km grid at 15-min temporal resolution. Within the flood region these composite fields are formed by combining measurements from five Doppler radars (see Table 2) in and close to the region, working at C- and S-band single polarized frequencies. Their coverage is somewhat limited because of beam blockage along the mountain range in the southern flood region but still accounts for approximately 90% of the flood catchment area (the Black Sea catchment). In Fig. 1, black shading depicts radars range while regions suffering from terrain beam blockage are shown in gray.
Currently, the OPERA network rain rates are based on the traditional Marshall–Palmer reflectivity–rain rate (Z–R) relationship only (Marshall and Palmer 1948; Z = 200R1.6). Despite the potential lack of fidelity in representing local gauge accumulation, this provides consistency across different regionally operated systems and allows for better understanding of comparison results when the product is compared against other independent datasets. It is worth mentioning that a number of storm-specific factors may contribute to overall uncertainty of the OPERA dataset. Some of them include Z–R variability within the storm, the radar signal attenuation, and the highly variable precipitation field (Moszkowicz et al. 1994; Krajewski et al. 2003; Miriovsky et al. 2004; Lee and Zawadzki 2005; Berne and Andrieu 2005). Although addressing this problem is among the top priorities of the OPERA project (Sandford and Gaussiat 2011), because of the amount of processed data and radar-type variability across the network, in its current stage the OPERA project does not quantify contributions from these error sources. However, it is expected that overall uncertainty is largely dominated by the assumption of a constant Marshall–Palmer DSD. To ensure the best possible quality, OPERA data in this study are filtered using accompanying quality-control flags, which resulted in data sample reduction of approximately 10% but increased the confidence of ground clutter removal. Missing pixels in the radar dataset are replaced by interpolating between the closest available (in time) measurements at the given grid point. This resulted in negligible changes in the results (less than 0.1% of rain accumulation over the domain). During manual inspection of the dataset, LAN interference, known to be often present in OPERA products (Lopez 2014), is noticed for one radar (Maly Javornik) in Slovakia. These spurious retrievals are replaced by interpolating values of interference-free, time-adjacent pixels. In most cases, the adjacent pixels had a value of zero that overall resulted in negligible changes to the 3-day rainfall accumulations.
b. Satellite data: GPROF 2014
Five conically scanning sensors in the GPM constellation [AMSR2 (Shimoda 2005), GMI (Hou et al. 2014), and SSMIS on board the F16, F17, and F18 satellites operated by the U.S. Air Force DMSP (Kunkee et al. 2008)] are utilized to provide satellite data. Rainfall rates from 29 overpasses during the flood event and 28 overpasses during the average, nonflood event over the region of interest are obtained from GPROF 2014 passive microwave retrieval at spatial resolution corresponding to 37 GHz of each sensor’s FOV with an average temporal sampling of the event of approximately 2.1 h. Based on the observations of the flood event and simulated channel and forward model errors, retrieval error is estimated to be less than 3% for applications in this study.
c. Surface gauge data
Surface 24-h rain accumulations from 25 rain gauges located within the flood region are used as a ground reference in this study. Table 3 lists gauge accumulations as given in surface synoptic observations (SYNOP) reports while their locations are shown in Fig. 2 (blue triangles). Although rain gauge observation is generally seen as the most accurate precipitation measurement, the complexity and amplitude of its error requires rigorous quality control before data can be used for scientific purposes. SYNOP rain gauge data can be affected by both systematic and random errors (Lopez 2013). Systematic errors are usually linked to raindrops splashing, high-wind conditions, wetting of the gauge walls, and loss through evaporation. Random errors are mainly caused by the discrete nature of the time sampling and by small-scale variations of the turbulent airflow around the gauge. In this study all chosen gauges are maintained by the National Hydrological and Hydrometeorological Services, undergoing the standard quality-control procedures recommended in terms of the latest global standard by the WMO. The selected subset of the gauge network is chosen to ensure the best-quality confidence and spatiotemporal coverage over the region. During the extreme event, meteorological conditions were such that significant errors are not expected (e.g., no high winds are detected throughout the region while precipitation was continuous). On the other hand, the isolated convection seen during the nonflood event may have easily resulted in significant random errors in SYNOP reports. These errors are, however, expected to be relatively small in the 24-h accumulation because of the averaging over 4–8 individual measurements (Ciach 2003). No significant differences in surrounding rain gauges (not included in this study) are reported for the flood event (no such report is expected for the nonflood event because of its mediocrity). Three- to six-hourly gauge readings are verified using monthly reports of national weather centers in the region and by comparing the reports from surrounding rain gauges (non-SYNOP ones). Missing data are found at one gauge location (Banja Luka, WMO ID 14542) for the first day of the flood event, and therefore this gauge is used in daily accumulation analysis only.
The ECA&D E-OBS daily rainfall dataset (Haylock et al. 2008; Lockhoff et al. 2014) is used for additional OPERA quality control. This high-resolution gridded precipitation product is based on combined monthly and daily gauge estimates using various interpolation and kriging techniques. Although only negligible differences are seen between this product and 24-h gauge accumulation values from the 25 SYNOP gauges used for this study, potential errors that may emerge from interpolation techniques (Haylock et al. 2008; Kirstetter et al. 2010) do not qualify this dataset for direct comparison with satellite and radar products. Techniques to overcome these problems exist (Lockhoff et al. 2014) but unfortunately do not apply to the Balkan region and quality of available data. Therefore, E-OBS dataset is used here only in the process of removing OPERA radar ground clutter and LAN interference (section 3a).
4. Dataset inspection
To allow inspection of each dataset and perform their intercomparisons, both satellite and radar data are first uniformly gridded into 0.2-km grids over the study area at the times of observations. Rainfall is then accumulated at each grid point, assuming that rain rates are constant between available measurements. To match OPERA’s temporal sampling, the time of the satellite overpass is rounded to the closest 15 min. This resulted in a virtual overlap of two satellite overpasses (GMI and F16 at 0400 UTC 16 May) that in reality were 6 min apart. Comparison between satellite and radar accumulations are made only where satellite measurements exist.
a. Ground radar to gauge comparisons
Gauge network observations of both the extreme and average 3-day raining events (described in section 2) are used to evaluate remotely sensed products. Despite the availability of close-to-instantaneous measurements at a number of the gauge locations (e.g., tipping-bucket measurements), their direct comparison to satellite estimates would be sparse and highly sensitive to spatial variability of the rainfall field and random errors. Therefore, rather than comparing satellite FOV to the gauge point measurements directly, a two-step approach is used. In the first step, ground radars are compared against collocated gauge measurements. In the second step (described in section 4b), satellite estimates are compared to ground radars.
Figure 3 shows 72-h accumulation maps for the extreme (14–16 May) and nonextreme or average (1–3 May) 3-day events (Figs. 3a and 3b, respectively). The color scale depicts radar estimates across the region, while gauge readings are labeled next to the station locations (details given in Table 3). Scatterplots of daily accumulations over the area for the same two events are given in Fig. 4.
Analyses show that the ground radars tend to underestimate gauge accumulations by a factor of 2 during the flood event but overestimate the same gauges by a factor of 1.3 during the average nonflood event. Similar results for May 2012 and May 2014 are reported in Lopez (2014), who compared the OPERA monthly mean products against ground gauges in this region. If rain gauges are treated as more accurate measurements, the fact that OPERA radar retrieval assumes constant particle size distribution (i.e., Marshall–Palmer with Z = 200R1.6) during both events implies that the difference in the precipitation regimes (e.g., DSDs) is the key variable to alter the radar to gauge ratio. This is further corroborated by the fact that radar to gauge biases are relatively constant over the wide range of rain rates during the flood event characterized by the tropical-like environment, as evident from the mean 0000 and 1200 UTC Belgrade soundings shown in Fig. 5. Conversely, variations in the bias on daily scales are highly correlated with precipitation type during the average nonflood event. In Fig. 4b during the first two days, the bias is opposite to that found on the third day, following the change in the precipitation regime from isolated intense storms to a more organized mesoscale convection (see section 2). Studies such as Petersen et al. (1999) and Cifelli et al. (2011) provide detailed understanding of drawbacks related to the use of radar products that rely on an average Z–R relationship (200R1.6) in tropical-like environment conditions seen during the Balkan floods event and a number of Front Range flooding events in Colorado. Therefore, based on 25 gauges collocated with radar measurements, an adjusted Z–R relationship for each 24-h interval of the two events was calculated (see Table 4) and used to form the gauge-adjusted OPERA estimates. For simplicity, this calculation keeps b, the exponent of the original Marshall–Palmer Z–R relationship (Z = aRb), constant. Significantly lower values of coefficient a during the flood event (Table 4) imply that the two events were characterized by different precipitation regimes. Sharma et al. (2009) show that an increase of the coefficient a is associated with transitioning from stratiform to convective regimes. Also, the coefficients of adjusted Z–R relationship during the flood event are similar to those used in Gochis et al. (2015) for tropical-like environment (Fig. 5) to match the Front Range flooding regimes in Colorado.
b. Satellite constellation to ground radar comparisons
Five microwave imagers made 29 overpasses over the region during the 3-day extreme precipitation event (14–16 May) and 28 overpasses during the regular nonflood event (1–3 May). Visual inspection of each overpass rain-rate field in its native spatial resolutions provided qualitative comparisons between satellite and ground radar estimates. An example is given in Fig. 6 where corresponding measurements of the OPERA network (Fig. 6, top) and the GMI (Fig. 6, bottom) sensor for 2015 UTC 15 May 2014 are shown. Black-shaded regions are the same as in Fig. 1, while the satellite swath is shaded in gray and outlined by a black and yellow line.
If one focuses on the northern half of the flood area to ensure the best performance of the OPERA radars and to avoid mountainous regions in the south, one can conclude that satellite and radar estimates match very well. Both precipitation system size and distribution of rain rates are in good agreement between the instruments, despite the fact that the two products come from sensors that utilize different vertical sampling, with satellite estimates based on atmospheric integrated column properties at a relatively large slant angle in contrast to radar beam volume sampling. Slant angle of the GMI radiometer results in some displacement of the precipitation features edges and convective cores but within expected ranges. A more detailed examination that includes regions outside the flood box, however, reveals that GMI often misses light rain rates (<0.2 mm h−1) as well as intense precipitation cores associated with isolated convection. The radiometer’s insensitivity to light rain is expected since the low rain-rate scenes are expected to be radiometrically very similar to the nonraining ones, while the underestimation in the deep convective cores is likely due to Bayesian averaging. Examples of side-by-side comparisons between the other four sensors and OPERA data focusing on the flood region only during the extreme event are shown in Fig. 7. Except for the higher spatial resolution of the GMI and AMSR2 sensors relative to the DMSP sensors, the same conclusions as in the GMI case hold. Overall, GPROF 2014 shows qualitatively good performance in capturing the spatial variability of the rainfall field when compared to ground radar measurements. The same conclusion was made upon examination of the nonextreme event (not shown here).
Quantitative comparisons between satellite- and ground-based measurements are given in Figs. 8 and 9. Figure 8 depicts timelines of the area mean total accumulation over the flood region at 15-min temporal resolution for each dataset during the flood (Fig. 8a) and nonflood event (Fig. 8b). Satellite observations (black line) are compared to 1) collocated ground radar measurements (red line), 2) ground radar measurements over the entire flood region using OPERA’s native (15 min) temporal resolution and a Z = 200R1.6 relationship (yellow line), and 3) gauge-adjusted radar measurements using 24-h Z–R relationships given in Table 4 (green line). The red line represents simulated satellite observations given by the ground radars, while the green line serves as a reference for “truth” for the size and scale of the events over the entire region of interest. The differences between the OPERA overpass match (red line) and OPERA native resolution (yellow line) exist because of their different spatial extent. The two lines address slightly different areas, being more apart when the satellite’s FOV captures only a small portion of the flood region.
It is evident from Fig. 8 that the radar observations at the satellite overpass times (red) are almost identical to 15-min radar full coverage over the region (yellow). This indicates that the constellation provides sufficient spatial and temporal sampling to capture all of the rainfall variability over the region for this event. The fact that satellite accumulation (black) in both events fairly closely follows in shape the gauge-adjusted accumulation (green), with a correlation of 0.99 between the two, indicates that the constellation’s sampling and performance is capable of addressing a life cycle of both extreme and nonextreme events. To support previous conclusions and point to the origin of the observed differences in rainfall accumulations, detailed comparisons of area mean rain rates are performed and presented in Fig. 9.
Studies that compare radar and radiometer products indicate that the differences in their rain-rate estimates stem primarily from the assumptions the algorithms use to relate the observed quantities to rainfall rates. As mentioned earlier, the satellite retrieval is built upon the observed ice-scattering signal. Similarly, OPERA radar rain-rate estimates rely on the Marshall–Palmer Z–R relationship. Therefore, to the first order, differences in the magnitude of the rain rate between the satellite and ground radar retrievals can be caused by 1) inappropriate choice of Z–R relationship in radar retrievals and 2) nonrepresentative storm structures populating the a priori database.
Based on Table 4 and Fig. 4, the fact that radar reflectivity is proportional to the sixth power of the rain drop diameter (Z ~ D6) suggests that the DSD was dominated by smaller drops (compared to Marshall–Palmer mean drop size) during the flood event and the third day of nonflood event. Consequently, radar retrieval underestimated gauge-adjusted rainfall accumulation in these well-organized systems, relating observed Z to lower rain rates. On the other hand, the first two days of the nonflood event, characterized by isolated intense convection, appear to consist of DSD with mean drop diameter somewhat larger than expected by the standard Marshall–Palmer. This difference in the DSD during the two events is likely linked to microphysical properties of isolated and organized convection regimes (Rosenfeld and Ulbrich 2003; Bringi et al. 2003). It is known that inappropriate hydrometeor profiles lead to underestimation by passive microwave radiometer algorithms (Kwon et al. 2008; Kubota et al. 2009; Ryu et al. 2012; Sohn et al. 2013; Shige et al. 2013, 2015; Taniguchi et al. 2013). While not proven for the events in this study, one can infer that the general findings related to precipitation systems organization hold, saying that more organized systems appear to have smaller drops and less pronounced ice-scattering signal for the same rain amount than the isolated convective storms of similar or greater top height. This would explain the satellite’s underestimates during the flood event and the last day of nonextreme event, both described as well-organized regimes with the presence of ice phase. However, the amplitude of this negative bias is related to both events’ intensity and a complex link between the environment and rainfall profiles in the database (contributions of the two are quantitatively described in section 4d). Figure 10 depicts the distribution of the a priori database profiles for the environmental conditions observed during the Balkan flood event over the vegetated surface type accounting for 60% of the area over 90% of the time. Colors represent the density of database entries within a given rain rate and Tb interval, while magenta and black crosses mark satellite-retrieved and ground-gauge-adjusted observed values, respectively. While the correct answers appear to be represented in the database, the algorithm favors lower rainfall rates, underestimating the observed rain by 50% (as seen in Fig. 8) for the given environmental conditions.
The mechanics of the Bayesian scheme itself account for two primary sources of bias: 1) a Bayesian pull toward the a priori database mean, which is more pronounced when the information content is low, and 2) incorrect or underrepresented microphysics in the a priori database relative to the extreme event. While the pull toward the database mean is expected by definition of the extreme event, the latter can be explained by the fact that, perhaps, the NEXRAD-based database builds on storms specific for the U.S. region with microphysics different from the one that took place during the Balkan flood event. In addition, the link between rainfall profiles and environmental conditions used to constrain the database (e.g., skin temperature and total column water vapor) may not be suitable for events such as Balkan flooding. Thus, regardless of the fact that the observed profiles exist in the database, the retrieval failed to recognize them as the most likely solution. Comparison of the observed Tb profiles with profiles that the a priori database links to the gauge-adjusted rain-rate values showed significant differences in ice-scattering signatures. Mismatch of approximately 10 K at high-frequency channels (i.e., 91 GHz) implied ~40% more ice in the column during the flood event than the database entries for the observed environmental conditions suggest. A more quantitative comparison between the bias contributors (i.e., Bayesian averaging vs microphysics) is given in the discussion section.
c. DPR to ground radar
To further address GPM’s microwave imager constellation potential, and to verify that precipitation regimes play a key role in the gauge to satellite discrepancies, DPR measurements are introduced (Seto et al. 2013). The DPR’s attenuation-based retrieval (Iguchi et al. 2000, 2009) adjusts its Z–R relationship to the observed precipitation regime through a number of steps that include effects of rain type, presence or absence of a bright band, and the phase state, all of which in essence relate a chosen Z–R relationship to the environment and thermodynamical processes of the system. It is worth mentioning that this algorithm, originally developed for the TRMM (Kummerow et al. 1998) Precipitation Radar (PR), is under an incremental development process for the use in the midlatitude rainfall systems. Thus, although adjustable, the initial Z–R relationship is likely biased toward tropical environment characteristic by smaller mean drop size (lower Z) and larger rain rate (high R) than typically seen in an average midlatitude regime.
Because of the relatively small sample size (only four overpasses during the flood period and five overpasses during the nonflood period), comparison between OPERA radar network and DPR products was extended beyond the flood region. Thus, DPR comparisons do not relate exclusively to the events but rather to a much larger, though from a synoptic perspective still similar, precipitation region. This broader comparison partially emphasizes the role that OPERA’s fixed Z–R relationship plays in defining the gauge to radar biases.
The comparisons of GMI and DPR to ground radar measurements over the OPERA domain for four (five) satellite overpasses for the extreme (average) event are shown in Fig. 11a (Fig. 11b), with the summary of pixel level analysis given in Table 5. Overall, the results show reasonably high correlation between the satellite and ground data, especially considering the fact that random satellite overpasses and differences in scanning geometry prevent the exact collocation between OPERA and DPR volumes. However, while DPR and OPERA reflectivities (not shown here) match very well in both events, their rainfall accumulations differ significantly. During the flood event the DPR to OPERA ratio is close to that of the gauge to OPERA, implying a good match between satellite and gauge data, but a similar ratio is also seen during the average event when DPR greatly overestimates gauges. A possible explanation may lie in the lack of the full dual-frequency impacts on the day-1 algorithm being used here.
A more detailed investigation of DPR precipitation profiles (using the GPM 2A.DPR product) addresses the contrast in precipitation regimes of the two events. According to DPR observations over the flood region only, the flood event is characterized by an average freezing-level height of 1700 m, a near-surface reflectivity ranging from 30 to 35 dBZ, and 95% of total rainfall being classified as stratiform. On the other hand, the stratiform portion of the total rain during the average event is 70%, with a mean freezing level at approximately 2700 m and near-surface reflectivity ranging from 27 to 32 dBZ. A high ratio of stratiform-to-convective rain and an increase in reflectivity below the bright band seen during the flood event are common for intense stationary flood-related regimes. Gochis et al. (2015) offer extensive analysis on the Great Colorado Flood of September 2013, showing precipitation profile signatures that are in many aspects similar to those presented here.
As suggested above, one can conclude that the radiometer underestimation of the flood event as well as the organized convection accumulations are caused by 1) a Bayes’ pull toward the a priori database mean and 2) a nonflood microphysics of the a priori database rainfall profiles. In the second case, the entire database appears to be biased toward different types of raining systems characterized by a different ice-to-rain relationship from the observed one. This scenario results in biased rain rates regardless of their value since the algorithm links observed Tb to rain-rate values that are not related to the scene being retrieved. For the first problem, on the other hand, the distribution of the database’s rain rate is such that the observed values are found only at its tail and thus, given the mathematics of the Bayesian retrieval, are underestimated.
Bayesian retrievals, by virtue of retrieving the expected value of the parameter, tend to gravitate toward the center of the distribution. This can cause a Bayesian scheme to underestimate extreme precipitation events. Having an incorrect a priori distribution of precipitation can likewise lead to over- or underestimation depending on biases in the a priori distribution. To investigate the contribution of Bayesian averaging versus errors in the a priori database upon the retrieval’s bias seen in the previous section, a synthetic set of rain profiles is generated from the database. This is accomplished by randomly selecting entries from the database with a given surface rain rate to match the observed (gauge adjusted) probability density function (PDF) of rain rates during the extreme event. These profiles thus have the correct PDF of rain for the extreme event while statistically preserving the microphysics of the database. The retrieval is then run using these synthetic observations, excluding the true answer, which is still contained in the entire database. Thus, any differences between the retrieved and original rain rates in this synthetic experiment must come from the Bayesian averaging. The comparison of gauge-adjusted and retrieved rain rates from this experiment is shown in Fig. 12. The retrieval underestimates the original values by 25%. The remaining bias in the retrieval, or roughly 45% of the overall bias, is therefore likely due to the structural differences between observed precipitation systems over the Balkan and those that built the NEXRAD database.
5. Conclusions and summary
The performance of the GPM passive microwave retrieval (GPROF 2014, versions 1–4) in an extreme precipitation event is tested to provide deeper understanding of its potentials and guidelines for its future development. Being an operational retrieval for GPM mission, GPROF serves multiple microwave imager sensors for both conical and cross tracking, with the goal of providing consistent precipitation observations at a wide range of regimes and scales across the globe. Analyzing a 3-day flood event that occurred in the central Balkan region, in addition to an average nonflood event of same duration, this study 1) provides qualitative and quantitative comparison of retrieval’s products against two sets of independent ground measurements and 2) offers some insight about the impact of regime-dependent microphysics upon these retrievals. The latter is likely to be the key for improving the accuracy of individual and combined satellite products.
The results show that the constellation of five cross-tracking sensors (AMSR2; GMI; and the SSMISs on board of F16, F17, and F18) provided sufficient sampling and coverage for the retrieval to closely reproduce rainfall rate and accumulation estimates given by ground radars. However, discrepancies between satellite, radar, and gauge 72-h rain accumulation estimates during the extreme precipitation event reveal that both satellite and ground radars underestimated accumulations relative to gauges by 60% and 50%, respectively. At the same time, relatively high correlations of 24-h accumulations (0.85) are seen between ground (OPERA) radars and gauges. Additional comparisons related to a more typical, nonextreme precipitation event of the same duration over the same area indicated satellite underestimate (12%) and radar overestimate (30%) relative to gauge accumulations. This ambiguous result is explained by OPERA’s exclusive use of Marshall–Palmer Z–R relationship (Z = 200R1.6), which is designed to represent midlatitude stratiform system DSD rather than DSDs related to tropical-like conditions seen during the Balkan extreme event. A similar explanation holds in the case of satellite retrieval where microphysics from typical rain events (built upon the NEXRAD observations) leave an inadequate link between rain rates and corresponding environmental conditions during this extreme event. While upcoming versions of the DPR retrieval will with no doubt implement midlatitude Z–R relationships, the role of the variability of the ice-to-rain ratio over the broad spectrum of microphysical and dynamical cloud properties related to specific environments remains to be carefully addressed in the future.
This research was supported by NASA Earth and Space Science Fellowship 2014 and PMM Grant NNX13AG31G. The authors wish to acknowledge Joshua King (CSU), as well as Milan Dacic (RHMSS/SEEVCCC), and Vladimir Djurdjevic (Institute of Meteorology, Faculty of Physics, University of Belgrade, Serbia) for their helpful comments and suggestions. Special thanks are extended to Dr. Elena Saltikoff (Finnish Meteorological Institute) and Sergio Pasquini (EUMETNET) for their help in providing access to OPERA data.
No official estimate available at the time of writing this paper.