450 likes | 553 Vues
Utilizing satellite, radar, and rain gauge data, this study aims to produce high-resolution daily rainfall maps for South Africa. It involves the validation of satellite rainfall fields and merging data sources to enhance accuracy. Various techniques including texture analysis and discriminant functions are applied to improve rainfall estimates. Preliminary validation indicates adjustments needed for warm rain conditions. The study merges gauge and radar data with satellite information to create the Spatially Interpolated Mapping of Rainfall (SIMAR). Continuous enhancements are pursued to refine rainfall estimates.
E N D
Daily Mapping of 24 hr Rainfall at Pixel Scale over South Africa using Satellite, Radar and Raingauge Data Geoff Pegram1, Izak Deyzel2, Pieter Visser2, Deon Terblanche2, Scott Sinclair1 & George Green3 Civil engineering, University of KwaZulu-Natal, DURBAN, RSA METSYS, South African Weather Services, BETHLEHEM, RSA Deputy Director, Water Research Commission, PRETORIA, RSA
Presentation Overview • Objectives • Satellite information as a data source • Producing a satellite rainfall map for South Africa • Validation of Satellite Rainfall Fields • Use Gauge and Radar information to augment • Producing the Merged Rainfall Field • Verification of the Merged Rainfall Fields • How do we improve the product? • A spin-off is the ground-truthing of satellite data
Objectives South Africa has limited resources: • Sparse raingauge network • Patchy C-band radar coverage (non-Doppler) SO … • Use Satellite data to derive a Daily Rainfall Field over South Africa at highest spatial resolution • Combine rainfall estimates from METEOSAT, Rain Gauges and Radar to produce a Spatially Interpolated Mapping of Rainfall (SIMAR) over South Africa • Constantly seek ways to improve these estimates
Producing a Satellite Rainfall Map for South Africa • Overview of the Multi-Spectral Rain-Rate (MSRR) technique • Flow diagram of the MSRR algorithm layout • Components of the MSRR algorithm, particularly classification by texture
METEOSAT data used in SIMAR + + VIS WV IR =
Multi-Spectral Rain-Rate Estimation A: Mask out non-raining information • Cirrus, sun, speckles • Separate topography – cold versus warm coastal rain • When available, use VIS, WV & IR data to define mask • Use texture analysis to identify potential rain • Use image processing techniques: median filtering and edge detection to sharpen mask and clean up B: Use IR to estimate cold, intermediate and warm (coastal) rain
Infrared and Water Vapor Spectral Difference Cloud Mask • Negative Infrared and Water Vapor spectral difference field • Spatial Correspondence to strong Radar echoes • Mask =1 for deep moist cold cloud areas
Exploit Texture to Improve Estimate • Compute the Grey-Level Co-occurrence matrix (GLCM) at every point in the field • Thence compute the Angular Second Moment (ASM) at every point in the field • Defines a Mask that yields improved Accumulated Rainfall Estimates – comparable to TRMM estimates • Mask using WV when available, else IR
Texture Analysis of Infrared & Visible images IR VIS • Grey Level Co-occurrence Matrix (GLCM) texture features • Correspondence between certain texture features of Infrared or Visible cloud images to moderate Radar echoes
Discriminant Function based on LDA delineates rainfall areas Masked VIS Radar • Linear Discriminant Analysis (LDA), trained on Radar data, delineates possible rainfall areas
The IR → Rain-rate Relationship Cool: Rc= 0.45(230-IR) Medium: Rm= 0.00303(267-IR)1.85 Warm Stratiform: Rw=[{alog(73.32-0.173.IR)/2000] 0.625
Rain-Rate Estimationalgorithm COLD CONVECTIVE < 218K MIDDLE LAYER 219-267K WARM CLOUDS 268-278K NO Coastal ADAPTED IR POWER LAW RAIN-RATE DEEP CONVECTIVE ACTIVITY Sufficient slope NO Z-R derived Rhcs Rhms Rhws 0 HALF-HOURLY MSRR FIELD - Rhs Rhs ACCUMULATION Recursive speckle filter Rh*s = Half-hourly satellite rainfield Image smoothing filter 24-hr MSRR - Rs
Improvement of Daily Satellite Rainfall Fields • IR masked field & Final rainfield estimate • improve the vast spatial and quantitative overestimation of rainfall fields due to Cirrus contamination • improve the estimated spatial structure of daily rainfall fields • improve the detection of warm rain conditions, using algorithms not specifically designed for convective rain systems.
Preliminary Validation • validated with 300 1x1 min gridded raingauge values from Radar interpolated raingauge fields.
Verification of Satellite Daily Rainfall Fields Coastal < 1000m • Generally overestimated • Warm rain needs adjustment
Producing the SIMAR Merged Daily Rainfall Field • Why merge the rainfall fields? • Characteristics of each data source • Explain the merging techniques • Discuss the operational implementation ofthe merging routines
The steps in making a SIMAR map • Collect 24-hour rainfall data (up to 8:00 am) • Clean 5-minute radar-rainfall images and accumulate into a 24-hour mosaic • Process available satellite images of IR, WV & VIS from METOSAT 7 to get 24-hour estimate of rainfall over RSA • Combine: Gauge-Radar, Gauge-Satellite estimates • Post the map on the web by 11:00 am
Availability of Ground-based Rainfall Sensors • Weather radars: 11 C-band – except one S-band • Rain-gauges: 290 ± daily reporting climatological stations
South African Radar Networksuperimposed on the Mean Annual Rainfall map 1300 km N-S 1600 km E-W Area 1.2 Mkm2 Radar range is an (ambitious!) 200km
Automatically reporting raingauges 290 ±some outside RSA via HYCOS
Radar Explained Variance Field - VR Over-ambitious estimation of radar accuracy with range Needs revision Note FFT wrapping
Merged 24h Radar/Gauge Rainfall Field:R|G = (RK*VR+GK*VG)/(VR+VG)
Mean Satellite Field smoothed from Satellite Estimates at Gauge locations by Splines - SZ
Satellite Bias Skill Score Field – SB: compare S|G with R|G in 9x9 blocks at gauges – interpolate with Splines
Final Merged Rainfall Field:R|G,R,S = {R|G*(VRorVG)+ S|G *SB}/{(VRorVG)+SB}
SIMAR Part of the introductory SIMAR web-page Available daily by 11:00 am with previous day’s rainfall maps
How do we improve this? • Refine the merging of radar with gauge data to obtain better ground-truthing fields
An alternative method to improve SIMAR? • The explained variance method tried to be “fair” to gauges and radar • If we believe the gauges, then we want to condition the radar field onto the gauge readings as we did with the satellite images to get the S|G fields • We call it “Conditional Merging”
(a) The rainfall field is sparsely observed on a regular grid at rain-gauge locations (b) The rainfall field is also observed by radar on the regular grid - RR Description of the Conditional Merging technique Adapted from Ehret (2002)
(c) The rain-gauge observation are Kriged to obtain the best linear unbiased estimate of rainfall on the radar grid - MG (e) A rainfall field that coincides with the rain-gauge readings, while preserving the mean field deviations of the radar field is obtained as RR-MR+MG (d) The radar pixel values at the rain-gauge positions are Kriged onto the remainder of the grid to give a mean field - MR Adapted from Ehret (2002)
Correlated Field Contaminated Field Kriged Field Merged Field Explained variance weighting
Correlated Field Contaminated Field Kriged Field Merged Field Conditional merging
And finally a real cross-validation field experiment • Compare straight Kriging and Conditional Merging on 40 raingauges on a 4600 km2 catchment • Use cross-validation – estimation of daily total at each gauge separately using the remaining data
10 km Layout of the Liebenbergsvlei gauge network
Summary • We have made a start • Our Department of Water Affairs trusts the SIMAR fields enough to routinely use them in their Flood ForecastingDivision • Ongoing improvements are being made