Usually it is possible to predict tide at a certain location. Then, assuming no interaction between tide and surge, the surge fluctuation can be obtained by subtracting tide from the observed water level.

If you are familiar with Python you can try tappy, it works very fine and supports nodal correction (18.61 yr nodal cycle)

http://sourceforge.net/p/tappy/wiki/Main_Page/

It would help if you supplied more evidence of the suggested failure of t-tide. Also, you don't state in sufficient detail the nature of your study region ... maybe the signals are simply not tidal there, in which case you would need to incorporate other physics. Maybe if you write up a few-page document with diagrams, people could help.

The predominant difference between tidal and surge phenomena is the fact that the tide originates from astronomical mechanisms (and is thus very much deterministic in nature), while surges and other phenomena are more or less random.

The solution is therefore, to apply some sort of Fourier spectral analysis on your time series of water level signals. All tidal effects present in the signal will pop up in the spectrum as very distinct sinusoidal spectral lines at one or more specific tidal constituents, while the other (non-tidal) signals will show no distinct lines but a more continuous character.

After having identified all relevant tidal constituents you can subsequently subtract them from the original signal, thus yielding all other (non-tidal) effects.

succes.

T_tide is excellent and quite robust. You do need to ensure that the data interval is uniform and that the unit of time is consistent with what is expected (and I forget what that is). Those are the two most common user errors I have seen.  

I agree with Herman Peters, do subject your time series water levels to Fourier spectral analysis which will separate the surge and tide based on respective frequencies.

The value of the spectral approach is limited when records are short or gappy, or have a lot non tidal enery. If you have a few months at hourly intervals you can use it to isolate the semidiurnal constituents and estimate the noise amplitude. T_tide uses this as an optional way to estimate uncertainty in harmonic amplitudes.

Kyung-Duck: Actually the precise tide prediction for my study region.

Clément: Thanks. I'll try that package.

Dan: I attached a document file showing tide gauge records and analysis results. Could you please have a look at it?

Herman: Please see the attached document file. It seems that your method is the same as t_tide methodology, am I right? Anyway, I will follow this routine to compare the results with t_tide.

James: The data is uniform and without gap. The record duration is more than 3 months. So t-tide is expected to show less sinusoid fluctuations in the residuals.

Cynthia: Thanks for your answer.

 There is evidently high frequency oscillations in the residuals.  I suggest they may be overtides.  These are not automatically included in the T_tide. See below.

However, if you want to remove all the energy with periods shorter than 24 hrs, I suggest a simple low-pass filter.  Look up[

 http://woodshole.er.usgs.gov/operations/sea-mat/bobstuff-html/pl33tn.html

Use this to get the nontidal low frequency and then apply t-tide with the shallow switch.

=================================

FROM T_TIDE :

Shallow water constituents are not used automatically. They can

be used but you must specify them manually using 'shallow' input

option to T_TIDE. You can get the names and frequencies of

shallow water constituents from the CONST structure returned by

T_GETCONSTS):

CONST=t_getconsts;

CONST.name(finite(CONST.ishallow))

Dear James,

Thanks for your help. pl33tn uses 'detrend' function which can be downloaded from the same website. But 'detrend' calls 'polyfit2' function which is undefined in MATLAB. I found many 'polyfit2.m' files on the web, but none of them works! Could you please help me in this regard?

Dear James,

Thanks for your help. pl33tn uses 'detrend' function which can be downloaded from the same website. But 'detrend' calls 'polyfit2' function which is undefined in MATLAB. I found many 'polyfit2.m' files on the web, but none of them works! Could you please help me in this regard?

You must have an old version or a non-standard detrend.m. 

All that should be doing is fitting a line or quardatic to the series.  If you just subtract the mean yourself and comment out the detrend option you should get it to run at least.  Your series didn't seem to have much of a trend anyway so the result should be good.

The most straight forward way is to use harmonic analysis to create a tide time series and then subtract that from the data to create a residual. However, in our experience, there will always be some "tidal frequency" left overs for various reasons. A better way is to Fourier transform your time series and patch over the tidal bands in frequency space with random phase oscillations that match the adjacent spectrum. Then re-transform to time domain. This approach was successful for the long San Francisco tide gauge record. See: Bromirski, Flick, and Cayan, 2003. Storminess Variability along the California Coast: 1858 - 2000, J. Climate 16(6), 982-993. Your record is short, so that may create difficulty narrowing the spectral peaks, but on the other hand noise form errors and glitches will be easier to deal with.

I  know how vexing this can be! There may be several other approaches. First of all, be sure there are no bad data points. I assume you have plotted the data, which is easy for just a few months. We analyzed hourly data from San Francisco starting in 1859, so fixing glitches was a big deal! Anyway, make any adjustments needed, including filling gaps. Walter Munk many years ago published a paper called "Super Resolution" of tides, or something like that. In it, he essentially least-square fit a set of tidal constituents to a water level series. Find out which constituents are important in your study region and use those for a fit. The residual will be an estimate of the surge. The basis of this approach is assuming the tide signals are noise free, and that the rest is non-tide (surge, etc) and noise.Hence the "super-resolution" idea where you get more information seemingly for free than you can from simply FFT. Let us know how it goes.