A new melting layer detection algorithm that combines polarimetric radar-based detection with thermodynamic output from numerical models

Transcription

1 ERAD THE EIGHTH EUROPEAN CONFERENCE ON RADAR IN METEOROLOGY A new melting layer detection algoritm tat combines polarimetric radar-based detection wit termodynamic output from numerical models Terry J. Scuur 1,, Alexander V. Ryzkov 1,, and Jon Krause 1, 1 Cooperative Institute for Mesoscale Meteorological Studies, University of Oklaoma, Norman, Oklaoma, USA NOAA/OAR/National Severe Storms Laboratory, Norman, Oklaoma, USA 1 Introduction Te classification of cold-season precipitation at te surface is complicated by te broad range of ydrometeor types tat migt result from termodynamic and micropysical processes tat occur below te eigt of te radar s lowest elevation sweep. Because of tis, te fuzzy-logic-based ydrometeor classification algoritm tat was deployed on te WSR-88D network (Park et al., 009), wic gives classifications on conical surfaces, often provides results in transitional winter weater events tat are not representative of te precipitation type observed at ground level. In response to tis deficiency, Scuur et al. (01) reported on initial work to create an algoritm tat combines termodynamic output from a numerical model wit polarimetric radar data to produce a surface-based classification. Te most recent version of tis algoritm, wic as evolved into an all-season algoritm tat also takes advantage (under certain warm- and cold-season conditions) of results from te fuzzy-logic-based sceme, is summarized in a separate paper at tis conference by Scuur et al. (014). In sort, te transitional winter weater component of te algoritm relies upon te use of termodynamic output from te Hig-Resolution Rapid-Refres (HRRR) model to generate a background classification tat is ten eiter accepted or rejected based on observational evidence (determined troug an examination of te polarimetric radar data) of te presence/absence of a melting layer (ML). As suc, te algoritm s performance relies eavily upon te accurate detection of weter or not a ML exists aloft. Tis in itself can present callenge since several different metods for detecting te ML exist, wit all of tem aving teir own advantages and disadvantages. Te ML detection algoritm tat is currently deployed on te WSR-88D network (referred to as te WSR-88D MLDA, Giangrande et al. 008) uses tresolds of radar reflectivity (Z), differential reflectivity (Z DR ), and correlation coefficient ( HV ) collected at ig elevation scans (4-10º elevation) to determine te ML top and bottom at locations close to te radar. However, te algoritm ten assumes tat te near-radar MLDA detections can be projected out to more distant ranges along eac azimut. Tis feature as proven to be unrealistic for most meteorological situations, frequently resulting in te extension of a ML into regions were one does not exist. More recently, Krause et al. (013) presented a tecnique tat uses polarimetric radar data at lower elevation angles (<4º elevation). Wile te Krause et al. metod, wic was developed primarily to determine were polarimetric rainfall estimation migt suffer from ML contamination, as te advantage of being able to provide ML designations at more distant ranges tan is possible wit te iger elevations used by te MLDA, te interpretation of te results is often complicated by te effects of beam broadening. Combined, tese complicating factors wit te radar-based algoritms point to a need for a ML detection metod tat 1) takes account of te advantages and disadvantages of existing algoritms, ) can be applied over te entire radar domain, and 3) capitalizes on additional information provided by termodynamic output from numerical models. In tis paper, we describe a new ML detection algoritm tat combines radar observations wit termodynamic output from te HRRR model to provide a ybrid ML designation over te entire radar domain. Radar-based, range-dependent Gaussian weigting functions for bot ig- and low-elevation ML designations, wic take into account inerent errors known to eac tecnique, are combined wit a model-based Gaussian weigting function tat depends on orizontal gradients in wet-bulb temperatures to create a blended map of ML detections. A separate, time-dependent weigting function is used to account for time lag in te model analyses by de-empasizing te contribution from te numerical model as te t from te radar volume and te most recent model analysis becomes large. Metodology Te blended ML is created troug te combination of output from tree distinct melting layer detection tecniques: 1) te ML detection algoritm of Giangrande et al. (008), wic uses polarimetric radar data from 4-10º elevation scans, ) te ML contamination product of Krause et al. (013), wic uses polarimetric radar data from scans < 4º elevation, and 3) te region of ML, ere defined as a wet bulb temperature (T w ) of 0ºC < T w < 4ºC, tat is output from te HRRR model. In te following discussion, we use te subscript to refer to te Giangrande et al. (008) ML product since it was derived from iger elevation scans, te subscript l to refer to te Krause et al. (013) ML product since it was derived from lower elevation scans, and te subscript m to refer to te ML as determined from te numerical model output. Te first step in te combination process is to determine wic ML detection metods indicate te presence of a ML at eac individual gate. For any individual radar ray at a given azimut and elevation (< 4 ), we assign tree indices (I m, I, and ERAD 014 Abstract ID NOW.P05 1 scuur@ou.edu

2 ERAD THE EIGHTH EUROPEAN CONFERENCE ON RADAR IN METEOROLOGY I l ) to a eac range gate along te ray. I m = 1 if te HRRR T w at te gate is between 0 and 4 C, oterwise I m = 0. I l = 1 if te gate falls into te ML contamination areaa as determined by Krause et al. (013), oterwise, Il = 0. Similarly, I = 1 if te gate falls into te ML contamination area as determinedd by Giangrande et al. (008), oterwise, I = 0. Since te existing ML detection algoritm (Giangrande et al. 008) produces two numbers for a given azimutal direction: H b and H t corresponding to te eigts of te bottom and top of te melting layer. Te geometrical projections of H b and H t on te radar ray delineate te segment of te ray were I = 1 and I = 0 outside of te segment. Te gates wit I, I l, and I m equal to 1 are generally different as illustrated in Fig. 1. Figure 1: Melting layer designation at lower antenna elevations. Te segments tat indicate melting layer for eac product are ten combined into a single melting layer product by computing an aggregate weigt between 0 and 1, were: WtW mi m Wl Il WI A W W W W t m W m, W l, and W are Guassian weigts associated wit te tree products, weree te weigts are determinedd as follows l (1) Wm 0.5exp 0.69 g g 0 () W l r fl ( r)exp r0 l (3) W f ( H b r ( )exp 0.69 r0 In () (4), r is a slant range and te variable g caracterizes orizontal gradients of T w at te eigt H 0 = 1.5 km. (4) Finally, we introduce, g dt ( 0 H ) dt dx ( w H ) 0 dy w W t t exp t 0 (5) (6) ERAD 014 Abstract ID NOW.P05

3 ERAD THE EIGHTH EUROPEAN CONFERENCE ON RADAR IN METEOROLOGY were W t is a time-dependent weigting function tat is used to account for time lag in te model analyses by de- becomes large. empasizing te contributionn from te numerical model as te t from te radar volume and te most recent model analysis Te parameters g 0, r 0l, and r 0 are adaptable and willl continue to be optimized in te process of testing te algoritm. In tis paper, g 0, r 0l, r 0 and t 0, are 0.03K/km, 100 km, 50 km, and 60 min respectively. Te factor fl is equal to 1 for r > r gc and it linearly decreases from 1 to 0 for r < r gc were r gc is te distance up to wic te impact of ground clutter (gc) contamination on te quality of te product l can be significant. For most radars, setting r gc to 30 km seems to be reasonable. Te purpose of introducing factor f l is to reflect te fact tat te quality of te l product migt be compromised in te presence of ground clutter. Similarly, te factor f is introduced to account for degradation of te product at very low melting layer (or low H b ). Hence, f = 1 if H b > gc and f linearly decreases to 0 for H b < gc. For tis study, gc is 1.0 km. Te expression (1) for te aggregation value A is constructed in suc a way tat A varies between 0 and 1. If a given range gate is classified as ML by all tree ML products ten A = 1. If te aggregation value is above certain tresold A 0 (we may start from A 0 = 0.5) ten te gate is classified as true ML. Note tat if te wet bulbb temperature field is uniform and g = 0 ten W m = 0.5 over te wole radar coveragee area regardless of te distance from te radar. Te range-dependent weigts W l and W are iger tan W m at closer distances and lower tan W m far away from te radar. Fig. illustrates ow te decision about te presence of te ML aloft migt be made based on te results of true ML designation at antenna elevations less tan 4. According to te suggested approac, tere is no need to examine a vertical column in order to identify ML aloft. Instead, all ML points identified at lower conical scans are mapped to te surface. A pure model designation is performed at longer distances were te lowest tilt oversoots te melting layer. At close proximity of te radar beforee te closest ML point identified via te combination of m, l, and, te coice sould be made based on te m and designations using te aggregation value WtWm I A W W t m m W I W. (7) Figure : Scematic sowing contributions, wit range, of eac metod of ML detection to te overall ybrid product. ERAD 014 Abstract ID NOW.P05 3

4 ERAD THE EIGHTH EUROPEAN CONFERENCE ON RADAR IN METEOROLOGY 3 Example Norteastern U.S. Blizzard of 8 February 013 On 8 February 013, te Norteastern U.S. experienced a istoric winter storm tat resulted in snow accumulations up to 100 cm in dept. Figure 3, at 16 UTC on 8 February 013, sows Z, Z DR, and HV from te KOKX (New York City National Weater Service Office, located on Long Island at Upton, NY) WSR-88D radar. From an examination of te Z DR and HV fields, it is clear tatt a widespreadd region layer of warm air was located to te sout of Long Island at tis time, wit te transition from warm- to cold-season precipitation taking place along a region tat spanned te entire lengt of te island. Precipitation type reports during tis event supported tis transition wit widespread reports of rain on te sout side of te island and snow (and occasional reports of sleet) to te nort. We use tis date/time to illustrate te classification procedures followed by te proposed ybrid ML detection algoritm. Figure 3: (a) Radar reflectivity (dbz), (b) differential reflectivity (db), and (c) correlation coefficient from te KOKX (Upton, NY) WSR-88D radarr at 16 UTC on 8 February 013. Figure 4 sows te weigts used by te ybrid ML algoritm at tis time, were W m is derived from te HRRR dataa at H 0 =1.5 km using equation (), W l is derived from equation (3), and W is derived from equation (4). As can be seen from an inspection of Figure 4, W drops quickly at relatively sort distances from te radar. Wile ML detection at ig elevation angles as been sown troug comparisons to sounding data to be igly accurate (Giangrande et al. 008), te assumed projection of tose ML designations to more distant ranges as been found, for many events, to be unrealistic. A sarp reduction of W wit range terefore works to effectively enancee te contribution from te metod at close ranges, but quickly reducee its contribution to te overall product at more distant ranges. In a similar manner, te weigts for W l effectively eliminate it from being a significant contributor at close ranges (were ground clutter may result in false detections), enance its contribution at intermediate ranges, and ten quickly reduce its contribution at more distant ranges were substantial beam broadening can affect te quality of te data. As noted earlier, te weigts for all 3 metods (model, ig elevation radar-based detection, and low elevation radar-based detection) are controlled troug adaptable parameters. At tis time, tese parameters ave yet to be extensively tested. Figure 4: Weigts derived from (a) equation () for te model-based ML detection, (b) equation (4) for te ig-elevation radar-based detection, and (c) equation (3) for te low-elevation radar-based detection. Te weigt profiles are overlaid on a map of te NE U.S. centered on te KOKX WSR-88D radar (corresponding te radar fields displayed in Figure 3). ERAD 014 Abstract ID NOW.P05 4

5 ERAD THE EIGHTH EUROPEAN CONFERENCE ON RADAR IN METEOROLOGY Figure 5 sows te 0.5º elevation locations were Im=1, m corresponding to were HRRR T w on te 0.5º elevation surface falls between 0 and 4ºC, I = 1, corresponding to were te range were geometric projection of te I detection eigts fall on te 0.5º elevation surface, and I l =1, corresponding to were te 0.5º elevation scan detects ML contribution using te metodology described by Krause et al. (013). At tis time, te ig elevation radar-based ML detection of Giangrande et al. (008), wic often as to resort to sounding data in an insufficient number of ML detections were found at elevation anles between 4 and 10º, gave a ML BOT and ML TOP of ~.0 and.5 km, respectively. Te resulting geometric projection of tose eigts to te 0.5º elevation surface provides a ring of I =1, for tis case, at a range were te corresponding W (Figure 4b) is so low tat it would not be a significant contributor to overall ML product at te 0.5º elevation scan. Figure 5: Blue delineates te 0.5º elevation locations were (a) I m = 1, (b) I =1, and (c) I l =1. Te I m, I, and I l profiles are overlaid on a map of te NE U.S. centered on te KOKX WSR-88D radar (corresponding te radar fields displayed in Figure 3). Finally, we sow Figures 6 and 7, wic demonstrates te weigted product, A, given by equation (1) for te 0.5º elevation (Figure 6). Similar projections of A at oter radar elevations < 4º are ten combined into an aggregated surface projection of te maximum A (A surf ) at eac location. Work in te near future will focus on optimizing te adaptable parameters and finding te tresold weigt for A surf tat provides te most realistic solution. Figure 6: Te value of A at 0.5º elevation derived from equation (1). ERAD 014 Abstract ID NOW.P05 5

6 ERAD THE EIGHTH EUROPEAN CONFERENCE ON RADAR IN METEOROLOGY Summary and Future Work In tis paper, we ave described te basic framework for a new ybrid ML detection algoritm tat is designed to come up wit an optimal ML solution by combining bot radar- and model based tecniques troug te use of a metodology tat attempts to take into account te advantages and disadvantages of eac existing algoritm. Radar-based, rangedependent Gaussian weigting functions for bot ig- and low-elevation ML designations, wic take into account inerent errors known to eac tecnique, are combined wit a model-based Gaussian weigting function tat depends on orizontal gradients in wet-bulb temperatures to create a blended map of ML detections. A separate, time-dependent weigting function is used to account for time lag in te model analyses by de-empasizing te contribution from te numerical model as te t from te radar volume and te most recent model analysis becomes large. A simple example of ow te tecnique may be applied to a winter storm event is demonstrated on a radar volume collected by te polarimetric Upton, NY WSR- 88D radar on 8 February 013. Future work will focus on an extensive study of several events to determine 1) optimal values for te adaptable parameters, ) te A surf weigt tat gives te most realistic results. Acknowledgement Funding was provided by NOAA/Office of Oceanic and Atmosperic Researc under NOAA-University of Oklaoma Cooperative Agreement #NA11OAR43007, U.S. Department of Commerce, and by te U.S. National Weater Service, Federal Aviation Administration, and Department of Defense program for modernization of NEXRAD radars. References Giangrande, S. E., J. M. Krause, and A. V. Ryzkov, 008: Automated designation of te melting layer wit a polarimetric prototype of te WSR-88D radar. J. Appl. Meteor. Climatol., 47, Krause, J., V. Laksmanan, and A. Ryzkov, 013: Improving detection of te melting layer using dual-polarization radar, NWP model data, and object identification tecniques. 36 t Conference on Radar Meteorology, Breckenridge, CO, American Meteorological Society, Boston, 6. Park, H.-S., A. V. Ryzkov, D. S. Zrnic, and K-.E. Kim, 009: Te ydrometeor classification algoritm for te polarimetric WSR-88D: Description and application to an MCS. Wea. Forecasting. 4, Scuur, T. J., Park, H.- S., A. V. Ryzkov, and H. D. Reeves, 01: Classification of precipitation types during transitional winter weater using te RUC model and polarimetric radar retrievals. J. Appl. Meteor. Climate, 51, DOI: /JAMC-D Scuur, T. J., A. V. Ryzkov, H. D. Reeves, J. Krause, K. L. Elmore, and K. L. Ortega, 014: Recent modifications to a new surface-based polarimetric Hydrometeor Classification Algoritm for te WSR-88D network, 8 t European Conference on Radar in Meteorology and Hydrology, Garmisc-Partenkircen, Germany, 7.5. Giangrande, S. E., J. M. Krause, and A. V. Ryzkov, 008: Automated designation of te melting layer wit a polarimetric prototype of te WSR-88D radar. J. Appl. Meteor. Climatol., 47, ERAD 014 Abstract ID NOW.P05 6