Evaluation of Precipitation from Weather Prediction Models, Satellites and Radars

Charles Lin

Department of Atmospheric and Oceanic SciencesMcGill University, Montreal, Quebec, Canada

Collaborators

• Slavko Vasić

• Isztar Zawadzki

• Barry Turner

Objective

• To characterize and inter-compare precipitation retrieved from models, radars and satellites using spectral analysis and frequency distribution

Methodology

• Spectral domain– Scale decomposition using Fourier modes and

wavelets

• Frequency domain– Frequency distribution

• Domain of analysis

– Central and eastern US: 2,160 km × 2,160 km

• Weather prediction models

– Canadian GEM/HIMAP (10 km resolution)

– US ETA, NCEP (12 km)

• Radar data

– US composite (5 km resolution)

• Satellite

– GOES (5 km resolution)

1-hour accumulated precipitation 21 days in 2003, 2004

13 days with satellite data

2003• September 12, 13, 18, 27 • October 14, 17, 25, 26, 28 2004• May 21, 22, 30, 31• June 11, 24• July 3, 4, 5, 6• August 19, 20

Scale decomposition analysis

Accumulated precipitation for 11h-12h- October 26, 2003 -

2,160 km

2,1

60

km

Scale decomposition: Fourier and wavelet

Spectral amplitudes at different wavenumbers (kH = 2n)

Lag correlation analysis, 1-h accumulation (R = radar; G = GEM)

Frequency Distribution Analysis

• 21 days of 24 hourly accumulated precipitation → over 500 hourly samples

• Domain of 2,160 km × 2,160 km at 12 km resolution → Total of 180 × 180 = 32,400 pixels

• Bin hourly precipitation of each pixel, with bin width δ

• Count frequency of precipitation for each bin → histogram

Frequency distribution: Near power law dependence

Radar: ±1 standard deviation

Quartile analysis (3 days)

Total frequency

Compare skill of model forecasts and radar nowcasts

Categorical skill scores for 24-h GEM/HIMAP and ETA model forecasts

POD: Probability of detection

FAR: False alarm rate

CSI: Critical success index

Skill of Precipitation Forecasts from NWP Models and Radar Nowcasts

Golding (1998), Austin et al. (1987), Wilson et al. (1998)

Theoretical limit of predictability

Nowcasting methods

Numerical weather prediction models

Experimental Design

6 9 12 15 18 21 24

Five 9-hour radar nowcasts:0300, 0600, 0900, 1200 & 1500 UTC

24-hour GEM-HIMAP & ETA forecasts

Time [h]

21-day averaged skill scores(1-hour accumulated precipitation, 2003-2004)

- Threshold: 0.1 mm/h -

Lin, Vasic, Kilambi, Turner and Zawadzki (2005)

Summary

• Scale decomposition and frequency analysis of model and radar precipitation

• Enhanced falloff of power of model precipitation at high wavenumbers compared to radar

• Significant loss in power of models at scales ≈ 10 Δx• Approximate power law dependence for frequency

distribution• Skill of 24-h model forecast is relatively constant, and is

better than radar nowcasts after a lead time of about 6-8 hours

Thank you!