Infilling Radar CAPPIs Geoff Pegram, Scott Sinclair, Stephen Wesson & Pieter Visser
What we’ve done … We can remove ground-clutter and have improved the estimation of rainfall by radar at ground level We have refined the merged fields of radar with raingauge data We think that the combined fields are good out to 75 km from the radar with a reasonably dense network of gauges, but we’re happy to take advice!
NATIONAL WEATHER RADAR NETWORK see Deon’s presentation Existing radars Radars added (2004) Planned radars
Problems with Radar CAPPI Data Parts of radar volume scan where data is unknown Rainfall estimates at ground level unknown Ground clutter contamination can be extensive Results in poor quality rainfall estimates
Summary of Infilling Strategy Choose Rainfall classification algorithm Devise Bright band correction algorithm Semivariogram parameters determined by rainfall type. Climatological semivariograms. Ordinary and Universal Kriging to extrapolate rain information. Universal Kriging utilised in mixed zone. Cascade Kriging to progressively infill data down to ground.
Rainfall Classification Rainfall separated into two zones: (1) Convective Zone (2) Stratiform Zone Criteria of classification set out in table below.
Examples of Rainfall Classification Classified Images Reflectivity Images 18 km 0 km CROSS SECTION X-X dBZ Classification X
Characteristics of Classified Rainfall Stratiform – low average height, low variability and intensity. Convective – considerable vertical extent, high variability and intensity. Increase of rainfall intensity nearer ground level
Bright Band Correction Bright Band – melting snow & ice crystals Need to correct bright band to obtain accurate rainfall estimates at ground level Proposed correction procedure: pixel by pixel approach Corrected Climatological Profile New Rainfall Estimate at Ground Level Rainfall Estimate at Ground Level Climatological Profile Affected by Bright Band with Extrapolation to Ground Level CAPPI level affected by bright band corrected Climatological Profile Correction Procedure CAPPI level affected by bright band Climatological Profile Affected by Bright Band Height (km) Typical Climatologial Profile 4 km 3 km 2 km 1 km Reflectivity (dBZ)
Bright Band Correction Testing of bright band correction Results: improved rainfall estimates at ground level 2km CAPPI before bright band correction 2km CAPPI pixels marked which are affected by bright band 2km CAPPI after bright band correction
Semivariogram Modeling Semivariogram model parameters computed for convective & stratiform rain in horizontal & vertical directions Reflectivity Image SILL RANGE 30km
Table of Average Parameters: Graphs indicating clustering of alpha and correlation length parameters by rainfall type (15 Rain Events over 4 different years) Table of Average Parameters:
Sensitivity Analysis of Stratiform, Horizontal Parameters Convective Cluster: Lc , c Stratiform Cluster: Ls , s Missing data infilled with different combinations of α and L that represent the spread of parameter values. No significant difference between Kriging estimates returned for spread of parameter values L, α + σα L, α - σα L, α L + σL, α L - σL, α L, α + σα L, α - σα L, α L + σL, α L - σL, α
Kriging to Infill Missing Rain Data KRIGING used to extrapolate/interpolate horizontal and vertical rainfall information to infill unknown data points Considered to be the optimal technique for interpolation of Gaussian data Computational Efficiency & Stability: Nearest 25 rainfall values used in Kriging Singular Value Decomposition (SVD) with trimming of small singular values to ensure computational stability
Summary: Three Rainfall Zones Stratiform Zone All controls stratiform. OK used to infill target point. Convective Zone All controls convective. OK used to infill target point. Mixed Zone Controls stratiform & convective. UK used to infill target point. stratiform pixel convective pixel target pixel
Validation: Universal & Ordinary Kriging Observed Rainfall Kriging Estimate Reflectivity (dBZ) Rainrate (mm/hr) Rainrate (mm/hr) Reflectivity (dBZ) All Errors Rainrate (mm/hr) RAINRATE ERROR MAPS Absolute Error Reflectivity (dBZ) & |Rainrate Errors| (mm/hr) 100 50 Stratiform Rainrate Errors (mm/hr) Convective Rainrate Errors (mm/hr) Mixed Rainrate Errors (mm/hr)
UK & OK Effectiveness UK & OK tested on three different rainfall zones on a variety of instantaneous images Effectiveness evaluated by comparing mean, and Σdifference2 of estimated & observed rainfall UK in mixed zone provides a superior estimate than OK and reduced Σdifference2
KRIGING directly to Ground Level Unexpected problems with CAPPI edges Higher Kriged values returned than expected and serious discontinuity also evident Example: 24 hour accumulation Rainfall Accumulation (mm) 100 80 60 40 20 Discontinuities Inflation of Kriged values
Radar Volume Scan Data After Cascade Kriging 3D CASCADE KRIGING EXAMPLE Radar Volume Scan Data Radar Volume Scan Data After Cascade Kriging
CASCADE KRIGING: Ground Clutter Ground Clutter contaminates radar volume scan data up to 5km above ground level. Ground Clutter 3km above ground level Ground Clutter infilled on 3km level Reflectivity estimation at ground level
Testing: Ground Clutter Infilling Original Reflectivity Image Ground Clutter Map Superimposed Ground clutter segments to be estimated Estimated reflectivity data Tested on 3D Bethlehem ground clutter map Ground clutter placed onto known rain Tested on three different rain events over 24hr period Convert to rain rate by Marshall-Palmer equation Store estimated and observed rain rate values and proceed to next image in sequence
Results: Ground Clutter Infilling Accumulations over 6, 12 and 24 hours show close correspondence between observed and estimated values
Testing: Rainfall Estimation at Ground Level MRL5 Weather Radar Bethlehem Raingauge Locations Liebenbergsvlei Catchment Polokwane Irene Ermelo Bloemfontein Bethlehem De Aar Durban 2 L Selection Range Radar Pixel Locations 1 km Rainguage Locations East London Port Elizabeth Cape Town Extrapolated radar estimates at ground level compared to raingauge estimates
Results: Rainfall Estimation at Ground Level Two rain events selected of different rainfall types – 12h & 24 h accumulations Results indicate fair estimation of rainfall at ground level We’ve got a handle on the errors
The Conditional Merging algorithm To combine radar and gauge data optimally: Krige the gauges to give best guess field, MG Krige the radar pixels at gauge locations, MR If RR is the measured radar rainfield, Conditional Merged Field is: RC = RR + MG – MR which coincides with the gauges and interpolates intelligently
Conditional merging Check spelling and wording here. 29/06/2005
Simulation experiment Check spelling and wording here. 29/06/2005
Simulation experiment Check spelling and wording here. 29/06/2005
A real cross-validation field experiment Compare straight Kriging and Conditional Merging on 45 rain gauges on a 4600 km2 catchment Use cross-validation – estimation of daily total at each gauge separately using the remaining data Check spelling and wording here. 29/06/2005
Layout of the Liebenbergsvlei gauge network Check spelling and wording here. 29/06/2005 Bethlehem
Comparison of daily mean errors Check spelling and wording here. 29/06/2005
Errors with range – how good is the radar? 22 new gauges 4 different days of accums
Rainfall 9 January 2005
Rainfall 12 January 2005
Rainfall 13 January 2005
Rainfall 21 January 2005
Concluding Remarks With intelligent extrapolation and climatoloical variograms we can get good ground estimates With conditional merging of radar and gauge data we can get good interpolation to adjust for errors in the Z-R formula Within 75 km from the radars, we can offer sound areas in varying climates and land cover in our expanding radar and gauge network Check spelling and wording here. 29/06/2005