SeaWinds Empirical Rain Correction Using AMSR

1
SeaWinds Empirical Rain Correction Using AMSR

• January 17, 2005
• Bryan Stiles, Svetla Hristova-Veleva, and Scott
  Dunbar

2
Outline

• Method Overview
• Data set description
• Variance computation change in objective function
• Rain correction methods
• Performance Summary
• Metrics
• Direction and Speed Histograms of DIRTH vectors
• 2-D NCEP/Retrieved Relative Direction Histograms
• Cross Track Bias (by liquid and speed)
• Speed Bias (by liquid and speed)
• RMS Direction Difference (by liquid and speed)
• Discussion

3
Method Overview

• The entire SeaWinds mission was processed 3 ways.
• Climatological attenuation correction only (SCAT)
• Physically based rain correction (PHY)
• Empirically based rain correction (EMP)
• The objective function was modified for all three
  cases.
• Log(var) term was put in.
• Variance was modified so that
• assuming the standard deviation of the
  backscatter correction b was 50
• noting that measurement noise was multiplied by
  the attenuation correction a.

4
Method Overview

• PHY
• Uses physical models of attenuation and
  backscatter to compute a and b from AMSR liquid,
  vapor, SST, and rain rate.
• s is the splash ratio as a function of rain rate
• EMP
• Estimates a and b as function of liquid, vapor,
  and SST using NCEP winds collocated with SeaWinds
  ?0 values.
• To avoid biases due to NCEP errors
• Scaled a to match physical liquid0 values.
• Scaled b so that minimum backscatter was 0.

5
Performance summary

• Rain free cases are not affected significantly by
  the corrections.
• Both rain corrections improve speed bias and
  reduce cross track direction preference.
• RMS direction performance is mixed
• Nearest RMS direction difference from NCEP is
  increased by the correction techniques, but that
  may be explained by
• The number of ambiguities decreases in the
  corrected cases for liquids over 1 mm.
• Selected RMS direction difference has little
  change
• One would expect improvement due to reduced cross
  track preference.
• Lack of improvement may indicate an additional
  directional noise imparted by the corrections.
• DIRTH RMS direction difference is significantly
  decreased especially for the empirical
  correction.
• DIRTH tends to smooth out directional noise in
  the corrected winds.

6
Performance summary (cont)

• RMS speed performance
• RMS speed differences (not shown) decrease due to
  speed bias improvement
• Speed variance increases especially for the
  empirical case.

7
1-D Direction and Speed Histograms

• Plot format
• NCEP Histograms were plotted together with the
  DIRTH vector histograms for each correction
  method.
• Direction and Speed Histograms were computed for
  varying
• Correction Strategy (line color)
• Geographic region (plot in slide)
• Liquid Range (slide)
• Percentage of Data in each liquid range is noted.
• Observations
• Corrections tend to match model direction
  histograms better
• Corrections tend to follow model wind speed
  trends by geographic region
• DIRTH creates cardinal direction spikes
  (investigating )

8
Rain Free, 86 of data
9
0.2-0.4 mm, 8.6 of data
10
0.4-0.8 mm, 3.3 of data
11
0.8-1.5 mm, 1.0 of data
12
1.5-3.0 mm, 0.4 of data
13
3.0-15 mm, 0.03 of data
14
2-D Direction Histograms

• Two dimensional histograms of retrieved direction
  and NCEP direction, relative to the s/c flight
  direction.
• Demonstrates the removal of rain-related
  artifacts e.g. cross-track directions.
• Histograms were computed for varying
• Correction method (slide)
• Liquid range (plot in slide)
• Choice of DIRTH, Selected, or Nearest (slide)
• SCAT-only histograms repeated as the top row of
  each slide for comparison.
• SCAT-only histograms differ for EMP and PHY
  slightly due to differences in flagging.

15
EMP, Nearest
16
PHY, NEAREST
17
EMP, Selected
18
PHY, Selected
19
EMP, DIRTH
20
PHY, DIRTH
21
Plot Formats

• Metrics from here on are
• Plotted as a function of
• Liquid (x-axis) (full range or 0-3 mm)
• Due to a bug liquids values on the x-axis are 4
  times the true values.
• NCEP speed (multiple plots in slide)
• Correction method (line color, cyanEMP, redPHY,
  blackSCAT, dotted blackSCAT w/o log(var))
• Computed for 200 orbits of SeaWinds data.
• Plots for full liquid range and 0-3 mm (99.97 of
  data) are on separate slides.

22
Speed Biases

• Metric Definition
• Selected speed - NCEP speed
• Performance Summary
• Nearest and DIRTH speed biases (not shown) are
  similar.
• Significant improvement for all but highest wind
  speeds.
• Even heavy rain cases show improvement.
• Correction imparts little or no change for rain
  free data.
• Slight change in rain free biases with addition
  of log(var).

23
Speed Bias, All Liquids (Liquid x-axis values
are 4X true liquid values)
24
Speed Bias, 1-3 mm(Liquid x-axis values are 4X
true liquid values)
25
Cross Track Direction Bias

• Metric Definition
• The average amount closer to the cross swath than
  NCEP in degrees.
• Angle between NCEP and cross swath minus the
  angle between selected and cross swath.
• A positive value indicates the cross track
  direction is preferentially retrieved.
• Performance Summary
• Corrections reduce rain induced preference for
  cross swath direction.
• Nearest and DIRTH performance is similar to
  Selected.
• Full liquid range and 0-3 mm plots are shown.

26
Cross Track Bias, All Liquids (Liquid x-axis
values are 4X true liquid values)
27
Cross Track Bias, 1-3 mm(Liquid x-axis values
are 4X true liquid values)
28
RMS Direction Difference

• Nearest, Selected, and DIRTH stats are plotted.
• Performance Summary
• Correction increases Nearest RMS direction
  difference in rain.
• Number of ambiguities are reduced.
• Correction noise is added.
• Selected direction difference - no change
• Correction noise competes with removal of cross
  track preference.
• DIRTH direction difference - improvement with
  rain correction
• Best case DIRTH spatially smooths correction
  noise.
• Worst case DIRTH smooths directional features in
  rain.

29
RMS Direction Diff, Nearest (Liquid x-axis
values are 4X true liquid values)
30
Number of Ambiguities (Liquid x-axis values are
4X true liquid values)
31
RMS, Direction Diff, Selected(Liquid x-axis
values are 4X true liquid values)
32
RMS Direction Diff, DIRTH (Liquid x-axis values
are 4X true liquid values)
33
RMS Dir. Diff, DIRTH, All Liquids (Liquid x-axis
values are 4X true liquid values)
34
Discussion

• What further validation is needed?
• What can change analysis tell us?
• Should change analysis look at
• DIRTH solution performance?
• Cross Track Direction Bias?
• What can we compare with besides NCEP? Buoys?