PyKrige internally supports the six variogram models listed below. Additionally, the code supports user-defined variogram models via the ‘custom’ variogram model keyword argument.
Where s is the slope and n is the nugget.
Where s is the scaling factor, e is the exponent (between 0 and 2), and n is the nugget term.
Variables are defined as:
\(d\) = distance values at which to calculate the variogram
\(p\) = partial sill (psill = sill - nugget)
\(r\) = range
\(n\) = nugget
\(s\) = scaling factor or slope
\(e\) = exponent for power model
For stationary variogram models (gaussian, exponential, spherical, and hole-effect models), the partial sill is defined as the difference between the full sill and the nugget term. The sill represents the asymptotic maximum spatial variance at longest lags (distances). The range represents the distance at which the spatial variance has reached ~95% of the sill variance. The nugget effectively takes up ‘noise’ in measurements. It represents the random deviations from an overall smooth spatial data trend. (The name nugget is an allusion to kriging’s mathematical origin in gold exploration; the nugget effect is intended to take into account the possibility that when sampling you randomly hit a pocket gold that is anomalously richer than the surrounding area.)
For nonstationary models (linear and power models, with unbounded spatial variances), the nugget has the same meaning. The exponent for the power-law model should be between 0 and 2 1.
A few important notes:
The PyKrige user interface by default takes the full sill. This default behavior can be changed with a keyword flag, so that the user can supply the partial sill instead. The code internally uses the partial sill (psill = sill - nugget) rather than the full sill, as it’s safer to perform automatic variogram estimation using the partial sill.
The exact definitions of the variogram models here may differ from those used elsewhere. Keep that in mind when switching from another kriging code over to PyKrige.
According to 1, the hole-effect variogram model is only correct for the 1D case. It’s implemented here for completeness and should be used cautiously.