The nonlinear patterns of North American winter temperature and precipitation associated with ENSO

1
The nonlinear patterns of North American winter
temperature and precipitation associated with ENSO

• Aiming Wu, William W. Hsieh
• Dept. of Earth Ocean Sciences,
• University of British Columbia
• and
• Amir Shabbar
• Meteorological Services of Canada
• Downsview, Ontario

2
ENSO El Niño Southern Oscillation
El Niño
La Niña
3
Atmos. Response to ENSO is nonlinear
Composite of Z500 and tropical precipitation
during El Niño (A) and La Niña (B) (from
Hoerling et al 1997 J. of Climate)

-

A

-

• La Niña El Niño
• Sign reversed
• Shifted eastward by 30-40(asymmetric)

B
-
4
Question

• If x is the ENSO index, how to derive the atmos.
  response y ƒ(x) ?
• linear regression (or projection)
• y a x

x
x
-
-


-

• Linear method cannot extract asymmetric patterns
  between x and x
• Need a nonlinear method

5
Nonlinear projection via Neural Networks (NN
projection)

• x, the ENSO index
• h, hidden layer
• y, output, the atmos. response

A schematic diagram
Cost function J y y is minimized to
get optimal Wx, bx, Wh and bh (y is the
observation)
6
Data

• ENSO index (x)
• 1st principal component (PC) of the tropical
  Pacific SSTA
• Nov.-Mar.
• 1950-2001,monthly
• SST data from ERSST-v2 (NOAA)
• Linear detrend
• standardized

• Atmos. Fields (y)
• surface air temp. (SAT) and precip.(PRCP)
• From CRU-UEA (UK)
• Monthly,19502001, 1??1?
• Nov.-Mar. North America
• Anomalies (1950-01 Clim)
• Linear detrend
• PRCP standardized
• Condensed by PCA
• 10 SAT PCs (90) retained
• 12 PRCP PCs (60)

7
Bootstrap

• A single NN model may not be stable (or robust)
• Bootstrap randomly select one winters data 52
  times from the 52-yr data (with replacement) ?
  one bootstrap sample
• Repeat 400 times ? train 400 NN models ? average
• of the 400 models as the
• final solution

400 NN models
Give a x ? NN model ? y ? (combined with EOFs) ?
atmosphere anomaly pattern associated with x
8
NN projecton in the SAT PC1-PC2-PC3 space

• Green 3-D
• Blue projected on 2-D PC plane
• C extreme cold state W extreme warm state
• Straight line linear proj.
• Dots data points

9
SAT anomalies

• as ENSO index takes on its
• (a) min.
• (d) max.
• (b) 1/2 min.
• (e) 1/2 max.
• (c) a-2?b
• (f) d-2?e
• Darker color ? above 5 significance

10
PCA on Lin. Nonlin. Parts of NN projection
NL NN LR
Linear regression
27
73
11

• PC1 of Lin. part vs. ENSO index ? a straight
  line
• PC1 of Nonlin. part vs. ENSO index ? a quadratic
  curve
• ?
• A quadratic
• response

12
A polynomial fit
?1 , ?2 are x, x2 normalized, x is the ENSO
index
SAT
13
PRCP anomalies

• as ENSO index takes on its
• (a) min.
• (d) max.
• (b) 1/2 min.
• (e) 1/2 max.
• (c) a-2?b
• (f) d-2?e
• Darker color ? above 5 significance

14
Lin. nonlin. prcp. response to ENSO
LR NL NN
78
22
15
Lin. nonlin. prcp. PC1 vs. ENSO index
16

• Summary and Conclusion
• N. Amer. winter climate responds to ENSO in a
  nonlinear fashion (exhibited by asymmetric SAT
  and PRCP patterns during extreme El Niño and La
  Niña events).
• The nonlinear response can be successfully
  extracted by the nonlinear projection via neural
  networks (NN), while linear method can not.
• NN projection consists of a linear part and a
  nonlinear part. The nonlinear part is mainly a
  quadratic response to the ENSO SSTA, accounting
  for 1/41/3 as much as the variance of the linear
  part.

17

• Thank you !