Re: Parameter of function--filwgts_lanczos

Also, for climate research, I recommend that you look at filter
response curves in the *time domain* as well as the frequency domain.
Climatologists may be interested in the cutoff quality in terms of
adjacent time periods. In the low frequency cutoff area, there may be
some rather poor performance that is hidden in the frequency domain
curve!

Currently the NCL filter applications page does not have an example of
a time domain response curve. Please use the following general
purpose evaluation program to see the time domain curve! Look at the
third plot in the output series.

http://www.esrl.noaa.gov/psd/people/dave.allured/data/ncl/demo/filter/plot-filter.lanczos.ncl

For the 421-weight filter, 2 to 7 years, notice how much out of band
signal gets through at the 8 year (96 month) cycle. Not so good, I
think, so you will need even more weights if you want to suppress the
8 year cycle!

--Dave

On Fri, Dec 14, 2012 at 4:29 PM, Dave Allured - NOAA Affiliate
<dave.allured@noaa.gov> wrote:
> Li Yang,
>
> 1. Number of weights.
>
> My informal rule is to always use a number of weights that is at least
> 5 times 1/fca, the reciprocal of the lower cutoff frequency in a
> digital filter, at least for climate analysis. The whole reason for
> using a digital filter is to get sharp cutoff. If there are not
> enough weights, then the filter will have a slow or mushy cutoff, and
> the signal that you are looking for will have significant
> contamination from out-of-band frequencies.
>
> You want a band pass filter for 2 to 7 years, with monthly data. With
> a band pass filter, the lower cutoff frequency is the one associated
> with the higher time limit on the band specification (7 years). Using
> the nomenclature on the filter example page that you referenced:
>
> dt = 1 month per time step
> t1 = 7 years = 84 months (low frequency cutoff, expressed in time domain)
> fca = dt/t1 = 1./84. = 0.0119 time steps (low frequency cutoff)
> nwt = 5 * 1/fca = 5/0.0119 = (420 rounded up) = 421 weights
>
> 2. End point problem.
>
> The only solution is to get additional data points outside of your
> study time range. With 421 weights and monthly data, you need about
> 18 additional years (210 time steps) before your study period, and 18
> additional years after your study period. For example, to get
> complete filter output for 1970 to 1990, you need valid time series
> for 1952 through 2008.
>
> I do not have experience with other filter types. Maybe someone else
> can answer your other question comparing filter types.
>
> Caveats:
>
> 1. When designing a digital filter, always examine carefully the
> filter response curve. See examples on the referenced web page.
>
> 2. Here are caveats and reference information recently posted by Dennis Shea:
> http://www.ncl.ucar.edu/Support/talk_archives/2012/2294.html
>
> 3. Please read other NCL discussions on this topic, there is more
> good advice. Enter "filter weights" into the NCL search box. HTH.
>
> --Dave
> * Reply only to list, please *
>
> On Fri, Dec 14, 2012 at 10:16 AM, Li Yang <liyang0711@gmail.com> wrote:
>> Hi,
>>
>> I have two question to ask you.
>> First is,
>> I want to use Band pass filters for 2-7 years in the monthly data. But I do
>> not know how to select the parameter of this function--filwgts_lanczos.
>> How to select the best parameter (e.g. nwt , how to select it?).
>> And I also want to solve the End point problems(that is End point is valid
>> point),when I use the function--filwgts_lanczos to filter . How do I do.
>> Second is,
>> In this website http://www.ncl.ucar.edu/Applications/filter.shtml,
>> the filwgts_lanczos and FFT all use to Band pass filters.But I do not kow
>> which is best .And if I use Band pass filters for 2-7 years,the
>> filwgts_lanczos or
>> FFT?
>>
>>
>> Best Wishes
_______________________________________________
ncl-talk mailing list
List instructions, subscriber options, unsubscribe:
http://mailman.ucar.edu/mailman/listinfo/ncl-talk
Received on Fri Dec 14 17:53:33 2012