DIVAnd (Data-Interpolating Variational Analysis in n dimensions) performs an n-dimensional variational analysis/gridding of arbitrarily located observations. Observations will be interpolated/analyzed on a curvilinear grid in 1, 2, 3 or more dimensions. In this sense it is a generalization of the original two-dimensional DIVA version (still available here https://github.com/gher-ulg/DIVA but not further developed anymore).
The method bears some similarities and equivalences with Optimal Interpolation or Krigging in that it allows to create a smooth and continous field from a collection of observations, observations which can be affected by errors. The analysis method is however different in practise, allowing to take into account topological features, physical constraints etc in a natural way. The method was initially developped with ocean data in mind, but it can be applied to any field where localized observations have to be used to produce gridded fields which are "smooth".
Please cite this paper as follows if you use
DIVAnd in a publication:
Barth, A., Beckers, J.-M., Troupin, C., Alvera-Azcárate, A., and Vandenbulcke, L.: DIVAnd-1.0: n-dimensional variational data analysis for ocean observations, Geosci. Model Dev., 7, 225-241, doi:10.5194/gmd-7-225-2014, 2014.
(click here for the BibTeX entry).
Under Linux you will also need the packages
netcdf which you can install under Debian/Ubuntu with:
apt-get install make gcc libnetcdf-dev netcdf-bin
You need Julia (version 1.4 or 1.5) to run
DIVAnd. The command line version is sufficient for
Inside Julia, you can download and install the package by issuing:
using Pkg Pkg.add(PackageSpec(name="DIVAnd", rev="master"))
It is not recommended to download the source of
DIVAnd.jl directly (using the green Clone or Download button above) because this by-passes Julia's package manager and you would need to install the dependencies of
To update DIVAnd, run the following command and restart Julia (or restart the jupyter notebook kernel):
using Pkg Pkg.add(PackageSpec(name="DIVAnd", rev="master"))
A test script is included to verify the correct functioning of the toolbox.
The script should be run in a Julia session.
Make sure to be in a directory with write-access (for example your home directory).
You can change the directory to your home directory with the
All tests should pass without error.
INFO: Testing DIVAnd Test Summary: | Pass Total DIVAnd | 427 427 INFO: DIVAnd tests passed
The test suite will download some sample data. You need to have Internet access and run the test function from a directory with write access.
The main routine of this toolbox is called
DIVAnd which performs an n-dimensional variational analysis of arbitrarily located observations. Type the following in Julia to view a list of parameters:
using DIVAnd ?DIVAndrun
DIVAnd_simple_example_4D.jl is a basic example in fours dimensions. The call to
DIVAndrun looks like this:
fi,s = DIVAndrun(mask,(pm,pn,po,pq),(xi,yi,zi,ti),(x,y,z,t),f,len,epsilon2);
mask is the land-sea mask, usually obtained from the bathymetry/topography,
(pm,pn,po,pq) is a n-element tuple (4 in this case) containing the scale factors of the grid,
(xi,yi,zi,ti) is a n-element tuple containing the coordinates of the final grid,
(x,y,z,t) is a n-element tuple containing the coordinates of the observations,
f is the data anomalies (with respect to a background field),
len is the correlation length and
epsilon2 is the error variance of the observations.
The call returns
fi, the analyzed field on the grid
DIVAndrun is the core analysis function in n dimensions. It does not know anything about the physical parameters or units you work with. Coordinates can also be very general. The only constraint is that the metrics
(pm,pn,po,...) when multiplied by the corresponding length scales
len lead to non-dimensional parameters. Furthermore the coordinates of the output grid
(xi,yi,zi,...) need to have the same units as the observation coordinates
DIVAndgo is only needed for very large problems when a call to
DIVAndrun leads to memory or CPU time problems. This function tries to decide which solver (direct or iterative) to use and how to make an automatic domain decomposition. Not all options from
DIVAndrun are available.
diva3D is a higher-level function specifically designed for climatological analysis of data on Earth, using longitude/latitude/depth/time coordinates and correlations length in meters. It makes the necessary preparation of metrics, parameter optimizations etc you normally would program yourself before calling the analysis function
If zero is not a valid first guess for your variable (as it is the case for e.g. ocean temperature), you have to subtract the first guess from the observations before calling
DIVAnd and then add the first guess back in.
epsilon2 and parameter
len are crucial for the analysis.
epsilon2 corresponds to the inverse of the signal-to-noise ratio.
epsilon2 is the normalized variance of observation error (i.e. divided by the background error variance). Therefore, its value depends on how accurate and how representative the observations are.
len corresponds to the correlation length and the value of
len can sometimes be determined by physical arguments. Note that there should be one correlation length per dimension of the analysis.
One statistical way to determine the parameter(s) is to do a cross-validation.
You can repeat all steps with a different validation data set to ensure that the optimal parameter values are robust. Tools to help you are included in (DIVAnd_cv.jl).
DIVAnd allows the calculation of the analysis error variance, scaled by the background error variance. Though it can be calculated "exactly" using the diagonal of the error covariance matrix s.P, it is too costly and approximations are provided. Two version are recommended,
DIVAnd_cpme for a quick estimate and
DIVAnd_aexerr for a version closer the theoretical estimate (see Beckers et al 2014 )
An arbitrary number of additional constraints can be included to the cost function which should have the following form:
J(x) = ∑i (Ci x - zi)ᵀ Qi-1 (Ci x - zi)
For every constrain, a structure with the following fields is passed to
yo: the vector zi
H: the matrix Ci
R: the matrix Qi (symmetric and positive defined)
Internally the observations are also implemented as constraint defined in this way.
On the server, launch the notebook with:
jupyter-notebook --no-browser --ip='0.0.0.0' --port=8888
where the path to
jupyter-notebook might have to be adapted, depending on your installation. The
port parameters can also be modified.
Then from the local machine it is possible to connect to the server through the browser.
Thanks to Lennert and Bart (VLIZ) for this trick.
Some examples in
DIVAnd.jl use a quite large data set which cannot be efficiently distributed through
git. This data can be downloaded from the URL https://dox.ulg.ac.be/index.php/s/Bo01EicxnMgP9E3/download. The zip file should be decompressed and the directory
DIVAnd-example-data should be placed on the same level than the directory
An educational web application has been developed to reconstruct a field based on point "observations". The user must choose in an optimal way the location of 10 observations such that the analysed field obtained by
DIVAnd based on these observations is as close as possible to the original field.
5 days ago