Open GoogleCodeExporter opened 9 years ago
Original comment by guillaum...@gmail.com
on 29 May 2014 at 7:25
Hi! I only recently discovered this toolbox, but am looking forward to trying it out and potentially contributing.
As for tidal analysis, I would look at Rich Pawlowicz's t_tide
package for Matlab:
I have just pushed a new branch https://github.com/aodn/imos-toolbox/tree/tidal_analysis with auto 1D tidal analysis from UTide and plot of residuals.
When loading a deployment, the tidal analysis can be called from the menu for the selected instrument (which has to have pressure data):
By default the tidal analysis is ran with iteratively reweighted least squares IRLS method (requires Statistics toolbox though) in auto mode (constituents are selected following the automated decision tree method with Rmin = 1). The IRLS, as opposed to the ordinary least squares OLS method, can minimize the influence of outliers, which is of a great benefit when dealing with instruments on an anchored mooring subject to current knock downs.
On top of being able to detect outliers and other interesting phenomenon like mooring knock downs, I'm hoping the residuals would help to detect or correct time and/or pressure drifts.
The following 1D tidal analysis have been performed in auto mode. For all of them, the 35 first main constituents have been automatically selected.
Tidal analysis of a pressure sensor @20m for a 120m deep site, that is subject to important knock downs, using OLS method:
Using OLS method, the residuals follow a sinusoidal pattern which shows important mismatch between observations and model. Manually selecting only relevant constituents based on diagnostic parameters may help to obtain a better reconstructed fit.
Tidal analysis of the same sensor using IRLS method:
Using IRLS method, there is a much better reconstructed fit with a lot higher SNR on all constituents than with using OLS method.
Tidal analysis with IRLS of a pressure sensor @120m for a 120m deep site:
Good reconstructed fit with no Pearson residual higher than 1 in absolute value.
Tidal analysis of the same sensor with introduced linear drift on clock of 5min over the deployment:
Tidal analysis of the same sensor with introduced quadratic drift on clock of 5min over the deployment:
In both cases the RMS of the residuals is slightly higher and the SNR for all constituents slightly lower than in the original data. So if we attempt to correct linear or non-linear clock drift, the tidal analysis should enable us to assess the improvement on the corrected data.
Tidal analysis of the same sensor with introduced linear drift on pressure of 3dBar over the deployment:
Diagnostics and residuals don't help to testify the linear pressure drift. Fortunately this might be a rare case in reality.
Tidal analysis of the same sensor with introduced quadratic drift on pressure of 3dBar over the deployment:
A U shape in the residuals clearly shows the lack of agreement between observations and model. If we attempt to correct such a drift, the U shape should disappear and we should see a more evenly distributed residuals vertically around 0.
Original issue reported on code.google.com by
guillaum...@gmail.com
on 27 May 2014 at 6:39