kriging.js is a Javascript library providing spatial prediction and mapping capabilities via the ordinary kriging algorithm.

Kriging is a type of gaussian process where 2-dimensional coordinates are mapped to some target variable using kernel regression. This algorithm has been specifically designed to accurately model smaller data sets by assigning a prior to the variogram parameters.

The first step is to link kriging.js to your html code and assign your coordinate and target variables to 3 separate arrays.

<script src="kriging.js" type="text/javascript"></script>
<script type="text/javascript">
	var t = [ /* Target variable */ ];
	var x = [ /* X-axis coordinates */ ];
	var y = [ /* Y-axis coordinates */ ];
	var model = "exponential";
	var sigma2 = 0, alpha = 100;
	var variogram = kriging.train(t, x, y, model, sigma2, alpha);
</script>

The train method in the kriging object fits your input to whatever variogram model you specify - gaussian, exponential or spherical - and returns a variogram object.

Notice the σ² (sigma2) and α (alpha) variables, these correspond to the variance parameters of the gaussian process and the prior of the variogram model, respectively. A diffuse α prior is typically used; a formal mathematical definition of the model is provided below.

Values can be predicted for new coordinate pairs by using the predict method in the kriging object.

  var xnew, ynew /* Pair of new coordinates to predict */;
  var tpredicted = kriging.predict(xnew, ynew, variogram);
  

The various variogram models can be interpreted as kernel functions for 2-dimensional coordinates a, b and parameters nugget, range, sill and A. Reparameterized as a linear function, with w = [nugget, (sill-nugget)/range], this becomes:

• Gaussian: k(a,b) = w[0] + w[1] * ( 1 - exp{ -( ||a-b|| / range )² / A } )
• Exponential: k(a,b) = w[0] + w[1] * ( 1 - exp{ -( ||a-b|| / range ) / A } )
• Spherical: k(a,b) = w[0] + w[1] * ( 1.5 * ( ||a-b|| / range ) - 0.5 * ( ||a-b|| / range )³ )

The variance parameter α of the prior distribution for w should be manually set, according to:

• w ~ N(w|0, αI)

Using the fitted kernel function hyperparameters and setting K as the Gram matrix, the prior and likelihood for the gaussian process become:

• y ~ N(y|0, K)
• t|y ~ N(t|y, σ²I)

The variance parameter σ² of the likelihood reflects the error in the gaussian process and should be manually set.