Interpolate environmental data onto a finer grid

Puts a coarse field onto a finer grid, so a 0.25 degree model variable can join 4 km satellite data without the satellite being coarsened to meet it.

Usage

downscale_grid(
  env_dat,
  to,
  vars = NULL,
  method = "nearest",
  idw_radius = NULL,
  idw_power = 2
)

Arguments

env_dat

an sf POINT object from accessEnvDat(), on a regular grid

to

the target grid: either a resolution in the CRS's units (one number for square cells, or c(x, y)), or another sf object whose grid to adopt — which is the form that puts two products onto one grid.

vars

columns to resample; NULL uses all covariate columns, as covariate_columns() reports them

method

one of "nearest", "bilinear", "cubic", "idw"; length one for all variables, or a named vector per variable

idw_radius

search radius for method = "idw", in the CRS's units. Defaults to twice the source cell size, which reaches the immediate neighbours; larger draws from further away and smooths more.

idw_power

distance exponent for method = "idw". Higher concentrates weight on the nearest points, approaching nearest in the limit.

Value

an sf POINT object on the target grid, with the same YEAR/MONTH/DAY columns as the input

What this does and does not do

Downscaling adds cells, not information. A 0.25 degree field rendered at 4 km has 4 km cells, but it still resolves nothing below 0.25 degrees, and no method here changes that — none of them consult any other data source, so there is nowhere for real fine-scale structure to come from.

That distinction is easy to lose downstream, which is why nearest is the default. It replicates each coarse value across the fine cells inside it, so the result is visibly blocky at the source resolution: the output looks like what it is. bilinear and cubic return a smooth field that looks like a finely-resolved measurement and is not one. They are the better choice when the field really is smooth at the source scale and blockiness would be an artefact — sea surface height, say — and the worse choice when someone later computes a gradient from the result, because that gradient is then a property of the interpolator.

• nearest — the value of the containing coarse cell. Invents nothing; blocky. Required for categorical fields, and applied to them automatically.

• bilinear — weighted by the four surrounding cell centres. Smooth, and NA anywhere a contributing neighbour is NA, so gaps grow by a cell.

• cubic — smoother again, over sixteen cells. Can overshoot past the source range near sharp edges, which for a bounded quantity like chlorophyll can produce negatives.

• idw — inverse distance weighting over the source points within idw_radius. Unlike the others it fills across holes rather than propagating them, which makes it the one to use on gappy satellite data where fill_satellite_gaps() is not an option.

See also

upscale_grid() for the other direction, fill_satellite_gaps()

Examples

if (FALSE) { # \dontrun{
# Model chlorophyll (0.25 deg) onto the satellite grid, blockiness intact
fine <- downscale_grid(chl_model, to = chl_satellite)

# Smooth, for a field that really is smooth at the source scale
ssh <- downscale_grid(ssh_model, to = 0.04, method = "bilinear")
} # }