Questions of Filtering options
Added by edith shin almost 14 years ago
Dear Manager
I have a questions of Filtering options.
First, I read the previous questions of others, but the answers are inconsistent.
One answer of 6-month filtering question is fmin=0.5 instead of 2, but the other answer of seasonal filtering question is fmin=4.
So i'm so confused which one is correct@.@
So I wanna ask you my problem.
I wanna bandpass filter of 2-8days in daily mean of 30 years data.
pls answer me what i should use in fmin & fmax.
And also i appreciate you your help if you explain me N &dT in documentaions in detail.
I read several times documentation and examples, but i coudln't get it;(
Finally, can i use bandpass filter in only 90 days data?
if possible, pls tell me how i decide fmin&fmax.
Thank you so much.
Edith
Replies (1)
RE: Questions of Filtering options - Added by Cedrick Ansorge almost 14 years ago
Dear Edith,
First, the arguments are given as frequencies with the unit [1/year], thus your frequency window is picked by setting
fmin = 365/8 fmax = 365/2. Your frequency range looks a bit as if you wanted to use the 'standard' blackman filter approach for synoptic data. Please Note, that our Filter has different characteristics as it uses a rectangular filter function in frequency space.
Second, concerning the definition of N and dT, I refer to the following lines of our online manual which you can find under [1]:
- the lowest frequency greater zero that can be contained in ifile is 1/(N*dT),
- the greatest frequency is 1/(2dT) (Nyquist frequency),
with N the number of time steps and dT the time increment of ifile in years. I.e. N is the number of time steps your file contains, and dT is the time spacing from time step to time step.
Third, technically yes; statistically you will always have to consider the ratio of the lowest frequency in the filter and the lowest frequency in your data. Using the 2-8 days filter with 90 days of data, I would say, it is marginally okay, although you should not rely to much on what you see there. Responses in the transform caused by the non-periodicity of data become larger and larger as you decrease the length of a time series. Also do not forget to detrend in this case!
Best Regards,
Cedrick
[1] https://code.zmaw.de/embedded/cdo/1.5.0/cdo.html#x1-5790002.15.2