P361

RECENT IMPROVEMENTS TO THE QUALITY

CONTROL OF RADAR DATA FOR THE OPERA

DATA CENTRE

R. Scovell∗1, N. Gaussiat2, and M. Mittermaier1

1Met O�ce, Exeter, UK.2Météo-France, Toulouse, France.

1. Introduction

OPERA is the operational programme forweather radar networking within EUMETNET,the grouping of European meteorological services.The current programme of activities, OPERA4,started in April 2013, after the successful comple-tion of the OPERA3 programme, and is describedfurther in a poster at this conference (P364) bySaltiko� et al. (2013).

In 2009, an OPERA3 project started to de-velop the OPERA Data Centre (ODC), which wasundertaken jointly between Météo-France (MF)and the UK Met O�ce (UKMO). It was to be-come the successor to the OPERA Pilot Data Hub(PDH), which ran from 2006 to 2011 and pro-vided radar rainfall composites at 4km resolution,with 15 minute updates and with a domain cov-ering the whole of Europe. In 2011, the ODCproject culminated in the delivery of an opera-tional service, called Odyssey, which is now hostedjointly at UKMO and MF and provides Europeanradar composites on the same domain as the PDHbut at 2km resolution with 15 minute updates.

∗Corresponding author address: Met O�ce,Fitzroy Road, Exeter. EX1 3PB. UK. e-mail:robert.scovell@meto�ce.gov.uk

Unlike the PDH, which received Cartesian radarproducts, it ingests volumetric radar data, includ-ing Plan Position Indicator (PPI) scans. It alsobene�ts from the use of a newer radar data ex-change format known as the OPERA Data Infor-mation Model (ODIM), developed by Michelsonet al. (2008), which de�nes strict rules for the en-coding of radar data and associated meta-data ineither of the HDF5 or WMO BUFR �le formats.Presently 18 National Met Services (NMSs) aresupplying ODIM-compliant radar data to Odysseyand this includes volumetric data from over 130radars, most of which operate at C or S-band fre-quencies.Three di�erent types of composite are produced

at each 15-minute time step as shown in Table 1.The names RATE, DBZH and ACRR correspond tothe ODIM names for the types quantity which areencoded.

Table 1: Composite product types.Name Description Units

RATE surface rainfall rate mm h−1

DBZH column maximum re�ectivity dBZACRR 1hr accumulation mm

For an example of the RATE product please see

1

(a) Without QC (b) With QC

Figure 1: Odyssey SURF composite with and without QC.

Figure 1. This is based on a weighted average ofdBZ over all radar pixels within a speci�ed hor-izontal range of each composite grid cell centre.A �xed Z − R relationship is used to convert theaveraged dBZ into mmh−1.

Most NMSs provide data with the least possi-ble pre-processing applied. This is recommendedby OPERA because it allows the Quality Con-trol (QC) to be done in a consistent and central-ized way on Odyssey. To date two QC methodshave been added to process each PPI scan beforecompositing. Firstly, a simple climatology-basedclutter �lter was introduced in 2011. Secondly, asuite of anomaly detection algorithms, providedby BALTRAD1, was introduced in March 2013.Figures 1a and 1b show an example of the SURFcomposite both with and without the QC applied.Further details of these algorithms are given in thenext few sections.

1BALTRAD is the Baltic Sea Region Programme (2007-2013) part-�nanced by the European Union, European Re-gional Development Fund and European Neighbourhoodand Partnership Instrument.

2. Odyssey clutter �lter

Figure 1a demonstrates that there is a signi�-cant amount of residual clutter in the compositenear to many of the radar sites. The clutter �l-ter's purpose is to mask out these echoes whileretaining the genuine rainfall signals as much aspossible.

A set of monthly echo count images are updatedwhen new PPI scans are processed by Odyssey.For each known PPI scan con�guration there isone echo count image stored which is indexed bya combination of the year, month, radar site codeand elevation angle (with a tolerance of 0.01◦ toallow for drift).

Figure 2 shows a monthly echo count image forthe UK Clee Hill radar scan at 1.0◦ elevation,which extends to 250km range. The image showswidespread regions of permanent ground clutter(where the echo count is at or near 100%) withinthe �rst 100km of the radar site.

By combining a clutter �lter with the composit-ing algorithm it becomes possible to reject pix-els entirely from a radar image and yet retain

2

Figure 2: Normalized monthly echo count accu-mulation for Clee Hill radar (1.0◦ elevation).

valid data in the composite by substituting nearbyradar pixels from the same radar or a di�erentradar. Radar pixels are rejected when the normal-ized echo count (a value between 0% and 100%)exceeds a predetermined clutter threshold whichis de�ned on a site-by-site basis. The normaliza-tion factor is equal to the highest recorded echocount in the image (divided by 100).

a. Evaluation

The �rst version of Odyssey used a clutterthreshold which was set conservatively at valueof 90% for all radars. Work was done later withinOPERA to evaluate the impact of the �lter andthe e�ect of the choice of the threshold.

This evaluation was carried out in 2012 by com-paring UK Met O�ce hourly gauge reports, froma network of tipping-bucket rain gauges, to theradar accumulation (ACRR) composites, for a do-main covering the UK, for a trial period of threemonths. This included 2160 hourly time stepsat each of approximately 300 co-located radar /gauge pairs. Using the categories of "rain" and"no-rain" for each of the two observation types, a2x2 contingency table was constructed ( see Table2 ) and then the Peirce's Skill Score, as describedin Jolli�e and Stephenson (2003), was calculatedaccording to Equation 1.

Table 2: Construction of contingency table.

Radar GaugeRain No Rain Total

Rain A B A + BNo Rain C D C + DTotal A + C B + D A + B + C + D

PSS =AD −BC

(A+ C)(B +D)(1)

The trial was repeated for each of 5 di�erentclutter threshold settings and the PSS is shownfor each threshold in Table 3. Based on these skill

Table 3: Peirce's Skill Score for di�erent clutterthresholds.

Threshold PSS100% 0.5190% 0.5960% 0.640% 0.5920% 0.55

scores, it was concluded that the clutter �lter hasa bene�cial e�ect on the composite and that forthe UK Met O�ce radars, a threshold of 60% isoptimal.It was noted that the scores can be sensitive to

the method of combining radar pixels in the com-positing algorithm. For example, when calculat-ing the weighted average to arrive at a compositepixel value, if there is at least one "rain" echo inthe average, even if the other pixels are mostly "norain" echoes, the composite pixel will be treatedas a "rain" pixel.As a result of this study, a �rst-guess value of

60% was set for the whole of the European radarnetwork but ideally site-speci�c tuning would bedone for all radars because each radar has dif-ferent characteristics; many have been processedwith Doppler clutter �lters and / or have mini-mum dBZ thresholds applied, e�ectively �lteringout some of the clutter. The potential for further

3

tuning using this method will be limited by theavailability of recent rain gauge data from denserain gauge networks in the regions of interest butit is possible that the NMSs themselves will supplytheir own values for the thresholds.

3. bRopo QC algorithms

The BALTRAD software module known asbRopo became available in 2012 and included asuite of anomaly detection algorithms which weredeveloped originally at FMI by Peura (2002). Thissoftware was deployed on the operational Odysseysystem in March 2013.The algorithms included are listed individually

in Table 4. Odyssey currently employs the SPECK,EMITTER and SHIP algorithms but for e�ciencyreasons only the SPECK �lter is used for PPI scansabove 2.0◦ elevation.Each �lter contributes to an overall quality in-

dicator (a value between 0 and 255) which is sub-sequently used with a chosen quality threshold todecide which pixels are to be rejected before com-positing.Further details of the SPECK, EMITTER and SHIP

algorithms are given in the following sections.

Table 4: bRopo �lters.

Name Type of AnomalyBIOMET birds and insects near the radarSPECK noise; distinct specks

EMITTER line segmentsSUN long line segmentsSHIP ships (and aircraft)

VERT GRAD sea waves and ducting e�ectsMETEOSAT suspiciously warm dataDOPPLER non-continuous Doppler data

a. EMITTER �lter

Interference from Radio Local Area Networks(RLANs) can have a serious impact on the qualityof the measured re�ectivity data. These e�ects areusually obvious and take the form of radial spokes,

e.g. see Figure 3a. bRopo deals with this problemby using a computer vision method to detect linesegments within the image (Peura (2002)). Fig-ure 3b shows some examples of spokes, markedas red, which have been identi�ed in unprocessedradar data. A useful feature of this algorithm isits ability to allow some pixels within an identi-�ed sector, i.e. those that are more likely to bemeteorological echoes, to be preserved.

The algorithm can be con�gured using three pa-rameters which constrain the identi�cation to havei) a re�ectivity above a minimum dBZ, ii) to belonger than a speci�ed number of radial bins andiii) to be wider than a speci�ed number of degrees.

b. SPECK �lter

Isolated individual pixels or small clusters ofpixels, known as speckle, can arise due to vari-ous sources, including noise in the radar receiverapparatus. bRopo provides an algorithm to iden-tify speckle which is again based on the computervision method of segment identi�cation. A maxi-mum area of the segment, in pixels, can be speci-�ed in addition to a minimum dBZ threshold.

c. SHIP �lter

Some types of anomaly arise as a result of re-turns over the sea from highly re�ective surfacessuch as ships and waves. The ship anomalies arelikely to occur within busy shipping channels, suchas in the Gulf of Finland, the English Channel /La Manche or the Strait of Gibraltar. bRopo triesto detect these as small but intensive echoes inindividual pixels by considering relative changesin dBZ between pixels, although some additionalprocessing is required to be able to distinguish be-tween convective storms and ship or wave echoesbecause they have similar signatures in the radarimagery. Figure 4a shows an example of a monthlyaccumulation (based on ACRR composite data) fora region covering the Gulf of Finland, includingthree Finnish radars. It can be clearly seen inthis image the presence of straight lines, where theaccumulation is greater, which are known to be

4

(a) Image containing RLAN interference spokes

(b) Image containing identi�ed spokes (in red)

Figure 3: Images showing the identi�cation ofRLAN interference spokes.

the location of the shipping channels. The secondimage, Figure 4b, shows a similar accumulationbut which included radar data processed using theSHIP �lter.

(a) Without QC applied.

(b) With bRopo QC �lter applied.

Figure 4: Images showing the identi�cation andremoval of echoes from ships in the Gulf of Finlandusing monthly accumulation data derived fromACRR composites.

4. Comparison to NWP

An assessment of the PDH was carried out byMittermaier et al. (2008) using the UKMO NorthAtlantic European (NAE) model as a reference.This model has a 12km resolution and a domainwhich completely encloses that of the PDH. Themethod of assessment involved the conversion ofthe radar data on to the NAE grid, and then ananomaly di�erence �eld ∆η was calculated be-tween each daily model accumulation M and the

5

corresponding daily radar accumulation R:

ηR =R− µR

σR

ηM =M − µM

σM

∆η = ηM − ηR

where the means µ and the standard deviations σare based on the spatial averages over the wholedomain for each single day. The daily anomalydi�erences were then averaged over a long timeperiod (typically 40+ days) with the aim to re-move the transient model forecast bias errors.

The results of Mittermaier et al. (2008) suggestthat there were a number of unresolved problemswith the quality of the PDH data. This includedproblems related to software compatibility issuesand the intermittent availability of some radars,range e�ects resulting from the reduced radar de-tection ability at long range, seasonal biases, inparticular cold season bias over the Nordic andBaltic region, de�ciencies in the model to pre-dict convective scale NWP, orographic e�ects seenover the Pyrenees and the west coast of Norway,Vertical Pro�le of Re�ectivity (VPR) e�ects andregions a�ected strongly by anomalous propaga-tion, e.g. over the North Sea in the UK. It wasalso noted later that there were some serious geo-location errors which arose because of the manydi�erences in the de�nition of local radar coor-dinate systems and because, unlike Odyssey, thePDH ingested a mixture of single site data andnational composites.

Figure 4 shows the results of a new study, usingthe same method of calculating anomaly di�er-ences, to compare Odyssey daily accumulations,derived from ACRR composites, to the more recentUKMO EURO4km NWP model, a 4km down-scaled model which encloses the Odyssey domain.The main focus was to discover what biases, ifany, are present in the data and how any suchbiases compare to the earlier PDH study. Theareas of red indicate a positive anomaly di�er-ence (a strong model anomaly or a weak radar

Figure 5: Anomaly di�erences for April 2013 av-eraged over a 30 day period.

anomaly) while the areas of blue indicate a nega-tive anomaly di�erence (a weak model anomaly ora strong radar anomaly). As can be seen in the �g-ure, much of the area is white, where the absoluteanomaly di�erence is less than 0.2, indicating agood agreement between radar and model. Thereare some strong red areas, for example over thenorth coast of the Iberian Peninsula, in the Scot-tish Highlands and on the west coast of Norway.This signature is similar to that of the PDH andis thought to correspond to cases where the modelis overestimating the orographic enhancement and/ or where the radar is simply not seeing this ef-fect. There is also a strong red anomaly over theAlps but this is thought to be due to the inabilityof the French radars to see into this region due tothe local terrain (note there are no Swiss or Italianradars in the Odyssey composite). There are someisolated blue "dots" in Eastern Europe which arethought to be due to ground clutter and there arein various places some longer blue lines which are

6

likely to be due to RLAN interference. Also note-worthy is a blue ring around some of the Danishradars - the cause is not known but the e�ect ap-pears to have been resolved since the study wasdone. Please note also that there is an edge e�ectaround the Romanian radars which appears in redbut is not a positive anomaly di�erence. It resultsfrom the inability of the composite to properly in-terpolate S-band radar data at long range.

The anomaly di�erence information is usefulfor planning future improvements to the qual-ity control and corrections. With this in mind,anomaly di�erences were calculated for the April2013 case both i) without bRopo + clutter QCand ii) with bRopo + clutter QC. Figures 6a and6b show a region over the west of Europe, pri-marily over France, before and after the QC. Inthe uncorrected di�erence �eld, there are somestrong isolated negative anomaly di�erences whichare known to be due to wind farms. These were,by choice, not removed from the PPI scan datareceived on Odyssey. In the second image, thesesignatures appear to have been removed.

Based on the images shown, it was concludedthat the e�ect of the QC is generally localised, asmight be expected, i.e. the widespread regionsof red or blue anomaly di�erences were largelyuntouched and the main di�erences were seen interms of the removal of the isolated regions ofstrong negative anomaly di�erence, likely to bepermanent clutter or RLAN interference. Also, itwas noted from the images that the �lters do notperform as well as might be expected in some re-gions, e.g. the RLAN interference in Figure 6a isnot modi�ed in �gure 6b. However, this may becorrected with further tuning of the bRopo algo-rithms, which has not yet been done for radarsoutside the Baltic Sea Region.

5. Further Work

In the OPERA4 programme there are a num-ber of activities starting which aim to improvethe both the QC and the Quantitative Precipita-tion Estimation (QPE) of Odyssey. This includes

(a) Without QC applied.

(b) With QC applied.

Figure 6: Images showing the impact of clutterand bRopo QC on the anomaly di�erences for aregion in the west of Europe. Note that some ofthe strong negative anomalies (blue) are thoughtto be due to wind farms.

7

improvements to the compositing algorithm, VPRcorrections, solar monitoring (for elevation and az-imuth pointing corrections) and improvements tothe way quality information is shared and sub-sequently used to give appropriate weightings toradar data that is used in the composite. Thereare also plans to repackage single-site data, whichhas undergone QC, into single-volume �les whichwill then be redistributed to NWP users.The �ndings described in this report suggest

that further tuning of the bRopo and clutter �l-ters on a site-by-site basis would be e�ective andthat a more rigorous veri�cation study would bet-ter identify areas where improvements to the QCare needed.Odyssey SURF composites have recently been

used for input into European nowcasting mod-els as part of the Hazard Assessment based onRainfall European Nowcasts (HAREN) project(http://haren-project.eu/ ).

8

REFERENCES

Jolli�e, I. and D. Stephenson, 2003: Forecast Veri�cation: A Practitioner's Guide in AtmosphericScience. John Wiley and Sons, 50-51 pp.

Michelson, D., R. Lewandowski, M. Szewczykowski, and H. Beekhuis,2008: EUMETNET OPERA Weather Radar Information Model forImplementation with the hdf5 File Format. Available online from:http://www.eumetnet.eu/sites/default/�les/OPERA_2008_03_WP2.1b_ODIM_H5_v2.1.pdf.

Mittermaier, M., B. Macpherson, M. Naylor, R. Scovell, D. Harrison, and P. Earn-shaw, 2008: Assessing the OPERA Pilot Data Hub European Radar Composite forNWP Applications. Met O�ce Technical Report No. 521. Availabile online from:http://research.meto�ce.gov.uk/research/publications/papers/technical_reports/reports/521.pdf.

Peura, M., 2002: Computer vision methods for anomaly removal. Proceedings of the Second EuropeanConference on Radar Meteorology (ERAD), Delft, Netherlands, 312�317.

Saltiko�, E., H. Leijnse, and P. Novak, 2013: OPERA 4 - THE NEW PHASE OF OPERATIONALWEATHER RADAR NETWORK IN EUROPE. Proceedings of the 35th AMS Conference on RadarMeteorology, Breckenridge, Colorado, USA.

9