dankelley
commented
7 years ago Since I wanted this in my own work, I've added but it is not in the namespace and it is not documented. Heck, you have to look in src/filtering.c to see how to use the thing! Therefore, I feel justified in pushing this to the develop branch. At the moment, it is just copying entries at the ends of the timeseries ... and, actually, I think that makes the most sense because it's easy to describe and therefore won't get users wound up into knots about the endpoints. Below is a sample code, cut/pasted from inside src/filtering.c
system("R CMD SHLIB filtering.c")
dyn.load("filtering.so");
#' Perform hamming filtering
#'
#' NOTE that this is not in the namespace. Also, the behaviour is not in a final
#' form, since this just copies over at the ends of x. (I'm not even sure
#' that is bad, come to think of it.)
#'
#' @param x a vector to be smoothed
#' @param n length of filter (must be an odd integer exceeding 1)
hamming <- function(x, n)
.Call("hamming", x, n)
x <- seq(-6, 6) + rnorm(13, sd=0.5)
y <- hamming(x, 5)
print(data.frame(x, y))
## next work in my testing, so I'm doing the formula right. But note
## that we need to normalize the filters ... why the heck are they
## not normalized in signal::hamming() or on wikipedia?
## signal::hamming(3) ## [1] 0.08 1.00 0.08
## signal::hamming(5) ## [1] 0.08 0.54 1.00 0.54 0.08
par(mar=c(3, 3, 1, 1), mgp=c(2, 0.7, 0))
plot(seq_along(x), x, type='b')
points(seq_along(x), y, col=2)
lines(seq_along(x), y, col=2)
grid()
richardsc
It seems that every time I want to smooth a time series, it's back to the beginning, reading documentation (including some in oce) that seems overly complicated. And then, if the task is a professional one, I'll need to explain how it was done^[1]. I'd prefer, for simple cases, to just say I used a N-point hamming filter, or maybe a hann filter.
And I don't want to have to choose to have NAs at the ends or to have terrible results if there is a trend (here, I'm talking about the two choices for
circularinfilter()). Heck, I also don't like having to choose whether to use thefilterthat R supplies, or the same name function that thesignalpackages provides.I want to be able to say, simply "the time-series was smoothed with a 5-point Hamming filter" and be done with it.
I'm starting to think that, as a user, I'd like to have a function called
hamming(), plain and simple, with two arguments,xandn. I think hamming is superior to hanning, but the latter could be added also. Honestly, I half-forgot that oce has this thing calledoceFilter, and I'm not sure what it does that's special, without reading the docs (maybe it handles NA better?)The other thing is for end points. I think a sensible scheme would be to use whatever filter coefficients are left for both sides. (That is, just start reducing filter length from the sides, dividing through to get the same sum of course.) So, the first and last points of
xoutwould be the same asx, and nearby points would be smoothed a little, and then a little more, until the interior, where the full filter would be used.I'm thinking out loud here. This would likely get done in C because this filter-trimming is not something I've heard of in the existing R functions. That means that creating it would take more than 10 minutes. It would be a Saturday morning thing.
Does this seem like something useful, to the other developers and users?
[1]: and that e.g. the matlab Butterworth filters don't work the same way as R ones, owing to treating starting conditions differently ... matlab is better when there are trends in the data