gpfield

gpfield is the spatial member of an agricultural-analytics R-package orchestra. It does change-of-support Gaussian-process regression for field and multi-environment-trial (MET) data: it smooths a noisy field surface, predicts at a support different from the one observed (a point process predicted as paddock-scale block averages, with the predictive variance propagated through the support map), and -- the part that matters in practice -- abstains when the data support or the estimated spatial range cannot bear the requested prediction resolution, rather than returning a confident wrong surface.

The Gaussian process is a self-contained exact solver in base R, so the package takes no heavy external GP dependency and installs and checks anywhere.

Related packages

gstat does variogram modelling and point or block kriging, including spatio-temporal kriging and simulation.
fields fits curves and surfaces to spatial data with splines and kriging, choosing smoothing and covariance parameters by cross-validation or REML.
GpGp fits Gaussian process models to large spatial and spatio-temporal data quickly with Vecchia's approximation.

gpfield fits an exact Gaussian process, whose cost grows with the cube of the number of observations, so it suits field trials rather than very large datasets. It adds a typed refusal when the estimated range or the data cannot support the requested prediction, a one-call smoother for row-column trial layouts (gp_field_smooth()), and output in the orchestra manifest that the other orchestra packages read.

Installation

# install.packages("remotes")
# remotes::install_github("max578/gpfield")
install.packages("gpfield", repos = c('https://max578.r-universe.dev', 'https://cloud.r-project.org'))

What it does

Fit an exact spatial (or spatio-temporal) GP with a squared-exponential or Matern covariance, hyperparameters estimated by marginal likelihood.
Change of support -- predict block (areal) averages from point observations, with within-block correlation propagated into the predictive variance, not averaged away.
Honest abstention -- a typed refusal with a machine-readable reason when the range or support cannot bear the request.
MET smoothing -- one call from a row-column trial layout to a smoothed surface.
Orchestra manifest -- emit a result in the federation's shared, versioned, integrity-hashed contract so it composes with the other members.

A minimal example

library(gpfield)

# A noisy field on a 10 x 10 plot grid.
set.seed(1L)
field <- expand.grid(row = seq_len(10L), col = seq_len(10L))
field$yield <- 3 + 0.3 * field$row + sin(field$col / 2) +
  rnorm(nrow(field), 0, 0.2)

# Fit the GP.
fit <- gp_fit(gpfield_spec(c("row", "col"), "yield"), field, seed = 1L)
fit
#> <gpfield_fit>
#>   kernel:      matern
#>   n training:  100
#>   length-scale (standardised): 4.34
#>   range (raw units):           34.31
#>   marginal var / noise var:    3.317 / 0.02171
#>   log marginal likelihood:     10.6

# Predict two paddock-scale block averages (change of support).
blocks <- block_support(
  lower = rbind(c(1, 1), c(6, 6)),
  upper = rbind(c(5, 5), c(10, 10))
)
gp_predict(fit, support = "block", blocks = blocks, seed = 1L)
#> <gpfield_prediction>
#>   support:   block
#>   locations: 2
#>   mean range: [4.725, 4.778]
#>   sd range:   [0.03812, 0.03812]

# Ask for a resolution the field cannot support: gpfield abstains.
gp_predict(fit, data.frame(row = c(1, 60), col = c(1, 60)))
#> <gpfield_abstention>
#>   scope:  gp_predict
#>   reason: range_too_short
#>   detail: query spacing 83.44 exceeds 1 x the estimated range 34.31; the field cannot be resolved this finely
#>     spacing = 83.44
#>     range = 34.31
#>     spacing_tol = 1

Contributing

Bug reports and suggestions are welcome through the GitHub issues.

Citation

citation("gpfield")