Overview of QPE/QPF techniques and hydrological applications Siriluk Chumchean Department of Civil Engineering Mahanakorn University of Technology Typhoon Committee Roving Seminar 2011, Malaysia (20-23 September 2011) 1
Talk Structure Lecture A1: QPE techniques and development Real-time radar rainfall estimates t Case studies Lectuer A2: QPF techniques and development QPF techniques (Radar-based nowcasting) Operation and research development systems QPF accuracy assessment Case study : Bangkok rainfall forecasting system Lecture A3: Hydrological Applications Cases studies 2
Real-time Radar Rainfall Estimation i An Overview 3
Why do we need weather radar? Measure rainfall continuously covering large area Provide fine spatial and temporal resolution Can be used to project the storm trajectory, i.e., provide a short-term storm forecast 4
Basic concept of measuring rainfall using radars r p = C * K * Z r 2 z = AR b where: A,b radar parameters; Z reflectivity; R - rainfall 5
Volume scan data Range from radar (km) 6 Height above grou und (km)
Rhine Pixel 1 km² Radar site Intensity class scale (0 200 mm/hr) 7
Simplified scheme of radar processing Measurement of rain drops Reception of the reflection of the impulse sent out Radar image in polar coordinates - continuous values Transformation Radar products, e.g. in cartesian coordinates - values in 256 classes 8
Polar and cartesian coordinates Near distance <-> far distance Due to the radar measurement in polar coordinates, the data near to the radar are denser than far from the radar. When transforming polar data to a cartesian square grid, close to the radar several polar pixels are used. The spatial resolution of the square grid data is less than that of the originally measured polar data. Far from the radar, one polar pixel may be the original for several cartesian grid pixels. Here, the density of the data is less and therefore measurement errors have a higher impact on the cartesian data. 9
PPI PPI reflectivity map is extracted from the raw reflectivity data from the beam at a particular elevation angle. 10
CAPPI Constant Altitude Plan Precipitation Indicator (CAPPI) 11
Problems and Limitations Are parameters measurable? Are parameters stable? Does reliability stay constant with distance away a from the radar? What about obstructions (buildings, mountains)? What about multiple raindrops and rain-clouds in the path of the beam? What about the effect of the earth s curvature? What about differences in rain profile on ground versus high elevations? r Z = AR b where A,b radar parameters Z reflectivity R - rainfall 12
Problems of using weather radar? 1. Rfl Reflectivity tiit measurement error ground clutter attenuation vertical reflectivity it profile beam geometry 2. Reflectivity-rainfall rate conversion error 3. Residual errors when compared with rain gauges 13
Some difficulties Ground-clutter clutter, bright-band band, hail, attenuation, range dependent bias not so difficult to remove as they can be noticed easily No screening PE Partial screening PE Complete screening 14
Some difficulties Distance vs height of radar beam 8000 Heig ht of the rad dar beam (b beam centr re) [m] 7000 6000 5000 4000 3000 2000 1000 0 1 0,5 0 0 50 100 150 200 250 Distance to radar location [km] 15
Some difficulties Increasing uncertainty in measured reflectivity as a function of range 1.0 tion Function (C CDF) 08 0.8 0.6 Increasing range Cummu ulative Distribut 0.4 0.2 Gauge rainfall Measured reflectivity 0.0-10 -5 0 5 10 15 20 25 30 35 40 45 50 55 Measured reflectivity (dbz) / Rainfall rate (dbr) 16
Some difficulties Range dependent errors 17
Some difficulties Range dependent errors 1 st Range 2 nd Range 1 st Range 2 nd Range 2 2 2 σ σ = GR = 0.1320 2 σ 0.1320 σ = GR = 0.0003 ( R 70 ) 0. 1320.0003 ( R 70 ) 0 + + 0. 1320 GR 70 70 GR 0 0 σ 2 GR i = 1 N { N [ () ()] () } 2 logrg i logrr i RG i > r i = 1 Anagnostou et al. (1999) 18
Some difficulties Attenuation example Rainfield attenuated (up to 6 classes) Rainfield without attenuation Marienfeld Radar Flechtdorf Radar Essen Einfalt (2004) 19
Some difficulties Bright-band Fabry et al. (1994) 20
Some difficulties..effect of storm types on Z-R relation Z = α 0 N o e Λ α Λ D D R = N e π ν ( D ) D 6 dd 4 3 o 2 0 D 3 dd Convective Stratiform Z = 35 dbz R = 7 mm/h Z = 35 dbz R = 10 mm/h 21
Z-R relationship for each rain-cloud Parameters a and b varies on rain drop size distribution. Cumulus Cumulonimbus Nimbostratus 22
Z-R relationship of each type of rain-cloud (Calibration) Radar refle ectivity (dbz) Cumulus Radar refle ectivity (dbz) Cumulonimbus Radar refle ectivity (dbz) Nimbostratus Rain gauge rainfall (mm/hr) Rain gauge rainfall (mm/hr) Rain gauge rainfall (mm/hr) Radar re eflectivity (db BZ) Cumulus Radar re eflectivity (db BZ) Cumulonimbus Rain gauge rainfall (mm/hr) Rain gauge rainfall (mm/hr) Chumchean et al. (2009) 23
Z-R relationship of each type of rain-cloud Rainy season Cumulus Cumulonimbus Nimbostratus Z=29R 1.5 Z=55R 1.5 Z=208R 1.5 Z = 30 dbz Z = 30 dbz Z = 30 dbz R = 10.58 mm/h R = 6.90 mm/h R = 2.85 mm/h Summer Cumulus Z=38R 1.5 Cumulonimbus Z=90R 1.5 Chumchean et al. (2009) Z = 30 dbz R = 8.93 mm/h Z = 30 dbz R = 4.99 mm/h 24
Some difficulties 8 9 Variations in Z~R 4 3 relationships for different 2 1 storm types ht (km) Heig 8 7 6 5 0 stratiform tif 0 5 10 15 20 25 Reflectivity (dbz) Height (km) 9 8 7 6 5 4 3 2 1 0 convective 0 5 10 15 20 25 Reflectivity (dbz) 25
Some difficulties Residual errors results in a mean-field-bias Exhibits persistence from one time-step to the next Very important in real-time estimation Difficult to remove using physical relations 26
Philosophy Estimated radar rainfall must first be corrected for reflectivity measurement errors and Z-R conversion errors based on the physical methods, and then a statistical method will be used to remove the average difference (mean field bias) between radar estimates at the rain gauge locations and the corresponding gauge rainfall amounts. 27
Real-time radar rainfall estimation method Reflectivity measurements Z-R conversion Rain gauge adjustment Record reflectivity field in 3D polar coordinate at operational temporal resolution Exclude reflectivity that are greater or lower than the maximum/minimum reflectivity thresholds Conversion to CAPPI Cartesian coordinate at altitude below bright-band level Remove the effect of ground and sea clutter Scaling correction Instantaneous storm classification Instantaneous convective/ stratiform reflectivity Account for the storm movement within an hour Z-R conversion Z=AR b Instantaneous convective/ stratiform radar rainfall Initial hourly radar rainfall estimates Accumulate into hourly radar rainfall Estimate hourly gauge-radar bias adjustment factor (AF) based on Kalman filtering approach proposed p by (Chumchean et al., 2003) [under review] Final hourly radar rainfall = AF initial radar rainfall estimates Chumchean et al. (2006) 28
Real-time radar rainfall estimation Scaling correction (to remove range-dependent bias) + Storm classification (to remove bias in Z~R relation due to the dominant storm type) + Correction of residual mean-field bias 29
Data C-band Kurnell radar 10-minute, 1km 2 resolution Analysis period November 2000 to April 2001 254 hourly rain gauges (SWC,SCA,BoM) 30
Correction of range dependent bias due to radar beam geometry 31
Application of scaling in measured reflectivity tribution Fu nction (CDF F) Cumm mulative Dis 1.0 0.8 0.6 0. 4 0.2 0.0 0-20 km 20-40 km 40-60 km 60-80 km 80-100 km 15 20 25 30 35 40 45 50 Measured reflectivity (db Z) Assume simple scaling holds for measured reflectivity 32
Scale transformation function The proposed scale transformation formula is : 1.0 Cummulati ve Distributi on Function n (CDF) 0.8 0.6 0.4 0.2 0.0 0-20 km 20-40 km 40-60 km 60-80 km 80-100 km 15 20 25 30 35 40 45 50 55 Transform ed reflectivity (db Z) 33
Range-dependent bias correction Using simple-scaling theory 1 : 27 8 Reflectivity (db BZ) 26 25 24 6 4 2 l (mm/h) gauge rainfall Rain 23 0-20 20-40 40-60 60-80 80-100 R ange interval (km ) 0 M easured reflectivity Transformed reflectivity R ain gauge rainfall 1 Chumchean et al, 2004, J Atmospheric and Oceanic Technology 34
Effectiveness of scale transformation function in correcting range dependent bias in radar rainfall 2.60 2.40 Gaug ge-radar ra atio (G/R) 2.20 2.00 1.80 160 1.60 1.40 Slope = 0.3 % Slope = 8.9 % 1.20 0-20 20-40 40-60 60-80 80-100 Range interval (km) 35
Effect of storm types on radar rainfall Estimates (Z-R conversion error) 36
Instantaneous pixel classification Apply Steiner et al.(1995) storm classification method to use for separation of instantaneous reflectivity of each pixel into convective or stratiform components convective Revise the classification criteria to be suitable for instantaneous reflectivity of the Kurnell radar 37
Parameter for pixel classification (after Steiner et al., 1995 )............ convective centre... background radius......... convective radius................... x............... minimum convective......... threshold*......... maximum stratiform...... threshold* convective radius Background radius 11 km * = proposed additional classification parameter 38
Classification results: calibration Using modified Parameters based on Steiner et al. (1995) Using modified pixel based classification approach 39
Classification results: verification Using modified Parameters based on Steiner et al. (1995) Using modified pixel based classification approach 40
Verification using VPR Height (km).. Height (km).. 11 10 11 10 9 8 7 6 5 4 3 2 1 0 9 8 7 6 5 4 3 2 1 0 (a) Stratiform Calibration: 20 Apr 01 at 14:35 UTC Verification1: 30 Jan 01 at 12:25 UTC Verification2: 10 Feb 07 at 18:35 UTC 0 5 10 15 20 25 30 35 (b) Convective Reflectivity (dbz) Calibration: 20 Apr 01 at 14:35 UTC Verification1: 30 Jan 01 at 12:25 UTC Verification2: 10 Feb 07 at 18:35 UTC 0 5 10 15 20 25 30 35 Reflectivity (dbz) 41
Correcting for mean field bias in real-time radar rainfall estimates using Kalman Filtering techniques? Rain gauge rainfall (G) = AF x Radar rainfall (R) AF = adjustment factor or mean field bias 42
Advantage of Kalman filtering technique over the simple G/R 1) It accounts for the noise in the measurements when updating the mean bias. 2) It provides an estimateof t the error in the computed tdbias. 3) It combines an estimate of the bias and its error variance made an hour earlier with the current measurements and its estimated measurement error variance to compute an updated bias estimate and new forecast for the next hour. Time update ˆ β t and - P t Measurement update βˆt and P t 43
Residual mean-field bias correction 1 Logarithmic Mean Field Bias (β) Assumed β exhibits Markovian dependence (Kalman process model assumed AR1) Kalman measurement 1 error model specified 0 Bias 6 5 4 3 2 β = t log 10 n 1 G i, t n i= 1 Ri, t K alman Filter Observed (G/R) as function of range 2-1 0 100 200 300 400 500 600 Time step (hours) 1 Chumchean et al, 2006, Journal of Hydrology, 2 Chumchean et al, 2003, Physics and Chemistry of the Earth, 28(3), 27-39. 44
Radar rainfall error variance model Chumchean et al. (2003) σ 2 Yi t = 0.015 G +, t 0.14 ; for r i 55 km σ 2 ( ri 55) = 0.015 G + 0.13 0.14 ; for r > 55 km i Y i +, t t p 45
Comparison of observed G/R and the Kalman-filter estimated bias (Calibration) 6 5 4 Kalman Filter Observed (G/R) lag 1 correlation coefficient of innovation sequence = 0.047 Bias 3 2 1 0 3 Nov 00 18 Dec 00 12 Dec 00 30 Jan-1 Feb01 19-22 Apr 01-1 0 20 40 60 80 100 120 140 160 180 200 220 240 Time step (hours) 46
Comparison of observed G/R and the Kalman-filter estimated bias (Validation) 6 5 K alman Filter Observed (G/R) lag 1 correlation coefficient of innovation sequence = 0.08 4 Bias 3 2 1 0-1 0 100 200 300 400 500 600 Time step (hours) 47
Application Stepwise application of error correction strategies Base reflectivity data free of effects of ground-clutter, bright-beam, hail Total rain gauges - 260 Cross-validation performed by leaving fraction of rain gauges to evaluate predictive uncertainty t Performance statistic - Root Mean Square Error (RMSE) 48
Results Climatological Z-R parameters of convective and stratiform rainfall (Z = AR 1.5 ) A parameter Calibration strategies No scaling correction Scaling correction Convective Stratiform Convective Stratiform No-classification 128 146 Hourly pixel classification 150 93 172 100 49
Results 4.10 390 3.90 3.70 GHNBO GHNBO+SC GHNBO+SC+ZR RMSE (mm m). 3.50 3.30 3.10 2.90 2.70 Without bias adjustment Kalman Filtering Sample G/R Gauge-based adjustment methods GHNBO Base data SC Scaling correction ZR Storm classification 50
Results a) No bias correction GHNBO GHNBO+SC GHNBO+SC+ZR b) sample G/R GHNBO GHNBO+SC GHNBO+SC+ZR c) Kalman filtering GHNBO GHNBO+SC GHNBO+SC+ZR 51
Conclusions A stepwise decrease in RMSE with added levels of error correction Correcting mean-field bias more important that correcting the other sources of errors Storm classification relatively more important than the correction of range dependent bias Kalman filtering much better than use of sample G/R correction, particularly l when calibration rain gauges are few 52
Conclusions Range dependent bias and storm-classification, though small, are certainly not insignificant Complete model (after mean-field bias correction) explains more than 50% variance of gauge rainfall This may be the best we will ever have remember we are comparing 1kmx1km grid averages to rainfall at a point! 53