Roberts - Roberts 1978 JGR
Roberts - Roberts 1978 JGR
Roberts - Roberts 1978 JGR
THOMAS D. ROBERTS
its perfectcutoff characteristic).(On page91 of Mooerset al. wherez is a complexvariable. Sincex• = 0 for k < 0,
[1968]the equationfor the weightingfactorshouldreadf(m)
= •{[1 + cos (•'m/60)] sin (0.7•'m/12)}/O.7•-m/12.) The con- Xo Xl XN-1
Squared Butterworth
Cosine- Lanczos
Ideal
............ Gaussian
10-1
10-2
,.
..
10-3
ß
ß
ß
ß
10-4
10-5
10 3 10- 2 10-1 10
Fig. 1. Power gain functionsfor four low-passfilters with cutoff frequencyat 1/40 hours. The cosine-Lanczos
and
Gaussianfilters each were given 60 weights.
cading second-orderfilters minimizes roundoff error, which the squareof the gain of the originalfilter. Note, however,that
can be a problem with higher-orderfilters. Stearnsalso uses thisfinalgainis no longerthat of a Butterworth filter,though
this method to designButterworth bandpassand high-pass it will approximatea Butterworthpower gain function.
filters.
AN EXAMPLE
Becausethe Butterworth filter is recursive(that is, past
valuesof the outputare usedasinput), thereis an unavoidable Figure 2a containsinformation about the differencein sea
phaseshift of the signal.This can be removedby filteringthe level at Kodiak and at GASS9 in the Gulf of Alaska, 400 km
signal forward and then backward in time through the same away. The hourly data start on April 21, 1976, and are of 3-
filter. In this casethe final power gain functionis, of course, months duration. Only the first 512 hours are used in this
2.00 -
1.50
1.00
0.50
- 0.50
- 1.50
- 2.00 I I I I I I I I t I
50 1O0 150 200 250 300 350 400 450 500
TIME (hours)
2.00
Butterworth & Cosine - Lanczos
Ideal
Gaussian
1.50
1.00
0.50
- 0.50
- 1.00
-1.50
- 2.00
50 100 150 200 250 300 350 400 450 500
TIME (hours)
Fig. 2b. The samedata of Figure 2a, after filtering. The first and last 60 hours of the data have bccn discarded.
ROBERTS AND ROBERTS:BRIEF REPORT 5513
4.00
Squared
Butierworth
Cosine- Lanczos
3.50 Ideal
ß ß ß ß ß ß ß ß, Gaussian
:,
ß ß
3.00
,:
2.50
2.00
I,U
rr 1.50
1.00
0.50
0.50
1.oo
[ 50 10.0 150 200 250
TIME (hours)
300 350 400 450 500
Fig. 3. Low-pass
filterimpulse
responses.
Eachfilterwasgivenasinputa 512-pointsequence
consisting
of all zeros,except
for the 200th point, which had a value of 50.
ideal filter (using a forwara fast Fourier transform (FFT), computertime betweenthe Butterworthfilter and the com-
discardingFourier coefficientscorrespondingto frequencies parisonfilters considered,and the differencebecomesmore
greater than 1/40 hours, and then usingan inverseFFT with pronouncedas more data points are filtered. It is not our
the remainingcoefficients).The numberof weights,m, and the purposeto recommend replacement by the Butterworthfilter
cutoff for the Gaussian filter are chosen to make its cost of other filters now in use;rather, we wish to make note of the
identical to that of the cosine-Lanczosfilter. The power gain cost comparison.
functionsfor thesefilters are shown in Figure 1.
Acknowledgments.The authorswishto expresstheir appreciation
To examinethe transientresponse,an impulseof magnitude to J. Bottero for providingthe computerprogramsfor the cosine-
50 at time 200 has been applied to each of the filters. The Lanczosfilter. This work waspartiallysupportedundercontract03-5-
results are shown in Figure 3. The Butterworth filter has 022-56betweenthe Universityof AlaskaandNOAA, Departmentof
Commerce.
slightlymore overshootthan the cosine-Lanczos filter; other-
wise,the responses are almostidentical.There is no overshoot REFERENCES
for the Gaussianfilter, but the impulseis still broadened.The
ideal filter exhibits an undamped oscillation which can be Brenner,N.M., Three Fortran programsthat perform the Cooley-
eliminatedby putting a shoulderon the transfer function(J. Tukey Fouriertransform,Tech.Note 1967-2,29 pp., LincolnLab.,
Mass. Inst. of Technol., Lexington, 1967.
Holbrook and R. Weisberg,personalcommunication,1978). Cadzow, J. A., Discrete-TimeSystems:An IntroductionWith Inter-
The resultsof filteringthe sealevel data are shownin Figure disciplinaryApplications,440 pp., Prentice-Hall,EnglewoodCliffs,
2b. The Butterworth and the cosine-Lanczos filtered data are N.J., 1973.
almost identical, and the ideal filtered data are similar. As can Kuo,
F.F.,Network
Analysis
andSynthesis,
2rid
ed.,51•pp.,John
Wiley, New York, 1966.
be predicted from Figure 1, the Gaussianfilter with only 60
Mooers,C. N. K., L. M. Bogert,R. L. Smith,and J. G. Patullo,A
weightsretainsan appreciablediurnal tidal component. compilationof observations from mooredcurrent metersand
Table 1 is a costcomparisonfor the filters.Costswill vary, thermographs (and of complementary oceanographic and atmo-
of course,from computer to computer, but the running time sphericdata),vol. 2, Ref.No. 68-5,98 pp., Dep. of Oceanogr.,
Ore.
State Univ., Corvallis, 1968.
comparisonis clear. The cost for the ideal filter can probably
Stearns,S. D., Digital SignalAnalysis,280 pp., HaydenBookCom-
be somewhatreducedby usinga more efficientFFT program pany, RochellePark, N.J., 1975.
(we usedBrenner's[1967] program FOURG). In view of the Tarbell, S. A., and A. W. Whitlach,A compilationof mooredcurrent
transient responsefor this filter, however,it was not explored data and associatedoceanographic observations,1970 Measure-
further. ments,vol. 13,Rep. WH01-77-18,189pp., WoodsHole Oceanogr.
Inst., Woods Hole, Mass., 1977.
CONCLUDING REMARKS