Estimating Bias of SatelliteBased Precipitation Estimates Relative to In Situ Measurements

1
Estimating Bias of Satellite-Based Precipitation
Estimates Relative to In Situ Measurements

• Thomas Smith1
• Phillip A. Arkin2
• George J. Huffman3
• John J. Bates1
• 1. NOAA/National Climatic Data Center/NESDIS
• 2. ESSIC, Univ. of Maryland, College Park
• 3. Science Systems and Applications, Inc. and
  NASA Goddard Space Flight Center

2
Precipitation Bias

• Bias adjustments are needed for satellite-based
  precipitation.
• Bias Evaluations estimates of satellite bias
  relative to gauge data.
• Bias Uncertainty how good are the bias estimates
  given the available gauges.
• Oceanic Precipitation Bias can satellite
  estimates be adjusted over oceans to minimize the
  bias.

3
Bias Evaluations

• Bias Satellite Gauge difference.
• Analysis of monthly differences using a
  large-scale optimum interpolation (OI).
• Bias spatial scales are needed for analysis.

4
Bias Correlation Scales

• Zonal Averages at each latitude.
• Red satellite bias w.r.t. gauges (land only).
• Black satellite to satellite difference scales
  (land and sea).
• Scales are similar, mostly 1000 km .
• Use constant scales of 750 km in both directions,
  to minimize excessive spreading of bias.

5
Examples Zonal Averages

• 1996-2003 Bias typically a few mm/day.
• OPI bias smallest.
• SSMI combined and GPI biases have opposite signs,
  some annual cycle.

6
Examples 1996-2003 Averages

• Positive bias over central Africa.
• Negative over east Asia and Hawaii.
• Different sign biases may cancel when data are
  combined.

7
Differences that Bias Corrections Make

• 1996-2003 Average uncorrected and corrected
  SSMIc.
• Most regions show little difference.
• Large differences over central Africa, N.W. North
  America.

8
Bias Uncertainty

• How well can the OI analysis reflect the bias?
  (Input data 2.5o and monthly OI uses data from
  12.5o centered region.)
• How does bias uncertainty change with the number
  of raw biases analyzed?
• What is the bias uncertainty when there are no
  gauges in a region?

9
OI Relative Error

• Set all raw biases 1 and analyze, compute error
  change with n number of raw biases.
• An exponential function estimates the error as a
  function of n.

Error
n
10
RMS Bias

• RMS Bias (RMSB) is the typical bias magnitude (no
  correction), from only near gauges, global.
• RMS difference between each satellite and all
  other satellites (RMSDs, land sea) is mostly
  comparable to RMSB. OPI and TOVS have lower bias
  than the others, which have similar magnitudes.

RMSB (mm/day) computed using n 20 for each
satellite,1996-2003 and globally. SMMI RMSB is
the same for all. The RMSD between each
satellite and all other satellites is also
shown. Satellite RMSB RMSDs OPI 0.9
2.0 AGPI 1.5 1.7 GPI 2.0
2.1 SSMI 1.8 2.0 SSMI/TOVS 1.5
1.8 TOVS 1.2 2.0
11
Bias Error from a Merged Analysis

• Merged error depends on how correlated the biases
  are.
• All uncorrelated
• All correlated
• Part correlated

12
Bias Correlations

• Measured spatial correlations from well-sampled
  regions (n 20).
• Negative correlations reduce bias error by
  canceling positive and negative errors.
• Most have weak positive correlations (r usually lt
  0.5, some bias error reduction from use of
  multiple satellites).

13
Test January 2000, all Satellites

• Observed correlations, equal weighting for all 8
  satellite estimates.
• No bias analysis ? no corrections (oceans).

• More satellite coverage ? less error (tropics).

14
High-Latitude Gauge Biases Work in Progress

• Bias errors from evaporation and blowing snow.
  They may offset each other, but they can also be
  as large as 30 of the mean (Groisman et al.).
• Gauge adjustments are recommended for high
  latitudes. Uncertainty should be factored into
  bias uncertainty.

15
Oceanic Bias Adjustments Work in Progress

• No oceanic gauges except scattered islands, near
  coasts, and on some buoys. An oceanic low-bias
  estimate needed for adjustments.
• TMI satellite estimates were found to have little
  bias compared to tropical Pacific buoys (Bowman
  et al. 2003).
• Over oceans, Bias adjustment relative to TMI (or
  a satellite with least bias relative to TMI).
  Uncertainty from uncertainty estimate of TMI or
  of the satellite.
• TMI limited in time (beginning 1998) and to
  between about 40ºS and 40ºN.

16
Summary and Conclusions

• Using available data, satellite precipitation
  bias is evaluated near gauges.
• Bias spatial scales are usually large (gt 1000
  km), but may be smaller near coasts.
• Bias may be reduced by direct adjustment and by
  combining analyses with uncorrelated biases.
• Oceanic estimates may be adjusted to a low-bias
  satellite estimate, such as TMI.
• Because of uncertainties in the base analyses
  (gauges, TMI) the absolute bias is difficult to
  define. Adjustment is relative to a best
  estimate to minimize bias.