Skip to contents

The articles and help pages use the terms below in these senses.

The terms, A to Z: area of applicability (AOA), autocorrelation range, backend, block, cell, cell’s true mean, CRS units, design effect (deff), dissimilarity index (DI), flat region, fold, level, lon/lat, nugget, sill, range, pooled value, range floor, row ID, size, support ceiling, variogram.

area of applicability (AOA)
The prediction locations where a model’s cross-validated score can be expected to hold: those whose dissimilarity index is at or below a threshold, the largest dissimilarity index of the training rows that is not an outlier. area_of_applicability() computes it.
autocorrelation range
The distance beyond which two observations are close to uncorrelated, in the units of the CRS the variogram is fitted in. estimate_sac_range() returns it as the effective range of an exponential variogram: three times the fitted range parameter, where the semivariance has covered about 95 percent of the rise from the nugget to the sill. A block or a buffer has to exceed it to keep held-out rows apart from their training data.
backend
One of the three model fitters behind the spatial_fit class: the random forest (fit_rf_model()), GWR (geographically weighted regression, fit_gwr_model()) and the Bayesian GP (a Bayesian model with a Gaussian-process term in the coordinates, fit_bayesian_spatial_model()). new_spatial_fit() wraps a model of your own in the same class.
block
A piece of the study area that a fold holds out whole: a cell of a rectangular grid, at least block_size on each side when that is given, or one of the polygons passed in blocks. make_folds(method = "block_kfold") builds the grid.
cell
A polygon of a tessellation: a Voronoi cell, a hexagon, a square or a Delaunay triangle. assign_features_to_polygons() puts each observation in one cell, and the poly_id column names it.
cell’s true mean
What a cell value estimates: the expected value of one further observation located in the cell the way its observations were. It is not the average of a surface over the cell’s area, which would give weight to ground that holds no observation. “Pooled values” in ?summarize_by_cell defines it.
CRS units
The unit of the coordinate reference system a function works in, such as the metre for a UTM zone. Distances, block sizes, bandwidths and ranges are given in it. A function that measures distances moves a lon/lat layer into a projected CRS that ensure_projected() chooses, whose unit is the metre (?"spatialkit-geometry").
design effect (deff)
The factor by which correlation among a cell’s observations inflates the variance of the cell mean as an estimate of the population mean; n / deff is then the cell’s effective sample size. summarize_by_cell(deff = ) applies it to the standard errors. At the default, deff = 1, the standard error is that of the cell’s own mean, which needs no such correction when the cell’s observations are spread through it.
dissimilarity index (DI)
The distance from a prediction location to the nearest training row in the space of the predictors, each centred and scaled by the training data’s own mean and standard deviation and weighted by importance when the model carries it, divided by the mean distance between training rows. Meyer and Pebesma (2021) define it, and area_of_applicability() computes it.
flat region
The levels whose criterion value is within tol of the optimum, which select_resolution() returns with the count it chooses. The criteria do not rise or fall steadily with the count, so the flat region is a set of levels that can leave out some in between. Report it with the count.
fold
One split of a cross-validation: the rows held out and the rows the model is trained on to predict them. make_folds() returns the folds, each a train and a test set of row IDs.
level
A candidate cell count on the ladder that resolution_profile() scores. select_resolution() chooses one of the levels.
lon/lat
Coordinates in degrees of longitude and latitude, in a geographic CRS. See CRS units for how distances are measured on such a layer.
nugget, sill, range
The variogram’s value at zero distance (the noise), the value it levels off at (the total variance), and the range, the distance over which it rises toward the sill. The exponential model estimate_sac_range() fits approaches the sill without reaching it; see autocorrelation range. sac_nugget() reads the nugget of an estimate_sac_range() result.
pooled value
A cell’s mean pulled toward the mean of the observations in the 8 nearest cells (nearest by the centroid of their observations). The own weight, the share of the pooled value that is the cell’s own mean, grows with the cell’s count. summarize_by_cell() returns pooled values beside the plain means by default (estimate = "pooled"), and “Pooled values” in its help gives the formula.
range floor
The fewest cells the autocorrelation range allows: the area of the points’ convex hull divided by the square of the range, rounded up, and at least 2. A cell wider than the range averages over more than one patch of the field. resolution_profile() reports the floor and by default does not apply it; select_resolution(criterion = "reliability") reads the levels from the floor up.
row ID
..row_id, the column by which folds, cross-validated predictions and the record of dropped rows name rows. The cv_*() functions number the rows in it when a layer has none. make_folds() names rows by this column when the layer has it, and by row position otherwise.
size
The mean number of observations a cell is to hold, obs_per_cell in resolution_profile(). A size s gives max(2, round(n / s)) cells, with n the rows that carry a response, or every point without one. The default profile scores one size, 4.5, and that sets the default count of cells for pooled values.
support ceiling
The most cells the data support on a ladder of counts: the scored rows (those with a response and every predictor, or every point without a response) divided by min_cell_n, 9 by default, rounded down. On a ladder of sizes the finest level takes its place. When a criterion’s optimum sits on the ceiling or on the range floor, that bound sets the count, and select_resolution() names it in edge.
variogram
The semivariance of a variable against the distance between pairs of points, where the semivariance is half the mean squared difference of the pairs at that distance. estimate_sac_range() fits an exponential model with a nugget to it, and plot() on the result (?plot.sac_range) draws the variogram, with the fitted model where a range was identified.

Next

This is the last article in reading order. The help pages that give these terms in full are ?area_of_applicability, ?estimate_sac_range, ?make_folds, ?resolution_profile, ?select_resolution, ?summarize_by_cell and ?"spatialkit-geometry". ?spatialkit lists every step of the pipeline with its functions.

References

Meyer, H. and Pebesma, E. (2021). Predicting into unknown space? Estimating the area of applicability of spatial prediction models. Methods in Ecology and Evolution 12(9), 1620–1633. https://doi.org/10.1111/2041-210X.13650