Computers & Geosciences 28 (2002) 457–465

Geophysical data analysis using Python$

Jon Sa! enza,*, Juan Zubillagab, Jesus
! Ferna! ndezb
a
Depto. de F!ısica Aplicada II, Universidad del Pa!ıs Vasco, Apdo. 644, 48080-Bilbao, Spain
b
Depto. de F!ısica de la Materia Condensada, Universidad del Pa!ıs Vasco, Apdo. 644, 48080-Bilbao, Spain
Received 3 November 2000; received in revised form 12 July 2001; accepted 13 July 2001

Abstract

A set of routines designed for geophysical data analysis that make extensive use of the numerical extensions to the
computer language Python are presented. The routines perform some typical tasks during multivariate analysis of
geophysical fields, such as principal component analysis and related tasks (truncation rules by means of analytical and
Monte Carlo techniques). Other functions perform singular value decomposition of covariance matrices and canonical
correlation analysis for coupled variability of geophysical fields. Other parts of the package allow access to a library of
statistical distribution functions, multivariate digital filters, time-handling routines, kernel-based probability density
function estimation and differential operators over the sphere for gridded data sets. As they rely on the numerical
extensions to the Python language, they are fast for numerical analysis. The programs make the analysis of geophysical
data sets both easier and faster. r 2002 Elsevier Science Ltd. All rights reserved.

Keywords: Principal component analysis; Digital filter; Probability density function; Differential operator; Python

1. Introduction written for data analysis by scientists will not be used

often. So, saving some CPU cycles when running them is
In recent years, there has emerged compelling not interesting if one considers the development time
evidence that CPU cycles are not the only useful required.
quantity to measure the adequacy and effectiveness of On the other hand, some programming languages are
information systems and technologies to help scientists extremely powerful and they are the mainstream of
in their data analysis needs. That is the reason why some computer programming, but only for professional
tools such as IDL or Matlab have become so popular, programmers. It is difficult to write programs using
despite the fact that they need more CPU cycles than C++, for instance, because of conceptual and syntac-
equivalent programs written in C, FORTRAN or even tical problems. Students (undergraduate or graduate) of
assembler to solve the same numerical problem. Con- Earth Sciences would do better to study other topics
sidering the whole development cycle of some programs, related to their discipline rather than learning the
which includes programmers’ time, it is doubtful that whole Standard Template Library to perform some
low-level languages are the best tools available, because simple computations. Another advantage of high-
it takes too much time to develop the programs. It is level languages is that coding is straightforward, because
difficult and error-prone to write some programs using the programmer has only to consider the algorithm
low-level languages. Finally, many of the programs itself, and may avoid being distracted by syntactic or
data-hiding concerns. Next, if the algorithm is to be
optimized, only those parts that are actually the
$
Code available from server at http://www.iamg.org/ most critical ones should be coded using low-level
CGEditor/index.htm. languages. In this context, it is interesting that the
*Corresponding author. high-level code can be easily extended and designed with
E-mail address: jsaenz@wm.lc.ehu.es (J. S!aenz). this goal in mind from the start.

0098-3004/02/$ - see front matter r 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 9 8 - 3 0 0 4 ( 0 1 ) 0 0 0 8 6 - 3
458 !
J. Saenz et al. / Computers & Geosciences 28 (2002) 457–465

In this work, we are presenting the routine PyClimate1 complex models on a UNIX workstation and analyze
as a Python package that performs some of the most the results with a visualization program on a laptop
typical tasks during data analysis in the field of climate running MS Windows or other operating systems.
analysis, although they can be used for any general- To overcome the problems associated with working
purpose geophysical data analysis work. The package in different computing environments, which often
has been developed by the authors making extensive use support the binary representation of numbers in
of the numerical extension to Python. Python is a different ways, there exist some data formats which
modern, free of charge, and multiplatform-interpreted provide transparent access to binary data. netCDF
object-oriented high-level language (Dubois et al., 1996; (network Common Data Format)3 is one of those
Lutz, 1996; Lutz and Ascher, 1999). It was developed by systems which allows the use of the same binary
Guido van Rossum while working for the Stichting data files in ‘‘little-end’’ or ‘‘big-end’’ machines (Cohen,
Mathematisch Centrum in Amsterdam and first released 1981) without the user having to worry about how
in 1991. The programs are easy to read, write and debug, the data was created or how it is being read, and without
which shortens the development cycle. The code is easily the additional computational and storage burden
extensible and embeddable, providing access to external (and loss of precision) associated to the conversion to
devices, libraries and system-calls. The numerical exten- ASCII. There are some conventions built upon the
sions rely on an efficient C implementation of the basic bare netCDF to be able to communicate between
array operations. Thus, even if Python is interpreted, different pieces of software. One of the most popular
numerical Python is compiled and performs fast for conventions in oceanographic and atmospheric data sets
numerical tasks, with support for several data types is the so-called Cooperative Ocean/Atmosphere
(Int8, Int16, Float32, Float64, Complex32, Complex64). Research Data Service (COARDS) Conventions.4
It has access to several libraries for scientific computa- These conventions are supported by several universities
tion such as Fast Fourier Transform, special functions, and agencies and are interpreted by some popular
random number generation or linear algebra. Detailed visualization software packages (GrADS, Ferret).
information about Python can be found on the Python These conventions enforce a certain order to the
Homepage.2 It is a no-cost alternative to similar dimensions of a variable (time, level, latitude, longitude,
commercial products (MatLab, IDL, Maple, etc.) for etc.) and some attributes, so that visualization programs
those people who cannot afford the licensing fees. Its are able to properly interpret the compliant netCDF
simplicity and no-cost facility make it a viable alter- files. PyClimate includes a function to replicate the
native for its use as a classroom language. Another needed parts (dimensions, variables, attributes) of an
advantage over MatLab and other high-level software is existing COARDS netCDF file so that the output file
that Python is a general-purpose programming language from a computation can be created using the same
and allows much more interaction with the operating structure with a single call. An example of this is
system. provided in Fig. 1.
This paper will show a detailed discussion of the One of the problems facing climatologists or oceano-
mathematical concepts and data analysis tasks included graphers is the need to position adequately their
in our package, PyClimate. A detailed description of observations in time. This is especially relevant in the
the functions and several examples are provided in the frequent situation of monthly averages, because there is
documentation included with the distribution. However, no simple definition of a month. It depends on the year
to give the reader a deeper perspective on the kind and the month itself (28–31 days). The approach in this
of programs that can be prepared using these tools, package is to use Julian Days and define one month as
some short examples will also be shown here. Prerequi- one-twelfth of one year, considering one tropical year
sites, compilation for several UNIX flavours, installa- equal to 365.242198781 days. There exists a second class
tion and test procedures are also described in the (JDTimeHandler) whose constructor parses the attri-
documentation accompanying the distribution, freely bute units of the time variable in a COARDS file. Next,
available in Internet. making use of an offset and scaling process, the
instances are able to recover the fields of a date structure
from the time variable expressed as a double. The
2. COARDS-compliant netCDF files and time-handling
routines

Nowadays, it is commonplace to work using different 3

University Corporation for Atmospheric ResearchFUni-
computing platforms. For example, a user can run data NetCDF. http://www.unidata.ucar.edu/netCDF.
4
University Corporation for Atmospheric ResearchF
1
PyClimate Official Site, 2001. http://www.pyclimate.org. NetCDF Conventions. http://www.unidata.ucar.edu/netCDF/
2
Python Language Website. http://www.python.org. Conventions.html.
 !
J. Saenz et al. / Computers & Geosciences 28 (2002) 457–465 459

Definition of a minimal netCDF COARDS file of daily geopotential height:

netcdf hgtdaily {
dimensions:
time=UNLIMITED;
Z=5;
lat=74;
lon=144;
variables:
double time(time);
time:units=”hours since 1950-01-01”;
float Z(Z);
Z:units=”Pa”;
float lat(lat);
lat:units=”degrees_north”;
float lon(lon);
lon:units=”degrees_east”;
short hgt(time,Z,lat,lon);
}

Create an output netCDF file structured as the daily data and print the dates:
from pyclimate.ncstruct import nccopystruct
from pyclimate.JDTimeHandler import JDTimeHandler
from Scientific.IO.NetCDF import NetCDFFile

# Open the input NetCDF file using K. Hinsen’s netCDF access

inc=NetCDFFile(“hgtdaily.nc”)
# Retrieve the time variable object
itime=inc.variables[“time”]

dims=(“time”,”Z”,”lat”,”lon”) # Dimensions to be copied

vars=(“time”,”Z”,”lat”,”lon”) # Copy their structures
vcopy=(“Z”,”lat”,”lon”) # Copy their contents
# onc is a netCDF file ready to be written to
onc=nccopystruct(“hgtmonthly.nc”,inc,dims,vars,vcopy)

# Print the year/month/day associated to each record

jdth=JDTimeHandler(itime.units)
# Iterate through the records of the input file
records=itime.shape[0]
for irec in xrange(records):
iy,mi,id=jdth.getdatefields(itime[irec],3)
print “%4.4d-%2.2d-%2.2d”%(iy,mi,id)

Fig. 1. Use of netCDF files and time extensions of PyClimate.

current implementation does not consider the slow drift 3. Cumulative distribution functions
in the number of days per tropical year, so, it is of little
use for geological time scales involving several millennia. The possibility of extending the Python interpreter
The package should be used with caution for dates with external C libraries is illustrated with an example of
before the introduction of the Gregorian calendar, the way we are providing access to the DCDFLIB.C 1.1
because it simply extrapolates this calendar without library from Python in the package PyClimate.
taking into account the date of its introduction, which DCDFLIB.C library is a free library designed for the
changed from country to country. An example of the use direct and inverse computation of parameters corre-
of this class to read time values in netCDF files is shown sponding to discrete cumulative density functions, by
in Fig. 1. Barry W. Brown, James Lovato and Katty Russell,
460 !
J. Saenz et al. / Computers & Geosciences 28 (2002) 457–465

Original C prototyope in DCDFLIB.C:

void cdfchi(int *which, double *p, double *q, double *x, double *df,
int *status, double *bound);

C structures and functions of pyclimate designed to allow access to DCDFLIB from Python:
typedef struct {
int which;
double p;
double q;
double x;
double df;
int status;
double bound;
} CDFChi;
extern int pycdfchi( CDFChi *sptr);

Use from the Python interpreter:

import pyclimate.pydcdflib
pycdf=pyclimate.pydcdflib
chi2=pycdf.CDFChi() # Create an instance of the Chi**2 object
chi2.which=2 # Kind of conversion, from p,q and dof, get x
chi2.p=0.95 # Assign p (q is automatically set to 1-p)
chi2.df=10 # Degrees of freedom
# This intermediate function calls the original one in DCDFLIB.C
pycdf.pycdfchi(chi2)
# This prints: 18.307038 0 (0 means no error in DCDFLIB.C)
print chi2.x,chi2.status

Fig. 2. Access to DCDFLIB.C from Python by means of layer of structures and function calls.

Department of Biomathematics, the University of analysis is a popular technique in oceanographic or

Texas, M.D. Anderson Cancer Center.5 This library atmospheric data analysis (Preisendorfer, 1988; Rich-
defines functions for several statistical distribution man, 1986; von Storch and Zwiers, 1999). Given a scalar
functions such as b, binomial, negative binomial, w2 ; field Xðr; tÞ as a data matrix X ¼ ½xi;j ; with subscript
non-central w2 ; normal, Poisson, and others. Starting i ¼ 1; y; n the number of temporal samples and j ¼
from the original C prototype of the DCDFLIB.C 1; y; m the observing sites, PyClimate provides func-
library, some data types and functions which provide tions to calculate the eigenvalues, eigenvectors and the
access from Python to the variables used by the library principal components or temporal expansion coefficients
are defined in the interface that has been developed for of the field. The eigenvalues ll are the variances
PyClimate. An example for the w2 distribution function accounted for by each mode of the decomposition, the
is shown in Fig. 2. eigenvectors ul are orthogonal ðuk  um pdkm Þ and the
temporal expansion coefficients zk ðtÞ can bePscaled to
represent the variance of each mode, that is, t zkt zlt ¼
4. EOF, SVD and CCA analysis ll dkl or can also be set to unit variance (in such a
situation, the eigenvectors are accordingly scaled). The
The analysis of the joint spatial and temporal input dataset is centered (sample mean removed) inside
variabilities of geophysical fields in terms of empirical the subroutine. There are some functions that implement
orthogonal functions (EOF) or principal component truncation rules like the Bartlett test on the equality of
the last p eigenvalues (Jackson, 1991). The use of the
5
M.C. Anderson Cancer CenterFStatistical Software. North et al. (1982) rule on the sampling degeneracy of
http://odin.mdacc.tmc.edu/anonftp/page3.html. the eigenvalues is also available. Finally, the last
 !
J. Saenz et al. / Computers & Geosciences 28 (2002) 457–465 461

possibility is the use of Monte Carlo techniques based temperature (SST) field and the atmospheric geopoten-
on temporal subsampling of the input dataset, assessing tial height field (Peng and Fyfe, 1996), SSTs and
the stability of the eigenvectors through the congruence tropospheric vertical wind shear (Shapiro and Gold-
coefficient (Cheng et al., 1994; Richman and Lamb, enberg, 1998) or ozone and equatorial zonal winds
1985). For geophysical data sets, the number of (Randel and Wu, 1996). PyClimate includes some
temporal samples is usually lower than the number of functions to compute the covariance and squared
spatial sites. Under these conditions, the covariance covariance fractions, the homogeneous and heteroge-
matrix is singular and the eigenvalue problem cannot be neous correlation maps and the singular vectors. It is
solved by means of standard linear algebra routines like also able to perform a quantitative analysis on the
LAPACK’s dsyev. To overcome this limitation, PyCli- stability of those vectors by means of a Monte Carlo
mate uses the singular value decomposition (SVD) of the analysis based on temporal subsampling.
data matrix to achieve the EOF decomposition (von The canonical correlation analysis (CCA) is another
Storch, 1995; Wunsch, 1997), which makes the problem widely used technique for the analysis of linearly
solvable under all circumstances (Golub and van Loan, coupled fields (von Storch, 1995). Unlike the SVD that
1996). This approach makes essential a judicious use of maximises the covariance of the fields under considera-
the truncation rules previously mentioned. A simple tion, the CCA maximises the correlation of the fields.
example of how the EOF analysis is performed on a PyClimate also provides a module to carry out this
dataset is shown in Fig. 3. analysis.
One of the simplest algorithms available for the
analysis of the linearly coupled variability of multi-
variate datasets is performed by means of the singular 5. Kernel-based probability density function estimation
value decomposition of the covariance matrix of both
the fields (Bretherton et al., 1992). Despite some debate Histograms are simple estimators of probability
about the method (Cherry, 1996; Hu, 1997; Newman density functions (PDF) in univariate and multivariate
and Sardeshmukh, 1995), it is widely applied. Some statistical analysis. However, they have some draw-
applications include, for instance, the sea surface backs, such as their lack of derivability or their poor

# Example on the use of svdeofs.py

import Numeric
import pyclimate.svdeofs
from pyclimate.readdat import readcols
pyeofs=pyclimate.svdeofs

# The chemical.dat example at Jackson, 1991. Read the data:

chdata=Numeric.array(readcols("chemical.dat",[1,2]))
# Get the eigenvalues, eigenvectors and principal components
z,lambdas,eofs= pyeofs.svdeofs(chdata)
# Get the variance fraction
varfrac= pyeofs.getvariancefraction(lambdas)
# Sampling errors according to North et al. (1982)
north= pyeofs.northtest(lambdas,len(chdata))
# Perform the Bartlett test
chis,chiprobs= pyeofs.bartletttest(lambdas,len(chdata))
# The next sentences show the way the eigenvectors are arranged
print "# Eigenvectors:"
for ivect in xrange(len(lambdas)):
print "# EOF %1d:"%(ivect,),
for icomponent in xrange(len(lambdas)):
print "%10.4f"%(eofs[icomponent,ivect],),
print

Fig. 3. Use of code to compute EOFs by means of SVD decomposition of data matrix.
462 !
J. Saenz et al. / Computers & Geosciences 28 (2002) 457–465

stability upon the selection of parameters like origin,

number of bins and bin-width (Silverman, 1986; von
Storch and Zwiers, 1999). Thus, several recent works use
kernel-based PDF estimation for the analysis of
observational data or modelling results (Corti et al.,
1999; Kimoto and Ghil, 1993; Robertson et al., 2000).
PyClimate currently provides different functions to
perform kernel-based univariate and multivariate
density estimation. All of them share the same kernel-
based estimation for univariate data,
 
1 Xn
x  Xi
f#ðxÞ ¼ K ; ð1Þ
nh i¼1 h

where KðuÞ are the different kernels, x the abscissa of the

Fig. 4. PDFs of Cold Tongue (CTI) and Arctic Oscillation
indices during last century. PDF, h the bandwidth of the estimation and Xi the
experimental points, i ¼ 1; y; n: Currently, there exist

import Numeric
import LinearAlgebra
import pyclimate.KPDF
from pyclimate.readdat import readdat
from Scientific.IO.NetCDF import NetCDFFile
N=Numeric
LA= LinearAlgebra
pyKPDF=pyclimate.KPDF

# Read the CTI/AO dataset, covariance matrix and sqrt(det(S))

ctiao=N.array(readdat("ctiao.dat")) # Read
residual=ctiao-N.add.reduce(ctiao)/len(ctiao) # Remove mean
S=N.matrixmultiply(N.transpose(residual),residual)/len(residual)
Sm1=LA.inverse(S)
sqrtdetS=N.sqrt(LA.determinant(S))
# Create the netCDF file to hold the PDF
onc=NetCDFFile("ctiaopdf.nc","w")
Xs=N.arange(-3,3.01,0.05) # X-Y grid
Ys=N.arange(-3,3.01,0.05)
# Define the netCDF dimensions
onc.createDimension("dimx",len(Xs))
onc.createDimension("dimy",len(Ys))
# Define the X and Y variables of the netCDF file
onc.createVariable("X",N.Float64,("dimx",))[:]=Xs[:]
onc.createVariable("Y",N.Float64,("dimy",))[:]=Ys[:]
# This variable will store the PDF. Declare it
PDF=onc.createVariable("PDF",N.Float64,("dimx","dimy"))
# Start computing the PDF
grid=pyKPDF.MPDF2DGrid2Array(Xs,Ys)
pdfvect= pyKPDF.MPDFEpanechnikov(ctiao,grid,1.,Sm1,sqrtdetS)
pdfvect.shape=PDF.shape
PDF[:,:]=pdfvect[:,:]
onc.close()

Fig. 5. Example of use of two-dimensional PDF estimator.

 !
J. Saenz et al. / Computers & Geosciences 28 (2002) 457–465 463

functions, which accomplish the estimation in Eq. (1) 1999). The analysis of ENSO is often performed after
using Epanechnikov, Biweight or Triangular kernels. filtering data in adequate frequency intervals (Tourre
For multivariate data, simple and Fukunaga (Eq. (2)) and White, 1997). Thus, digital filters are ubiquitous
estimators have been implemented, in geophysical data analysis to remove from a broad-
band signal those frequency components that are
ðdet SÞ1=2 X
n
f#ðxÞ ¼ Kðh2 ðx  Xi ÞT S1 ðx  Xi ÞÞ; ð2Þ irrelevant to the problem P at hand. The filters are
nhd i¼1 usually linear, that is, Yt ¼ nk¼n ak Xtþk : Making use
of the array operations of Numerical Python, it is easy
where x is the d-dimensional point where the PDF is to to code this kind of filter in an efficient manner. To
be evaluated, Xi ; i ¼ 1; y; n are the observed points achieve a good performance, the implementation in
in the d-dimensional space. The data are previously PyClimate iterates only once over the whole dataset.
scaled by means of the inverse of the covariance matrix First, the filter coefficients ak are computed. Next, the
in the Fukunaga estimator, so, h is a one-dimensional linear combination of records is calculated by means of
bandwidth common to all axis and S is the sample matrix operations. There exist several books devoted to
covariance matrix. For simple multivariate estimators, the design of filters. The main goal is to achieve sharp
instead of the covariance matrix, the identity matrix I edges in the transference function while reducing the
is used. Currently, the estimator works only for d ¼ 2; 3 Gibbs’ oscillations. Currently, PyClimate supports
and Epanechnikov and multivariate Gaussian kernels. Kolmogorov–Zurbenko (Rao et al., 1997; Eskridge
Fig. 4 shows a sample PDF computed using h ¼ 1 et al., 1997) and Lanczos (Duchon, 1979) filters. Other
from the observed monthly Cold Tongue Index (CTI) filters can be added by subclassing the LinearFilter class
and Arctic Oscillation Index (AO), which measure the rewriting the constructors, which simply define the
*
El Nino-Southern Oscillation (ENSO) and the strength coefficients ak :
of the annular mode of the extratropical atmospheric
circulation in the Northern Hemisphere (Thompson
and Wallace, 1998). Fig. 5 shows the program used
7. Differential operators on the sphere
to compute the PDF. The module is completely written
using C to achieve better performance. There are
There exist several libraries with functions to compute
some auxiliary functions to generate the grids as linear
differential operators on spherical coordinates, like
arrays for multidimensional cases. There is no need to
SPHEREPACK,6 but they are oriented to the develop-
use complicated combinations of other Python and
ment of computer-intensive models using low-level
Numerical Python functions such as map, replicate
programming languages. However, in geophysics, it is
and concatenate to obtain those structures, whilst
a typical task to use differential operators in spherical
improving the overall performance of the PDF esti-
coordinates for the analysis of gridded data sets. The
mation.
gradient of geopotential height in the atmosphere or
ocean is needed to compute the geostrophic wind or
currents. The divergence of the troposphere-integrated
6. Multivariate digital filters moisture transport gives an estimation of the evapora-
tion less precipitation over the surface, if the time
When analysing geophysical data sets, there are often evolution of the vertically integrated moisture content is
different scales of motion involved, and they reflect disregarded. Finally, the curl of the surface wind stress
different physical processes. When analysing oceano- over the ocean divided by the density and the Coriolis
graphic data, it is important to distinguish between parameter allows the computation of the vertical
internal gravity waves and Rossby waves (Gill, 1982) in velocity at the base of the Ekman layer, to name a few
the analysis of observed data or to filter spurious aliased examples of popular uses of these differential operators.
energy from model results (Jayne and Tokmakian, PyClimate includes functions to compute the horizontal
1997). For extratropical atmospheric motions, the component of the gradient,
variability in the 2–10 days is due to baroclinic  
instability of the atmospheric flow. Conversely, the so- ~h F ¼ 1 qF 1 qF
r ; ; ð3Þ
called low-frequency variability in the monthly time- a cos f ql a qf
scale, is usually attributed to tropical diabatic forcing of the divergence of a horizontal vector field,
the atmosphere (Lau, 1997) or non-linear interaction of  
waves of different scales (Handorf et al., 1999; Hansen ~  vh ¼ 1
r
qu qðv cos fÞ
þ ð4Þ
et al., 1997). For the analysis of tropical convection, it is a cos f ql qf
a common practice to remove the joint effects of the
Madden–Julian Oscillation and the mixed Rossby- 6
Spherepack: A Model Development Facility. 1998. http://
gravity waves in the Tropics (Matthews and Kiladis, www.scd.ucar.edu/css/software/spherepack/.
464 !
J. Saenz et al. / Computers & Geosciences 28 (2002) 457–465

# Multiply each row of the v component by cosine(lat)

dummy=array(v)*self.clats[:,NewAxis]
# This will hold the resulting array
cdifv=zeros(v.shape,Float64)
# Meridional directions
cdifv[...,1:-1,:]=dummy[...,2:,:]-dummy[...,:-2,:]
cdifv[...,0,:]=2*(dummy[...,1,:]-dummy[...,0,:])
cdifv[...,-1,:]=2*(dummy[...,-1,:]-dummy[...,-2,:])
# Divide by the increments in latitude
cdifv=cdifv/2./self.dlat
# Now, divide by R*cos(lat)
cdifv=cdifv/R/self.clats[:,NewAxis]

Fig. 6. Array-based computation of qðv cos fÞ=qf by means of second-order centered differences.

and the vertical component of the rotational of a Acknowledgements

horizontal vector field,
  This work was made possible, thanks to grants
~
vÞz ¼ 1
ðr
qv

qðu cos fÞ
; ð5Þ CLI98-0236 (Plan Nacional de I+DCICYT). Jesus !
a cos f ql qf
Fern!andez is granted by the ‘‘Departamento de Educaci-
where a is the radius of the earth, l longitude and f ! Gobierno Vasco’’. The authors are grateful for the
on,
latitude. The numerical derivatives are based on second- comments and suggestions made by two anonymous
order centered finite-difference schemes in the central reviewers.
part of the domain and second-order backward/forward
schemes at the boundaries. These operations are only
supported for regular two-dimensional latitude/long-
itude grids ordered as required by COARDS
conventions; fk;l ¼ f ðfk ; ll Þ: As an example, Fig. 6 References
shows the implementation of the second term on the
right-hand side of Eq. (4). Bretherton, C.S., Smith, C., Wallace, J.M., 1992. An inter-
comparison of methods for finding coupled patterns in
climate data. Journal of Climate 5 (6), 541–560.
8. Concluding remarks and prospects Cheng, X., Nitsche, G., Wallace, J.M., 1994. Robustness of
low-frequency circulation patterns derived from EOF
A set of routines written by the authors and designed and rotated EOF analyses. Journal of Climate 8 (6),
for data analysis in the atmospheric or oceanographic 1709–1713.
Cherry, S., 1996. Singular value decomposition and canonical
sciences has been presented. Although they have been
correlation analysis. Journal of Climate 9 (9), 2003–2009.
designed with these fields in mind, some of the problems Cohen, D., 1981. The holy war and a plea for peace. Computer
considered are of wider interest and they can be used for 14 (10), 48–54.
general data analysis problems in several branches of Corti, S., Molteni, F., Palmer, T.N., 1999. Signature of recent
geophysics. The routines are written in C or Python and climate change in frequencies of natural atmospheric
can be accessed from the Python interpreter, which is a circulation regimes. Nature 398 (6730), 799–802.
free language with a numerical extension that makes it Dubois, P.F., Hinsen, K., Hugunin, J., 1996. Numerical
suitable for scientific data analysis. Python. Computers in Physics 10 (3), 262–267.
The routines are publicly accessible under the GPL Duchon, C.E., 1979. Lanczos filtering in one and two
license and are distributed freely on the Internet. dimensions. Journal of Applied Meteorology 18 (8), 1016–
1022.
PyClimate and other existing packages, which are
Eskridge, R.E., Ku, J.Y., Rao, S.T., Porter, P.S., Zurbenko,
written or make extensive use of Python (CDAT7 and I.G., 1997. Separating different scales of motion in time
mtaCDF8), make this programming language a valuable series of meteorological variables. Bulletin of the American
tool for the analysis of atmospheric and oceanic data. Meteorological Society 78 (7), 1473–1483.
Gill, A.E., 1982. Atmosphere-Ocean Dynamics, 1st edn.
7
Climate Data Analysis Tool. http://www-pcmdi.llnl.gov/ Academic Press, San Diego, CA, 662pp.
software/cdate. Golub, G.H., van Loan, C.F., 1996. Matrix Computations, 3rd
8
mtaCDF Documentation. http://www.ifm.uni-kiel.de/to/ edn. The John Hopkins University Press, Baltimore, MD,
sfb460/b3/Products/mtaCDF.html. 694pp.
 !
J. Saenz et al. / Computers & Geosciences 28 (2002) 457–465 465

Handorf, D., Pethoukhov, V.K., Dethloff, K., Eliseev, A.V., Preisendorfer, R.W., 1988. Principal Component Analysis in
Weisheimer, A., Mokhov, I.I., 1999. Decadal climate Meteorology and Oceanography, 1st edn. Elsevier, Am-
variability in a coupled model of moderate complexity. sterdam, 425pp.
Journal of Geophysical Research 104 (D22), 27253–27275. Randel, W.J., Wu, F., 1996. Isolation of the ozone QBO in
Hansen, J., Sato, M., Ruedy, R., Lacis, A., Asamoah, A., SAGE II data by singular-value decomposition. Journal of
Beckford, K., Borenstein, S., Brown, E., Cairns, B., the Atmospheric Sciences 53 (17), 2546–2559.
Carlson, et al., 1997. Forcing and chaos in interannual to Rao, S.T., Zurbenko, I.G., Neagu, R., Porter, P.S., Ku, J.Y.,
decadal climate change. Journal of Geophysical Research Henry, R.F., 1997. Space and time scales in ambient ozone
102(D22), 25679–25720. data. Bulletin of the American Meteorological Society 78
Hu, Q., 1997. On the uniqueness of the singular value (10), 2153–2166.
decomposition in meteorological applications. Journal of Richman, M.B., 1986. Rotation of principal components.
Climate 10 (7), 1762–1766. Journal of Climatology 6 (2), 293–335.
Jackson, J.E., 1991. A User’s Guide to Principal Components, Richman, M.B., Lamb, P.J., 1985. Climatic pattern analysis of
1st edn. Wiley, Chichester, UK, 569pp. three- and seven-day summer rainfall in the central United
Jayne, S.R., Tokmakian, R., 1997. Forcing and sampling of States: some methodological considerations and a regiona-
ocean general circulation models: impact of high-frequency lization. Journal of Climate and Applied Meteorology 24
motions. Journal of Physical Oceanography 27 (6), (12), 1325–1343.
1173–1179. Robertson, A.W., Mechoso, C.R., Kim, Y.-J., 2000. The
Kimoto, M., Ghil, M., 1993. Multiple flow regimes in the influence of Atlantic sea surface temperature anomalies on
Northern Hemisphere winter. Part I: methodology and the North Atlantic Oscillation. Journal of Climate 13 (1),
hemispheric regimes. Journal of the Atmospheric Sciences 122–138.
50 (16), 2625–2643. Shapiro, L.J., Goldenberg, S.B., 1998. Atlantic sea surface
Lau, N.-C-, 1997. Interactions between global SST anomalies temperatures and tropical cyclone formation. Journal of
and the midlatitude atmospheric circulation. Bulletin of the Climate 11 (4), 578–590.
American Meteorological Society 78 (1), 21–33. Silverman, B.W., 1986. Density Estimation for Statistics and
Lutz, M., 1996. Programming Python, 1st edn. O’Reilly, Data Analysis, 1st edn. Chapman and Hall, London, 175pp.
Cambridge, MA, 904pp. Thompson, D.J.W., Wallace, J.M., 1998. The Arctic oscillation
Lutz, M., Ascher, D., 1999. Learning Python, 1st edn. O’Reilly, signature in the wintertime geopotential height and tem-
Cambridge, MA, 384pp. perature fields. Geophysical Research Letters 25 (9), 1297–
Matthews, A.J., Kiladis, G.N., 1999. The tropical–extratropical 1300.
interaction between high-frequency transients and the Tourre, Y.M., White, W.B., 1997. Evolution of the ENSO
Madden–Julian oscillation. Monthly Weather Review 127 signal over the Indo-Pacific domain. Journal of Physical
(5), 661–677. Oceanography 27 (5), 683–696.
Newman, M., Sardeshmukh, P.D., 1995. A caveat concerning von Storch, H., 1995. Spatial patterns: EOFS and CCA. In: von
singular value decomposition. Journal of Climate 8 (2), 352– Storch, H., Navarra, A. (Eds.), Analysis of Climate
360. Variability. Applications of Statistical Techniques. Springer,
North, G.R., Bell, T.L., Cahalan, R.F., Moeng, F.J., 1982. Berlin, pp. 227–258.
Sampling errors in the estimation of empirical ortho- von Storch, H., Zwiers, F.W., 1999, Statistical Analysis in
gonal functions. Monthly Weather Review 110 (7), Climate Research, 1st edn. Cambridge University Press,
699–706. Cambridge, 484pp.
Peng, S., Fyfe, J., 1996. The coupled patterns between sea Wunsch, C., 1997. The vertical partition of oceanic horizontal
level pressure and sea surface temperature in the mid- kinetic energy. Journal of Physical Oceanography 27 (8),
latitude North Atlantic. Journal of Climate 9 (8), 1770–1794.
1824–1839.