Improving The Quantitative Precipitation Estimate For .

Eastern Region Technical AttachmentNo. 2015-06December 2015Improving the Quantitative Precipitation Estimate forHydrometeors Classified as Dry Snow by Polarimetric RadarsKirk R. Apffel*, Aaron Reynolds, and David ZaffNOAA/National Weather ServiceBuffalo, New YorkABSTRACTBetween 2011 and 2013, National Weather Service (NWS) Weather Surveillance Radar 1988Doppler systems (WSR-88D) were upgraded with a dual polarization capability. Thepolarimetric upgrade is a significant enhancement that provides new and improved informationabout precipitation type, intensity, and size. Much work has gone into improving quantitativeprecipitation estimates (QPE), but the dual polarization QPE system only uses a modifiedversion of the legacy reflectivity - rain relation for returns classified as dry snow that applies amultiplication factor of 2.8 to take into account the lower reflectivity returns associated with drysnow. NWS Forecast Office Buffalo, NY was upgraded with dual polarization during April 2012and together with surrounding offices noticed an overestimation in the dual polarization QPEfor several cold season events when the radar beam was above the melting layer. This studyused gauge-to-radar comparisons for 722 hourly cases to test whether the coefficient wascausing the overestimation. The results showed that the default coefficient of 2.8 was too highand led to a high bias in QPE. The mean dual polarization QPE was nearly double the gaugemeasured precipitation. When the coefficient was lowered to 1.4, the mean dual polarizationQPE was still 19% higher than measured precipitation, but much improved over the initialvalues.*Corresponding author address: Kirk R. Apffel, NOAA/National Weather Service, 587 Aero Drive, Cheektowaga,NY 14225. E-mail: kirk.apffel@noaa.gov

1. IntroductionFrom 2011 to 2013 the NationalWeather Service (NWS) network ofWeather Surveillance Radars (WSR-88D)across the United States were upgraded withthe addition of a dual polarization (DP)capability. This enhancement made itpossible to gain new and improvedinformation about precipitation type,intensity, and size (National Oceanic andAtmospheric Administration 2013).Following the installation of DP radar,forecasters at Weather Forecast Office(WFO) Buffalo, NY (BUF) and Cleveland,OH (CLE) noticed a high bias inQuantitative Precipitation Estimates (QPE)for several cool season events (Jamison andLaPlante, personal communication 2012).Although, much work has gone intoimproving QPE, the initial implementationof the DP QPE system still utilizes a basicreflectivity to rainfall (liquid equivalent)relationship for returns classified as drysnow rather than employing any of thepolarimetric variables. An initialassessment of QPE found that thepolarimetric radar precipitation estimationsystem may overestimate QPE when thelowest radar slice samples above the meltinglayer (Cocks et al. 2012).The 2.8 coefficient is applied to echoesclassified as dry snow to account for lowerreflectivity returns usually associated withthese hydrometeors. As previouslymentioned, initial assessments have shownthis to be excessive regardless of the surfaceprecipitation type. The Buffalo areareceives a variety of snowfall types withsynoptic, lake effect, and hybrid events. Drysnow QPE is critical to estimating snowfallrates and for river forecasts whenprecipitation melts and reaches the groundas rain. However, until a more robustalgorithm can be developed for dry snow,QPE will depend on using a legacy PPSEcorrection factor.Working with the Radar OperationsCenter (ROC), WFO Buffalo undertook astudy to quantitatively evaluate the DP drysnow QPE. The goals of this study were todetermine if DP QPE could be improved byadjusting the 2.8 coefficient, and to developa methodology for determining a moreappropriate value.2. Data CollectionFirst, the WFO BUF gauge networkwas carefully assessed in order to findreliable precipitation data. Gauges had toprovide reliable hourly data for allprecipitation types and record precision to ahundredth of an inch. Selection of sitesutilized local knowledge of gauge type,exposure, and track record for availabilityand data quality. Only gauges between 10km and 100 km of the Buffalo WSR-88Dradar were used (see Fig. 1). This was doneto avoid close sites which fall in the cone ofsilence and distant sites prone toovershooting and sampling issues associatedwith broader beam widths at longer ranges.Close proximity to Canada, Lake Erie, andOntario further limited candidate locations.The DP QPE algorithm uses the preDP legacy precipitation processing systemequation (PPSE) for returns classified as drysnow (e.g., the convective Marshall-Palmerrelationship). When the DP HybridHydrometeor Classification (HHC)algorithm classifies an echo as dry snow, theDP QPE system multiplies the legacy PPSEby 2.8 to derive the QPE (Giangrande andRyzhokov 2008). The QPE for dry snow isdetermined using the following relationship(Ulbrich and Lee 1999):QPE 2.8R(Z),where R(Z) (0.017)Z0.7142

These criteria yielded a total of 13 gauges tobe used for this study.conditions can cause unshieldedprecipitation gauges to significantlyunderestimate precipitation (Theriault et al.2012). In order to mitigate this impact,events with winds in excess of 4 m/s wereeliminated at the 9 gauges without windshields.Events were selected from coldseason months between October and Apriland the data was collected for two coldseasons, between October 2012 and April2014. Potential events were considered if atleast 5 of 13 gauges received greater than0.10 inch of precipitation. From this subsetof events precipitation at each gaugelocation was checked to see if the radarclassified the return as “dry snow” using theDP HHC for one continuous hour. ArchivedLevel II radar data and other products suchas the HHC, DP QPE, and gaugeprecipitation data were collected from theMulti-Radar Multi-Sensor (MRMS) Systemwebsite at http://nmq.ou.edu/ (Zhang et al.2011 and 2015). Brief periods of missing,undetermined, or anomalous data were quitecommon. Therefore, at least 90% of thehour had to be classified as dry snow toqualify as a preliminary hourly case.Other factors were also consideredbased on the data collected. Blowing snowcan cause a gauge to incorrectly reportprecipitation from snow blown into thegauge. Heated tipping bucket gauges cansometimes get clogged or not melt heavysnow fast enough to measure (Rasmussen etal. 2012). Data exhibiting these issues wereeliminated. As a final check, questionablegauge data was compared to measurementsfrom nearby cooperative observer orCoCoRaHS measurement. These criteriayielded 722 hourly cases which occurred on33 different days.Preliminary results from thisresearch provided strong evidence forlowering the coefficient for dry snow. On 6February 2014, the coefficient was changedto 1.4 for real-time DP QPE productgeneration. 102 of the 722 cases occurredafter the switch of the coefficient.Preliminary cases were furtherscreened for accuracy, keeping in mindgauge limitations in certain environments.A co-located or nearby site was used todetermine wind speed and groundprecipitation type. Windy surface3

Figure 1. A map showing the location of the 13 precipitation gauges used for this study andrange rings at a distance of 10 km and 100 km from the Buffalo WSR-88D.3. Resultsbased QPE still showed a positive bias (19%above gauge data), but this was a notableimprovement when compared to the casesusing the original default coefficient of 2.8(Fig. 3 and Table 1).DP QPE and gauge precipitation canbe used to calculate a coefficient for eachhourly case. This was done by multiplyingthe gauge measured precipitation by 2.8 andthen dividing the result by the DP QPE. Ofthe 722 cases, 620 cases with a coefficientof 2.8 resulted in a positive bias of 101%,that is, the radar based QPE was nearlydouble that of gauge data (Fig. 2 and Table1). For cases after 6 February 2014, the newcoefficient of 1.4 was used for thiscalculation. For these 102 cases, radarThe calculated coefficient varied foreach event depending on the type ofprecipitation sampled at surface gauges.Since the radar is sampling the precipitationabove the surface, many of the HHCidentified cases of dry snow resulted in otherprecipitation types at the surface as thehydrometeors fell through a melting layer4

below the radar beam. Table 2 shows thecalculated coefficient for all rain, all snow,and mixed precipitation events. Thecoefficient for all snow events was higherthan other events, with an averagecoefficient of 1.78. For events whichprecipitation was classified as snow at radarlevel, but melted before it reached theground the coefficient was 1.33 and thecoefficient for mixed precipitation cases was1.30. It is uncertain why the coefficient islower for rain and mixed precipitation cases,but one factor may be that these type ofevents have a melting layer in closeproximity. This could result in anoverestimation if bright band returns areincorrectly classified as dry snow. All rainand mixed precipitation cases were morecommon than all snow, with all snow eventsonly making up 13% of the cases.precipitation at lower levels, which iscommon in lake effect snow. However, it isimportant to note that despite the closeproximity to Lake Erie and Lake Ontario,the majority of cases used for this studywere synoptic scale events. Based onarchived forecast discussions, only about30% of the cases were impacted by lakeeffect, with the remaining cases synopticscale events.As an example, the ROC postprocessed radar data for an event on 28October 2012 is used to compare QPEcalculated using different coefficients.Figure 4 shows the DP storm total QPEusing the default 2.8 coefficient for thisevent. There is a noticeable ring of higherQPE about 120 km from the radar which isabout where melting layer is. Figure 5shows the same event using 1.5 as thecoefficient. Notice that there is lessdiscontinuity at and above the melting levelin the QPE.Sites closer to the radar had aslightly lower coefficient than stationsfurther from the radar (Table 3). In somecases the radar may be overshooting heavier5

Figure 2. A scatterplot graph which compares DP QPE to gauge measured precipitation for 620hourly cases with a 2.8 coefficient categorized by precipitation type at ground level.Table 1. A comparison of coefficients used for hydrometeors classified as dry snow for DP QPEand the calculated error for each one.Dual-polNumber ofDual-pol QPEGauge% ErrorCoefficientHourly Cases(inches)Measuredfor 41025.164.3519%6

Figure 3. A scatterplot graph which compares DP QPE to gauge measured precipitation for 102hourly cases with a 1.4 coefficient categorized by precipitation type at ground level.Table 2. The mean calculated coefficient for hydrometeors classified as dry snow separated bydifferent type of precipitation at ground level.Event TypeHourly CalculatedCases CoefficientAll Rain2321.33All Snow941.78Mixed3891.30Total7151.35Table 3. The mean calculated coefficient for hydrometeors classified as dry snow separated bydistance from the Buffalo WSR-88D radar.Event TypeHourly CalculatedCasesCoefficientClose to Radar ( 75 km)2841.23Far from Radar (75-100 km)4891.42Total7151.357

Figure 4. DP Storm Total Precipitation QPE from the Buffalo radar on October 28th, 2012between 1004Z and 2002Z. The default 2.8 multiplier is used for dry snow and ice crystals. Notethe melting layer discontinuity ring.Figure 5. DP Storm Total Precipitation QPE from the Buffalo radar on October 28th, 2012between 1004Z and 2002Z. This is the same event but using a multiplier of 1.5 for snow and icecrystals. Note the reduction of the melting layer discontinuity ring.8