Compared to traditional single-polarization radar, dual-polarization radar has a number of advantages for quantitative precipitation estimation because more information about the drop size distribution and hydrometeor type can be gleaned. In this paper, an improved dual-polarization rainfall methodology is proposed, which is driven by a region-based hydrometeor classification mechanism. The objective of this study is to incorporate the spatial coherence and self-aggregation of dual-polarization observables in hydrometeor classification and to produce robust rainfall estimates for operational applications. The S-band dual-polarization data collected from the NASA Polarimetric (NPOL) radar during the GPM Iowa Flood Studies (IFloodS) ground validation field campaign are used to demonstrate and evaluate the proposed rainfall algorithm. Results show that the improved rainfall method provides better performance than a few single- and dual-polarization algorithms in previous studies. This paper also investigates the impact of radar beam broadening on various rainfall algorithms. It is found that the radar-based rainfall products are less correlated with ground disdrometer measurements as the distance from the radar increases.
Building upon the success of Tropical Rainfall Measuring Mission (TRMM), the National Aeronautics and Space Administration (NASA) and Japan Aerospace Exploration Agency (JAXA) have embarked on the Global Precipitation Measurement (GPM) mission (Hou et al. 2014). The GPM Core Observatory satellite, carrying the first spaceborne Dual-Frequency Precipitation Radar (DPR) operating at Ku and Ka band and the GPM Microwave Imager (GMI), was launched on 27 February 2014. The active DPR and passive radiometer on board GPM extend the observation range attained by TRMM from tropics to most of the globe and provide accurate measurement of rainfall, snowfall, and other precipitation types. Through improved measurements of precipitation, the GPM mission is helping to advance our understanding of Earth’s water and energy cycle, as well as climate changes. As an indispensable part of the GPM mission, ground validation helps to develop the radar and radiometer retrieval algorithms by providing insight into the physical and statistical basis of precipitation (Chandrasekar et al. 2008). In the prelaunch era, several international validation experiments had already generated a substantial dataset that could be used to develop and test the prelaunch GPM algorithms. After launch, more ground validation field campaigns were conducted to further evaluate GPM data products as well as the sensitivities of retrieval algorithms. Among various validation equipment, ground-based dual-polarization radar has shown great advantages to conduct precipitation estimation over a wide area in a relatively short time span. Therefore, radar is always a key component in all the validation field experiments. In addition, polarization diversity has great potential to characterize precipitation microphysics through drop size distribution and to identify different hydrometeor types. From the operational point of view, a dual-polarization radar system can also provide measurements that are immune to absolute radar calibration and partial beam blockage. Currently, all the radar sites comprising the U.S. National Weather Service (NWS) Weather Surveillance Radar-1988 Doppler (WSR-88D) network, also known as Next-Generation Weather Radar (NEXRAD) network, are operating in dual-polarization mode. Therein, enhancement of quantitative precipitation estimation (QPE) is one of the main considerations for the dual-polarization upgrade.
In a number of previous studies, radar reflectivity at horizontal polarization Zh, differential reflectivity Zdr, and specific differential phase shift Kdp have been combined to derive rainfall products. For example, Chandrasekar et al. (1993) attempted to minimize the standard error of rainfall-rate estimation by selecting rainfall relations according to rainfall intensities. Cifelli et al. (2003) have developed a blended algorithm to estimate rainfall rate based on thresholds on the values of Zh, Zdr, and Kdp. Though there is no standard criterion that can be applied to determine which estimator to employ for a given set of dual-polarization measurements, the hydrometeor-classification-based rainfall algorithms have been fairly successful in recent years (e.g., Cifelli et al. 2005; Giangrande and Ryzhkov 2008). At Colorado State University (CSU), an optimization algorithm has been developed by Cifelli et al. (2011) using hydrometeor identification results to guide the choice of particular radar rainfall relations, namely, , , , and , where R indicates rainfall rate. This optimization algorithm is often referred to as CSU-HIDRO (Cifelli et al. 2011) or as the CSU Dual-Polarization (CSU-DP) algorithm (Seo et al. 2015). It is a key component of the Dual-Polarization Radar Operational Processing System (DROPS) developed at CSU. Therefore, the optimization algorithm in Cifelli et al. (2011) will be referred to as “DROPS1.0” in this paper. DROPS1.0 has been used in a number of previous studies. For instance, Cifelli et al. (2011) demonstrated the encouraging performance of DROPS1.0 in the high plains environment with the data collected from the S-band CSU–University of Chicago–Illinois State Water Survey (CSU-CHILL) radar and a network of rain gauges in Denver, Colorado. Pei et al. (2014) used the DROPS1.0 rainfall algorithm to study the impacts of raindrop fall speed and axis radio errors. Seo et al. (2015) showed that DROPS1.0 was superior to a single-polarization-based rainfall algorithm during the NASA GPM Iowa Flood Studies (IFloodS) field experiment in central and northeastern Iowa, especially for intense rainfall estimation. Chen and Chandrasekar (2015a) have implemented and evaluated DROPS1.0 in an urban environment in Dallas–Fort Worth, Texas, which showed better performance than both WSR-88D single- and dual-polarization rainfall products for the precipitation events presented in their study.
However, the bin-by-bin-based fuzzy-logic hydrometeor classification algorithm implemented in DROPS1.0 or any other hydrometeor-classification-based rainfall system (e.g., Cunha et al. 2015) may not be sufficient for operational applications, especially when the radar data quality is low. That is, for a given range gate, the hydrometeor classification result can be noisy if the input radar measurements are noisy since the information from adjacent gates is not used, which will severely affect the subsequent rainfall estimation. A smooth and clear output is desirable in an operational environment to ease the interpretation by the end users. In addition, the rainfall relations in DROPS1.0 are derived based on simulated drop size distribution (DSD) data (Cifelli et al. 2011), which may not be sufficient to represent real rainfall microphysical properties. This paper aims to improve DROPS1.0 for more accurate and robust rainfall estimation, especially in an operational environment. Compared with DROPS1.0, the improved dual-polarization radar rainfall algorithm (hereafter refer to as DROPS2.0) presented in this paper has incorporated a region-based hydrometeor classification methodology. In addition, the specific rainfall relations have been upgraded based on real DSD observations collected during the NASA IFloodS field campaign.
This paper also attempts to quantify rainfall errors introduced by radar beam broadening. Although the dual-polarization techniques provide us with better means for radar system calibration, data quality control, and rainfall estimation, the geometry of radar measurements combined with the variability of spatial distribution of precipitation still pose challenges. A number of studies have been devoted to the correction of range-dependent errors in rainfall estimates obtained from gridded radar reflectivity data (e.g., Chumchean et al. 2004). Nevertheless, the quantification of dual-polarization radar rainfall errors introduced by beam broadening and beam tilting in native radar polar coordinates is relatively rare. One of the challenges that limits such kinds of research is the ad hoc deployment of ground validation instruments (e.g., rain gauges or disdrometers) relative to radar. During NASA IFloodS field experiment, a variety of ground-based instruments have been deployed to collect high-quality in situ precipitation data. Among them, 14 autonomous particle size and velocity (Parsivel) unit (APU) disdrometers were deployed along the NASA Polarimetric (NPOL) S-band radar azimuthal radials at different ranges (W. Petersen and W. F. Krajewski 2016, unpublished manuscript), and tipping-bucket gauges were collocated with 10 APU disdrometers (see Fig. 1). The disdrometer locations, relative to NPOL radar, are listed in Table 1. This unique instrument layout provides us with an ideal environment to investigate the impact of beam broadening on radar rainfall estimation. Therefore, we take this opportunity to quantify rainfall errors of a few rainfall algorithms, namely, DROPS1.0, DROPS2.0, and the WSR-88D default (hereafter referred to as NEXRAD Z–R), as well as the dual-polarization rainfall relation proposed by Giangrande and Ryzhkov (2008) that is adopted by NEXRAD. Our intent is not to provide an exhaustive evaluation of various rainfall products for the whole campaign. Rather, the purpose is to demonstrate the performance of DROPS2.0 and to quantify the influence of radar beam broadening on the accuracy of rainfall-rate estimates in the context of IFloodS.
The remainder of this paper is organized as follows. Section 2 details the DROPS2.0 rainfall algorithm, including its error structure, as well as the region-based hydrometeor classification methodology implemented in DROPS2.0. Rainfall events and datasets used for demonstration purposes are presented in section 3, and section 4 provides the evaluation results, with particular attention paid to the radar–disdrometer rainfall comparisons with different estimators at different ranges. In section 5, we summarize the main findings of this paper.
2. DROPS2.0 rainfall algorithm
The architecture of DROPS2.0 rainfall algorithm is similar to DROPS1.0, which has been described extensively in Cifelli et al. (2011). In the following, we present important features of DROPS2.0, as well as the specific rainfall relations used by DROPS2.0.
The logic of DROPS2.0 is shown in Fig. 2, which includes three main steps:
data quality control and Kdp estimation,
region-based hydrometeor classification, and
In the data quality control step, the adaptive algorithm developed by Wang and Chandrasekar (2009) is implemented to estimate Kdp and to remove ground clutter and nonmeteorological echoes [see also Chen et al. (2017)]. Compared to traditional Kdp estimation methods (e.g., Hubbert and Bringi 1995), the adaptive technique in Wang and Chandrasekar (2009) does not smooth the peak Kdp by averaging over a long path. In addition, the nonmeteorological echoes are identified based on the characteristics of differential phase ψdp and copolar correlation coefficient ρhv measurements. The quality-controlled dual-polarization radar measurements then serve as input to the hydrometeor classification module. In DROPS2.0, the region-based hydrometeor classification proposed by Bechini and Chandrasekar (2015) is applied. A brief description of this methodology is provided here. For details, the reader is referred to Bechini and Chandrasekar (2015). The input radar data for hydrometeor classification include Zh, Zdr, Kdp, and ρhv. The vertical profile of temperature T observed from a nearby sounding station is also used as an optional input. The overall structure of this region-based classification methodology is depicted in step 2 in Fig. 2. First, traditional bin-based fuzzy-logic approach with four general blocks (i.e., fuzzification, inference, aggregation, and defuzzification) is implemented to get initial classification results (Liu and Chandrasekar 2000; Lim et al. 2005; Dolan and Rutledge 2009; Chandrasekar et al. 2013). The temperature profile is then adjusted based on the quality of wet ice classification, which is essentially the average confidence of all the bins identified as wet ice based on the inference rule (Bechini and Chandrasekar 2015). Second, a modified K-means clustering technique is applied to incorporate the spatial contiguity and microphysical constraints. Then, the connected component labeling algorithm is employed to derive connected regions (Gonzalez and Woods 2002), and the final classification is performed over connected regions where unique labeling of regions populated with adjacent bins are assigned to the same hydrometeor type. In total, 11 hydrometeor types are classified: large drops (LD), drizzle (DR), rain (RA), heavy rain (HR), rain–hail mixture (RH), hail (HA), graupel (GR), wet ice (WI), dry ice (DI), crystals (CR), and dendrites (DN). Ground clutter and nonmeteorological echoes are also classified as clutter (CL). Compared to conventional fuzzy-logic method, this region-based approach is appealing in terms of operational application and easy interpretation. Figure 3 illustrates sample NPOL radar observations and corresponding hydrometeor classification results for a range–height indicator (RHI) scan at 2343 UTC 29 May 2013.
For the sake of precipitation estimation, we keep a similar idea with DROPS1.0 in Cifelli et al. (2011) by narrowing down the hydrometeor classes to three categories: liquid, rain–hail mixture, and others, where liquid includes LD, DR, RA, and HR; rain–hail mixture is RH; and others includes HA, GR, WI, DI, CR, DN, and CL. Rainfall estimation is then conducted based on the hydrometeor categories and thresholds on Zh, Zdr, and Kdp. At S-band frequency, the thresholds on Zh, Zdr, and Kdp are 38 dBZ, 0.5 dB, and 0.3° km−1, respectively (Bringi et al. 1996; Petersen et al. 1999; Cifelli et al. 2011). However, the thresholds may need to be adjusted based on the measurement quality that can vary from system to system, due to a number of factors such as radar signal processing algorithms.
To attain the specific rainfall relations particularly suitable in the IFloodS domain, we use the DSD measurements from the 14 APU disdrometers (see Fig. 1) for simulation purposes. The Parsivel DSD data are essentially the number of raindrops in a 32 × 32 size versus fall velocity matrix (Tokay et al. 2014). In total, 13 772 one-minute-averaged DSDs are used for DROPS2.0 rainfall algorithm development. This DSD dataset (training data) consists of nine precipitation days, including a couple of severe multicellular convective thunderstorms and a few widespread stratiform rain cases. The dual-polarization radar moments (i.e., Zh, Zdr, and Kdp) are simulated at S-band frequency, using the T-matrix method (Waterman 1965). The drop shape model used in the simulation is the one proposed by Brandes et al. (2002). The temperature information is obtained from a local sounding station. We also compute the rainfall rates directly from the DSD data using the following equation:
where R is rainfall rate (mm h−1), Dn is raindrop mean diameter (mm), Sn is diameter spread (mm), is the number of drops, and is rain drop terminal velocity (m s−1) at level n.
The diameter level and spread are specified for a given type of disdrometer (Tokay et al. 2014). Equation (1) is essentially the discrete form of the definition of rainfall rate given by Bringi and Chandrasekar (2001):
where D is raindrop diameter, is the number of raindrops, and is the drop terminal velocity at sea level. In this study, the fall velocity measurements from disdrometers are not used because of the inaccuracy, particularly at larger size and faster fall speeds [for details, see Tokay et al. (2014)]. Instead, the model-based relation from Atlas et al. (1973) is adopted when calculating rainfall rate from DSD data.
Then, nonlinear regression is applied between rainfall rates and dual-polarization measurements in order to get the specific rainfall relations as follows:
where Zh (mm6 m−3) is the reflectivity factor at horizontal polarization and is differential reflectivity in linear scale.
In addition, we compare the results of DROPS2.0 with the standard NEXRAD Z–R relation [Eq. (4a)], DROPS1.0, and the relation used by WSR-88D (Giangrande and Ryzhkov 2008; Berkowitz et al. 2013). The sequence of equations used in DROPS1.0, which was developed based on simulated DSD data (Bringi and Chandrasekar 2001), is repeated as follows (Cifelli et al. 2011):
It should be noted that we are not fully implementing the blended WSR-88D rainfall methodology [for details, see Giangrande and Ryzhkov (2008)]. Instead, only the rainfall relation used in liquid regions is referred since the current operational version of WSR-88D QPE system only estimates the amount of liquid precipitation (Berkowitz et al. 2013), in which case Eq. (5) is adopted:
Hereafter, Eq. (5) will be referred to as NEXRAD DP relation.
In the implementation of DROPS1.0, NEXRAD Z–R, and DP relations, the data quality control in step 1 in DROPS2.0 is conducted first. In addition, NEXRAD Z–R and DP relations are only applied when the precipitation type is classified as “liquid” (based on step 2 in DROPS2.0), and zeros are assigned for the regions where nonliquid precipitation types are identified. To investigate the parameterization error structure of various rainfall algorithms, another DSD dataset (testing data) is used to quantify the parameterization errors, particularly for liquid precipitation estimation. Although collected in the same field experiment, the testing dataset is independent from the training data used to derive Eq. (3). The normalized standard deviation σp of rainfall-rate estimates (for liquid regions), defined as follows, is computed at different rainfall intensity ranges:
where and are, respectively, the estimated rainfall rates with radar rainfall relations in Eqs. (3)–(5) and rainfall rates computed directly from APU DSD (testing) data using Eq. (1). The variable stands for standard deviation. The angle brackets stand for sampling average. Figure 4a shows the scattergram of rainfall rates estimated using DROPS2.0 rainfall relations in Eq. (3) versus rainfall rates directly computed from DSD data using Eq. (1), whereas Fig. 4b illustrates σp due to parameterization of various rainfall algorithms. From a theoretical perspective, for liquid precipitation estimation, we can strive to achieve the error rates in Fig. 4b provided that the measurement errors can be eliminated by spatial or temporal averaging. It is also seen that although the rainfall relations in DROPS2.0 have better performance than those in DROPS1.0 at light rainfall region (less than 10 mm h−1), the parameterization error structures of DROPS1.0 and DROPS2.0 are quite similar at moderate and heavy rainfall regions, and both of them show better performance than the NEXRAD Z–R or NEXRAD DP relation. In section 4, the performance of various algorithms will be demonstrated and evaluated with radar data collected in an experimental environment.
3. Rainfall events and dataset description
To demonstrate the performance of the aforementioned rainfall algorithms, S-band NPOL radar data collected during the IFloodS field experiment are used (Wolff et al. 2015). During the IFloodS experiment, the NPOL radar was deployed at an ideal location to fill the gap of WSR-88D low-elevation coverage, and for proximity to local river basins. NPOL was operated in several modes, including the two-sweep (i.e., 0.7° and 1.4°) full plan position indicator (PPI), RHIs over the APU disdrometers, PPI sector (PPS) scans of precipitation systems over principal river basins, and “bird bath” scans that can be used for monitoring Zdr biases. The RHI sector scans covered an azimuth range of 8° above locations of ground-based instrumentation (e.g., APU disdrometers). The PPI and RHI scan tasks were repeated once every 3 min when precipitation was detected anywhere within NPOL’s coverage range, and they were performed throughout the campaign. Other scans, such as PPS and bird-bath scans, were scheduled between rain scans on an event-by-event basis, among which three options for PPS scans were considered depending on echo-top height and range to NPOL radar in order to obtain high-resolution rainfall mapping over the local river basins. Details of the scanning strategies can be found in W. Petersen and W. F. Krajewski (2016, unpublished manuscript). In this paper, we select three precipitation events that are characterized by different meteorological features. In this section, the three rainfall events are briefly described, and APU data processing will be detailed for quantitative validation purpose.
a. Rainfall events
1) 20 May case
This event, characterized by a mesoscale convective system (MCS), began in the form of strong, tornadic storms near the KDMX radar (NEXRAD deployed in Des Moines, Iowa). From the evening hours of 19 May to the very early morning hours of 20 May local time (central daylight time, i.e., UTC − 5 h), a few isolated cells developed in IFloodS domain. In particular, a strong line of convection was observed to the west of the NPOL radar, moving to the east, shortly after 0000 UTC 20 May. Followed by an asymmetric MCS from the southwest, this convective line passed over the NPOL site around 0145 UTC. For validation purposes, we only make use of the NPOL data collected during 0200–0500 UTC 20 May, when the rainfall was significantly impacting the disdrometer network. During this period, NPOL was conducting regular RHI and full surveillance PPI scans. In addition, a few PPS scans were conducted over the Turkey River basin (northeast of the NPOL radar) and the disdrometer network near the 130° radial (southeast of the NPOL radar) when strong convection moved to the regions of interest. In this research, the lowest (i.e., 0.7° elevation) PPI as well as PPS scans over the disdrometer network were used to generate various rainfall products. It is worth mentioning that, because of the high winds (~31 m s−1) at the radar site, it was decided to stow NPOL radar antenna in the vertical position at 0126 UTC, and it was restored at 0145 UTC. This may slightly affect radar data processing and subsequent rainfall estimation, which will be detailed in section 3b.
2) 25 May case
This is a typical stratiform precipitation case. Rain showers were observed in the IFloodS domain all through the night of 24 May. The stratiform precipitation became more widespread in the morning of 25 May, especially to the south and east of the NPOL site, and it lasted until late afternoon. Several flood and flash flood watches and warnings were issued in the IFloodS and nearby regions. NPOL radar was fully staffed again after maintenance on the previous day. It had been continuously conducting PPI and RHI scans during this event. PPS scans near the 130° azimuth were also scheduled around 1200–1800 UTC in coordination with instrumented disdrometer array as precipitation was focused there. It is a good case for horizontal variability studies of precipitation properties. Similar to the previous event, the lowest PPI and PPS sweeps (i.e., 0.7° elevation) collected during 1200–2100 UTC 25 May are used to derive rainfall products when fairly uniform precipitation coverage was observed over the disdrometer network in the NPOL domain.
3) 29 May case
With a major MCS passing through the entire IFloodS domain, this is another well-documented case with severe weather and heavy rain. Besides the regular PPI and RHI scans, NPOL performed many hours (around 1715–2100 UTC) of dedicated PPS scans for high-temporal-resolution rain mapping. It is noted that, around 2100–2400 UTC, the strong convective cells moved to the southeast of the NPOL coverage domain, where the disdrometer arrays were deployed. Therefore, the lowest PPI scan data collected during this period are utilized for rainfall analysis. In addition, single RHI scans over the disdrometer radial were also conducted regularly in order to investigate the vertical structure of precipitation. Nevertheless, characterization of the vertical structure/distribution of rainfall is beyond the scope of this paper.
b. Radar and APU data processing
During IFloodS, a number of disdrometers and tipping-bucket rain gauges were deployed within NPOL radar coverage to provide in situ validation data (see Fig. 1). In a number of previous studies, it has been shown that the use of rain gauges can introduce significant errors to high-temporal-resolution QPE systems due to the limitations on sampling time and bucket volume resolution, particularly in light rainfall cases (e.g., Habib et al. 2001; Chen and Chandrasekar 2015b). Therefore, we intend to use only APU data for radar rainfall evaluation. During the three precipitation events, some of the APUs had malfunctions. Only the APUs that were working fine all through the three precipitation events are used in this paper, namely, APU02, APU03, APU05, APU06, APU08, APU09, APU11, APU13, and APU14. Each APU, equipped with a Parsivel unit (version 2) developed by OTT Hydromet in Germany, is an optical disdrometer that can measure raindrop size and falling speed (Tokay et al. 2014). During the field experiment, the APU sampling resolution was configured to 1 min. With the drop size distribution, rainfall rate can be computed using Eq. (1). For the sake of evaluation, the consecutive 1-min APU rainfall-rate data are aggregated to get 5-, 15-, 30-, 45-, and 60-min rainfall accumulations.
The data quality control and Kdp estimation procedure in section 2 is applied to NPOL radar data before implementing various rainfall relations. The estimated NPOL radar rainfall rates corresponding to APU rainfall observation time are used to produce matched radar rainfall amounts. For the time frames when APU (or radar) did not report rain, zeros are assigned. The radar–APU rainfall pairs are then used for quantitative evaluation. For the time frames when there was no NPOL radar data/scan (not often), a piecewise cubic Hermite interpolating polynomial (PCHIP)-based interpolation methodology (Fritsch and Carlson 1980) is applied in order to get radar rainfall-rate data that match APU rainfall measurement.
4. Evaluation, results, and discussion
a. Sample products and evaluation metrics
In Fig. 5, sample NPOL radar observations and corresponding rainfall-rate estimates using different algorithms are provided for the event of 20 May 2013. Similarly, Figs. 6 and 7 show sample observations and corresponding rainfall products for the event of 25 and 29 May 2013, respectively.
The radar observations represent a unit of illuminated volume in polar coordinate with resolution of 0.98° × 150 m, whereas APUs are providing pointwise measurements. In this paper, the radar range gate closest to APU location is selected for the purpose of pointwise quantitative evaluation. The discrepancies caused by wind drift on radar–APU comparisons are neglected. Assuming the APU measurements are the “ground truth,” a set of metrics are computed for rainfall estimates at different time scales at each APU location. The evaluation metrics include the normalized mean absolute error (NMAE), Pearson correlation coefficient (CORR), and root-mean-square error (RMSE), defined as follows:
where the angle brackets stand for sample average and and denote the estimated rainfall amount at different time scales (i.e., 5, 15, 30, 45, or 60 min) from NPOL radar and APU, respectively.
b. Performance of different rainfall algorithms
Scrutinizing Tables 2–5, it can be concluded that DROPS2.0 generally has the best performance in terms of NMAE and RMSE. Surprisingly, NEXRAD Z–R has lower NMAE compared to NEXRAD DP relation for most of the cases, although NEXRAD DP generally has higher CORR. DROPS1.0 provides better QPE than NEXRAD Z–R or NEXRAD DP relation, except at the locations of APU13 and APU14, but its performance is not as good as DROPS2.0. As expected, DROPS1.0 has the worst performance among various rainfall algorithms at APU13 and APU14, especially for the stratiform precipitation event (25 May 2013). This is because of the range limitation (within or above melting layer) of DROPS1.0 (Chen and Chandrasekar 2015a; Seo et al. 2015). A more detailed comparison between DROPS1.0 and DROPS2.0 will be provided in section 4d.
It is also found that there is no big statistical difference among the four rainfall algorithms in terms of CORR. That is, all of the four algorithms can provide QPE with high CORR with respect to APU rainfall observations. However, it should be noted that the CORR has an increasing trend as the rainfall accumulation time increases from 5 to 60 min. On the other hand, the NMAE has a decreasing trend. This can be attributed to the reduction of random error in radar measurements due to temporal and spatial averaging. The RMSE increases as the rainfall accumulation time increases (i.e., rainfall amount gets larger). To further demonstrate the rainfall performance, we also show scatterplots of the radar–APU rainfall comparisons at a sample APU location (i.e., APU03) for the three cases combined in Fig. 8. Corresponding evaluation results are shown in Fig. 9. Scatterplots of the individual events are not shown because they show essentially the same results as those in Figs. 8 and 9.
c. Range impact on QPE performance
From a theoretical point of view, the advanced dual-polarization techniques have made it generally possible to obtain rainfall algorithms less sensitive to DSD. However, from an operational point of view, the geometry of radar measurements combined with the variability of the spatial structure of precipitation still limits the radar rainfall accuracy, especially for the regions far from radar (Ryzhkov 2007; Gorgucci and Baldini 2015). In this section, we take advantage of the unique instrument layout during the NASA IFloodS field campaign to quantify the rainfall error introduced by range impact. As shown in Fig. 1, APU05, APU06, and APU08 are almost collocated; APU09 and APU11 are very closely deployed; and APU13 and APU14 are collocated. Therefore, for the sake of comparison, we take the mean of ranges and rainfall evaluation results for those APUs that are closely deployed. Figure 10 shows a conceptual diagram illustrating the radar beam broadening effect. The mean ranges, radar center beam heights, and radar range gate volumes at these ranges are also indicated in Fig. 10.
Figure 11 shows the NMAE of different rainfall products as a function of range for all three events combined, whereas Fig. 12 shows the CORR results as a function of range. It can be seen from Fig. 11 that although the NMAE has no monotonic upward trend, it does have a general increasing trend for the distances after about 25 km. The decreasing trend of CORR is very clear in Fig. 12, which means the radar-estimated rainfall and APU-observed rainfall are less correlated as the range goes farther. As expected, outliers are seen in Figs. 11 and 12 (i.e., the performance of DROPS1.0 at the longest range), which again demonstrates the range limitation in the operational application of DROPS1.0. In general, it is suggested that DROPS1.0 should be used only within 100 km from radar. At long ranges (beyond 100 km), a single Z–R-based method shows similar performance (even better for longer rainfall integration time) compared to dual-polarization methodologies, which contradicts the theoretical analysis in Fig. 4b. This poses a challenge for weather radar to capture the complex spatial and temporal variabilities of precipitation at long distances.
We also want to note that the RMSE results are not shown to investigate their range dependencies because they are more dependent on the rainfall intensities compared to ranges.
d. DROPS1.0 versus DROPS2.0
A more detailed comparison between DROPS1.0 and DROPS2.0 is provided as follows. From the theoretical point of view, Fig. 4b shows that DROPS2.0 has very similar performance for liquid precipitation estimation compared with DROPS1.0, although it shows better performance for light rainfall regions (less than 10 mm h−1). However, when solid precipitation is involved, it is a challenging task to determine which rainfall relation to apply for a given set of dual-polarization measurements.
Figure 13 shows sample observations from NPOL radar at 1253 UTC 25 May 2013, as well as corresponding hydrometeor classification and rainfall-rate estimation results using DROPS1.0 and DROPS2.0. This case is selected because it includes partial beam blockage (i.e., a few beams near the 208 azimuth degree radial are partially blocked), and the stratiform regions shown in Fig. 13 are close to melting layer. In DROPS1.0, traditional fuzzy-logic-based hydrometeor classification methodology is implemented. However, this bin-by-bin-based classification method has great limitations when applied for “noisy” radar data caused by ground clutter, partial beam blockage, and/or brightband contamination. In this particular case, the poor performance of DROPS1.0 is mainly due to the melting layer contamination. At far distances, especially for the stratiform rain events, the radar beam partially overshoots liquid precipitation. In the regions close to or within the melting layer, the bin-by-bin-based classification approach is not able to properly identify the mixed-phase precipitation (i.e., hydrometeor classification results will be noisy). As a result, the estimated rainfall-rate field is noisy, which leads to low-quality rainfall accumulation products (see Table 3). The poor performance of DROPS1.0 at far ranges can also be explained by the outliers in Figs. 11 and 12. Although Figs. 11 and 12 are based on the rainfall estimates for all three events combined, the results beyond the 70-km point are basically from two events (i.e., 20 and 25 May 2013) since APU13 and APU14 did not observe any rain during the event of 29 May 2013. The noisy rainfall-rate field that resulted from brightband contamination in individual precipitation events significantly affects the overall performance of DROPS1.0. Previous studies suggested that either a partial (coefficients adjusted) Z–R relation could be adopted or the vertical profile of reflectivity (VPR) could be corrected to avoid overestimation introduced by brightband contamination (e.g., Bellon et al. 2005; Matrosov et al. 2007). In this paper, to avoid the complex mixed-phase precipitation as well as the range degradation, DROPS1.0 does not conduct rainfall estimation for the regions where the radar beam is above the melting layer. In DROPS2.0, the region-based approach takes into account the hydrometeor classification quality and the correlation with adjacent range bins. Benefiting from the hydrometeor microphysical constraints applied in the new classification scheme [for details, see Bechini and Chandrasekar (2015)], DROPS2.0 is able to provide a clean classification that allows adopting appropriate rainfall relations in case of limited melting layer contamination. Rainfall estimates can be significantly improved through the high-quality classification of different hydrometeor types. The improvement can be seen in Figs. 13g and 13h.
In addition, as shown in Fig. 13e, the hydrometeor type is misclassified by DROPS1.0 in the areas where the radar beams were partially blocked (low ρhv region in Fig. 13d near the 208 azimuth degree radial). As a result, the rainfall rate is not estimated by DROPS1.0 in such regions (see Fig. 13f), which is difficult for operational interpretation. This is recovered in DROPS2.0 (see Fig. 13h) by exploiting information from surrounding areas.
Radar has been used for rainfall estimation since its earliest application in meteorology. Traditionally, rainfall estimation with radar has been accomplished by relating the backscattered power to rainfall rate through the so-called Z–R relations. However, it has been found that Z–R relations greatly depend on drop size distribution (DSD), which varies across different rainfall regimes, even within a single storm. It is a challenging task to find an ideal Z–R relation for a given region to represent the local rain microphysical properties of different types of storms in different seasons. Since the introduction of polarization diversity in meteorological radar (Seliga and Bringi 1976), a large amount of research effort has been expended on polarimetric radar system and its weather applications. Particularly, through the dual-polarization radar measurements, including reflectivity at horizontal polarization Zh, differential reflectivity Zdr, specific differential propagation phase Kdp, and copolar correlation coefficient ρhv, radar data quality can be considerably enhanced. In addition, different hydrometeor types can be identified, and DSD information can be retrieved in more effective ways with the dual-polarization observations (Chandrasekar et al. 2013). A number of dual-polarization radar rainfall algorithms have been developed in previous studies. Although there is still no standard algorithm for radar rainfall estimation, the hydrometeor-classification-based rainfall methodologies are widely used in the weather radar community. Such rainfall systems typically consist of three modules, namely, data quality control, classification of different hydrometeor types, and precipitation quantification with appropriate rainfall relations. However, the rainfall products derived based on traditional bin-by-bin-based fuzzy-logic classification methods are not sufficient for operational applications, especially if the radar data are noisy. That is because the hydrometeor classification result will be noisy and unstable when the input radar data quality is low since the classification quality and correlation with adjacent range gates are not taken into account. In addition, traditional hydrometeor-classification-based rainfall methods severely suffer from brightband contamination due to the challenges of mixed-phase precipitation classification in melting layer.
In this paper, we proposed an improved S-band radar rainfall algorithm termed DROPS2.0, which, essentially, is also a hydrometeor-classification-guided approach. The advanced classification technique implemented in DROPS2.0 exploits the spatial information content of dual-polarization radar observations. Compared to traditional fuzzy-logic classification, it also considers the spatial coherence, the quality of classification itself, and the self-aggregation propensity of radar measurements (Bechini and Chandrasekar 2015). This hydrometeor identification methodology, in case of limited melting layer contamination, can also provide a classification that allows adopting appropriate rainfall relations. As a contribution to GPM ground validation activities, the proposed dual-polarization radar rainfall methodology is available operationally. It has been implemented, demonstrated, and evaluated with NPOL radar data collected during the IFloodS field campaign. It is shown that DROPS2.0 generally has excellent performance compared to the traditional fuzzy-logic-based dual-polarization approach and single-polarization-based Z–R relation. DROPS2.0 is very robust, and it was continuously working fine during the IFloodS campaign.
In addition, the impact of radar beam broadening on various rainfall algorithms has also been investigated within the framework of the GPM IFloodS field experiment. It is found that the radar-estimated rainfall is less correlated with ground truths (disdrometer measurements) when the beam goes farther from radar. It can be a challenging task to use weather radar to characterize the complex spatial and temporal variabilities of precipitation at long distances (e.g., beyond 100 km). Since S-band is one of the standard radar operating frequencies in many countries over the world, the methodology proposed in this paper can potentially be applied in a broader domain with operational dual-polarization radars.
This research was primarily supported by the National Science Foundation Hazards Science, Engineering and Education for Sustainability (SEES) program under Grant AGS-1331572. Additionally, we acknowledge that the APU and NPOL data were kindly provided by NASA. The authors thank Drs. Matthew R. Schwaller and Walter A. Petersen (all from NASA) for their field support of the instrument deployment. The authors are grateful to David B. Wolff, David A. Marks, and Jason L. Pippitt (all from NASA) for their feedback on the use of DROPS2.0.
This article is included in the IFloodS 2013: A Field Campaign to Support the NASA-JAXA Global Precipitation Measurement Mission Special Collection.