Re: [ferret_users] LSL_Lowpass filter

["William S. Kessler" <william.s.kessler@xxxxxxxx>]

I rarely use LSL_lowpass, but perhaps you should try a simpler filter first, at least to gain insight.

For signal processing where the frequencies sought are precisely known, cutoff filters are necessary, but for noisy geophysical data there probably is little difference in the type of lowpass filter used. I presume you are not looking for a signal with period exactly 7 years, but for some more broad-band process like ENSO. In that case a filter targetted to precise frequencies would be inappropriate and maybe misleading. 

An advantage of simpler filters is that the results are easier to understand, as you are seeing.

I would try the triangle filter (@SPZ), which has nice properties: no ringing, and the frequency response is easy to understand and plot. (See messages from me about this by searching SPZ in the user group archives; e.g. on 6 Jun 1997 and 20 Feb 2011).

If nothing else, you can use the results of the triangle filter to compare tests with more precise filters. Overlay time series of your tests with those from @SPZ. Are the precise filters better or just more involved to describe when you write your paper?

But to answer your question about data loss at the edges: All of these filters act on a finite length of points and require existence of that length to make the computation. At the ends (or at blank values in the middle), the filter length extends past the edge of the data and the computation can’t be done, so the result loses those points. Consider the simplest possible filter, the running mean or boxcar (@SBX). If you want to take a 3-point running mean, what should an algorithm do at, say, the rightmost endpoint? There is a value at the point, and one to its left, but none to its right. A 3-point mean is not defined. But, in the same SPZ search I mentioned above, see a suggestion from Ryo about solutions to this (2 Sep 2005).

Billy K

> On Apr 7, 2016, at 5:49 AM, saurabh rathore <rohitsrb2020@xxxxxxxxx> wrote:
> 
> Hi ferreters,
> 
> i have a simple query regarding the use of LSL_lowpass filter as it is mentioned like this in the documentation LSL_LOWPASS(A, cutoff_period, filter_span).
> where, 
> A is my monthly SST data from 1979-2014 i.e. 432 months, specification of A is given below 
> 
> currently SET data sets:
>     1> ./had_sst_anm_1979_2014.nc  (default)
>  name     title                                                                  I          J         K          L         M         N
>  SST_ANM  SST_1979_2014-TEMP_CLI[GT=SST_1     1:290    1:120     ...       1:432     ...         ...
> 
> i want to do filter for 7 years i.e. 84 months for decadal variability, so i am confused what should be the values i used for my cutoff_period and filter_span. For my analysis I used cutoff_period = 84 and filter_span = 85. Apart from this by using filter_span = 85 i am loosing 3 years data from beginning and Ending i.e. i am getting plot from 1982 to 2011 and when i am reducing the filter_span like  45, 25 and so on plot is reducing smoothness but the data loss is reducing i.e. i am getting values closer to 1979-2014. so kindly tell me what should be my filter_span to be used to get a good result.
> 
> regards
> 
> -- 
> 
> 
> REGARDS
> 
> Saurabh Rathore
> ​Research Scholar​ (Earth System Sciences & Technology)
> Centre For Oceans, Rivers, Atmosphere & Land Science Technology
> Indian Institute Of Technology, Kharagpur
> contact :- 91- 8345984434
> ​E-mail - rohitsrb2020@xxxxxxxxx​