Skip to contents

Build Compositional Geostatistical Model

gc_ilr_model.Rd

Constructs a gstat model for the ILR-transformed variables using either Independent Univariate Kriging or Linear Model of Coregionalization (LMC).

Usage

gc_ilr_model(
  ilr_params,
  variogram_model,
  data = NULL,
  model_type = "univariate"
)

Arguments

ilr_params

A list from gc_ilr_params() containing mean, cov, names.

variogram_model

A vgm() object defining the base variogram structure.

data

Optional sf object with ILR columns for conditioning. If provided, Conditional Simulation is performed. If NULL (default), Unconditional Simulation is performed.

model_type

Character string specifying the approach: "univariate" (default) or "lmc". Univariate is numerically stable and standard practice. LMC includes cross-covariance terms between ILR dimensions.

Value

A gstat object representing the model.

Details

Independent Univariate Kriging (model_type = "univariate"): Models each ILR dimension separately without cross-covariance terms. This is:

  • Numerically stable (avoids positive-definite issues in LMC fitting)

  • Standard practice in compositional geostatistics

  • Robust across different datasets

  • Efficient for large problems

The ILR transformation already decorrelates the data significantly, so ignoring spatial cross-correlation between ILR coordinates has minimal impact.

Linear Model of Coregionalization (model_type = "lmc"): Includes cross-covariance terms between all pairs of ILR dimensions. This is:

  • Theoretically more complete

  • More numerically complex

  • Useful when cross-correlation structure is important

For conditional simulation, the conditioning data must be passed to this function (not to predict()). The model then automatically uses that data during prediction.