First of all, some visual tools is used to explore the data. The data is plotted on original scale as well as on a first order trend.
From the first set of plots, we can see from the partial plots that the data is highly correlated with the coordinates. After we removed the first order trend (plot2), this correlation is greatly eliminated. Therefore in the following modeling procedures, we will always remove the first order trend first.
Again a first order trend is assumed when fitting the semivariogram. The parameters for the trend is \(\hat\beta_0 = 607.770661\), \(\hat\beta_1 = -1.278442\) and \(\hat\beta_2 = -1.138741\). The empirical semivariogram is plotted.
## variog: computing omnidirectional variogram
Then we tried to fit a model to the semivariogram. We considered exponential model and gaussian models, both with a first order trend, and with or without nugget effect. The results are summarised in the following table.
| model | AIC |
|---|---|
| exponential with nugget effect | 930.392205792884 |
| exponential with nugget effect | 929.732456215654 |
| gaussian with nugget effect | singular |
| gaussian with nugget effect | 946.099836017318 |
Finally the exponential model without nugget effect is chosen. The model fit to the empirical semivariogram like this.
Fianlly we want to do a universal kriging on 100 points on each axis, that’s 10000 points in the domain. The model is fit using the first order trend and semivariogram in model 2.
## krige.conv: model with mean given by a 1st order polynomial on the coordinates
## krige.conv: Kriging performed using global neighbourhood
From the prediction plots, we can see that the south-western corner are higher than the north-eastern corner. The prediction error are higer on the north-western corner because that’s where we have very few data points.