GeoStat-Framework/GSTools: v1.3.0 'Pure Pink'

Release Notes

A big step forward for GSTools. We now support geographical coordinates, directional variograms, auto-binning, arbitrary dimensions, normalizers and trends and much much more.

Installation

You can install GSTools with conda:

conda install -c conda-forge gstools

or with pip:

pip install gstools

Documentation

The documentation can be found at: https://gstools.readthedocs.io/

What's new?

Topics

Geographical Coordinates Support (#113)

• added boolean init parameter latlon to indicate a geographic model. When given, spatial dimension is fixed to dim=3, anis and angles will be ignored, since anisotropy is not well-defined on a sphere.
• add property field_dim to indicate the dimension of the resulting field. Will be 2 if latlon=True
• added yadrenko variogram, covariance and correlation method, since the geographic models are derived from standard models in 3D by plugging in the chordal distance of two points on a sphere derived from there great-circle distance zeta:
  • vario_yadrenko: given by variogram(2 * np.sin(zeta / 2))
  • cov_yadrenko: given by covariance(2 * np.sin(zeta / 2))
  • cor_yadrenko: given by correlation(2 * np.sin(zeta / 2))
• added plotting routines for yadrenko methods described above
• the isometrize and anisometrize methods will convert latlon tuples (given in degree) to points on the unit-sphere in 3D and vice versa
• representation of geographical models don't display the dim, anis and angles parameters, but latlon=True
• fit_variogram will expect an estimated variogram with great-circle distances given in radians
• Variogram estimation
  • latlon switch implemented in estimate_vario routine
  • will return a variogram estimated by the great-circle distance (haversine formula) given in radians
• Field
  • added plotting routines for latlon fields
  • no vector fields possible on latlon fields
  • corretly handle pos tuple for latlon fields

Krige Unification (#97)

• Swiss Army Knife for kriging: The Krige class now provides everything in one place
• "Kriging the mean" is now possible with the switch only_mean in the call routine
• Simple/Ordinary/Universal/ExtDrift/Detrended are only shortcuts to Krige with limited input parameter list
• We now use the covariance function to build up the kriging matrix (instead of variogram)
• An unbiased switch was added to enable simple kriging (where the unbiased condition is not given)
• An exact switch was added to allow smother results, if a nugget is present in the model
• An cond_err parameter was added, where measurement error variances can be given for each conditional point
• pseudo-inverse matrix is now used to solve the kriging system (can be disabled by the new switch pseudo_inv), this is equal to solving the system with least-squares and prevents numerical errors
• added options fit_normalizer and fit_variogram to automatically fit normalizer and variogram to given data

Directional Variograms and Auto-binning (#87, #106, #131)

• new routine name vario_estimate instead of vario_estimate_unstructured (old kept for legacy code) for simplicity
• new routine name vario_estimate_axis instead of vario_estimate_structured (old kept for legacy code) for simplicity
• vario_estimate
  • added simple automatic binning routine to determine bins from given data (one third of box diameter as max bin distance, sturges rule for number of bins)
  • allow to pass multiple fields for joint variogram estimation (e.g. for daily precipitation) on same mesh
  • no_data option added to allow missing values
  • masked fields
    • user can now pass a masked array (or a list of masked arrays) to deselect data points.
    • in addition, a mask keyword was added to provide an external mask
  • directional variograms
    • diretional variograms can now be estimated
    • either provide a list of direction vectors or angles for directions (spherical coordinates)
    • can be controlled by given angle tolerance and (optional) bandwidth
    • prepared for nD
  • structured fields (pos tuple describes axes) can now be passed to estimate an isotropic or directional variogram
  • distance calculation in cython routines in now independent of dimension
• vario_estimate_axis
  • estimation along array axis now possible in arbitrary dimensions
  • no_data option added to allow missing values (sovles #83)
  • axis can be given by name ("x", "y", "z") or axis number (0, 1, 2, 3, ...)

Better Variogram fitting (#78, #145)

• fixing sill possible now
• loss is now selectable for smoother handling of outliers
• r2 score can now be returned to get an impression of the goodness of fitting
• weights can be passed
• instead of deselecting parameters, one can also give fix values for each parameter
• default init guess for len_scale is now mean of given bin-centers
• default init guess for var and nugget is now mean of given variogram values

CovModel update (#109, #122, #157)

• add new rescale argument and attribute to the CovModel class to be able to rescale the len_scale (usefull for unit conversion or rescaling len_scale to coincide with the integral_scale like it's the case with the Gaussian model) See: #90, GeoStat-Framework/PyKrige#119
• added new len_rescaled attribute to the CovModel class, which is the rescaled len_scale: len_rescaled = len_scale / rescale
• new method default_rescale to provide default rescale factor (can be overridden)
• remove doctest calls
• docstring updates in CovModel and derived models
• updated all models to use the cor routine and make use of the rescale argument (See: #90)
• TPL models got a separate base class to not repeat code
• added new models (See: #88):
  • HyperSpherical: (Replaces the old Intersection model) Derived from the intersection of hyper-spheres in arbitrary dimensions. Coincides with the linear model in 1D, the circular model in 2D and the classical spherical model in 3D
  • SuperSpherical: like the HyperSpherical, but the shape parameter derived from dimension can be set by the user. Coincides with the HyperSpherical model by default
  • JBessel: a hole model valid in all dimensions. The shape parameter controls the dimension it was derived from. For nu=0.5 this model coincides with the well known wave hole model.
  • TPLSimple: a simple truncated power law controlled by a shape parameter nu. Coincides with the truncated linear model for nu=1
  • Cubic: to be compatible with scikit-gstat in the future
• all arguments are now stored as float internally (#157)
• string representation of the CovModel class is now using a float precision (CovModel._prec=3) to truncate longish output
• string representation of the CovModel class now only shows anis and angles if model is anisotropic resp. rotated
• dimension validity check: raise a warning, if given model is not valid in the desired dimension (See: #86)

Normalizer, Trend and Mean (#124)

• new normalize submodule containing power-transforms for data to gain normality
• Base-Class: Normalizer providing basic functionality including maximum likelihood fitting
• added: LogNormal, BoxCox, BoxCoxShift, YeoJohnson, Modulus and Manly
• normalizer, trend and mean can be passed to SRF, Krige and variogram estimation routines
  • A trend can be a callable function, that represents a trend in input data. For example a linear decrease of temperature with height.
  • The normalizer will be applied after the data was detrended, i.e. the trend was substracted from the data, in order to gain normality.
  • The mean is now interpreted as the mean of the normalized data. The user could also provide a callable mean, but it is mostly meant to be constant.

Arbitrary dimensions (#112)

• allow arbitrary dimensions in all routines (CovModel, Krige, SRF, variogram)
• anisotropy and rotation following a generalization of tait-bryan angles
• CovModel provides isometrize and anisometrize routines to convert points

New Class for Conditioned Random Fields (#130)

• THIS BREAKS BACKWARD COMPATIBILITY
• CondSRF replaces the conditioning feature of the SRF class, which was cumbersome and limited to Ordinary and Simple kriging
• CondSRF behaves similar to the SRF class, but instead of a covariance model, it takes a kriging class as input. With this kriging class, all conditioning related settings are defined.

Enhancements

• Python 3.9 Support #107
• add routines to format struct. pos tuple by given dim or shape
• add routine to format struct. pos tuple by given shape (variogram helper)
• remove field.tools subpackage
• support meshio>=4.0 and add as dependency
• PyVista mesh support #59
• added EARTH_RADIUS as constant providing earths radius in km (can be used to rescale models)
• add routines latlon2pos and pos2latlon to convert lat-lon coordinates to points on unit-sphere and vice versa
• a lot of new examples and tutorials
• RandMeth class got a switch to select the sampling strategy
• plotter for n-D fields added #141
• antialias for contour plots of 2D fields #141
• building from source is now configured with pyproject.toml to care about build dependencies, see #154

Changes

• drop support for Python 3.5 #146
• added a finit limit for shape-parameters in some CovModels #147
• drop usage of pos2xyz and xyz2pos
• remove structured option from generators (structured pos need to be converted first)
• explicitly assert dim=2,3 when generating vector fields
• simplify pre_pos routine to save pos tuple and reformat it an unstructured tuple
• simplify field shaping
• simplify plotting routines
• only the "unstructured" keyword is recognized everywhere, everything else is interpreted as "structured" (e.g. "rectilinear")
• use GitHub-Actions instead of TravisCI
• parallel build now controlled by env-var GSTOOLS_BUILD_PARALLEL=1, see #154
• install extra target for [dev] dropped, can be reproduced by pip install gstools[test, doc], see #154

Bugfixes

• typo in keyword argument for vario_estimate_structured #80
• isotropic rotation of SRF was not possible #100
• CovModel.opt_arg now sorted #103
• CovModel.fit: check if weights are given as a string (numpy comparison error) #111
• several pylint fixes (#159)