terra and raster extract output differs for categorical values

[tomroh]

For the raster package:

x <- raster::raster(<my-file>)
y <- sf::st_buffer( sf::st_as_sfc(list(sf::st_point(c(<Longitude>, <Latitude>))), crs = 4326),  dist = 10)
terra::extract(x, sf::as_Spatial(y))

Returns the integer values of the categorical values in the dataset.

z <- terra::rast(<my-file>)
terra::extract(z, terra::vect(y))

Returns a factor values but the levels are not the same integer levels within the raster dataset.

Per https://stackoverflow.com/questions/69533813/terra-extracting-incorrect-value-from-categorical-raster, this can be resolved by:

levels(z) <- NULL
terra::extract(z, terra::vect(y))

I could not find this anywhere in the documentation. Can an argument be added that turns off the "factoring" the values? Or can the returned factor include the correct integer values?

[rhijmans]

What would be a good use-case where you would want the "code" (integer level) rather than the actual "value" (factor label)?

[tomroh]

The "code" can be a key in a larger database where the "code" is used as the link in a relational data model. i.e. extracting the "values" and trying to join them to another table would break the model.

[rhijmans]

Example data

library(terra)
set.seed(0)
r <- rast(nrows=10, ncols=10)
values(r) <- sample(3, ncell(r), replace=TRUE) + 10
levels(r) <- data.frame(id=11:13, cover=c("forest", "water", "urban"), letters=letters[1:3], value=10:12)

Default

extract(r, cbind(0,0))
#  cover
#1 water

Solution 1: Set the active category to 0

activeCat(r) <- 0
extract(r, cbind(0,0))
#  id
#1 12

Solution 2: Remove the levels

activeCat(r) <- 1
levels(r) <- NULL
extract(r, cbind(0,0))
#    cover
#1      12

I could also add an argument "raw=TRUE" to extract, as you suggest. That used to be there (as factors=TRUE), but the large number of options and their interactions, was making things a bit complicated. I would prefer not to go back there, but it may be preferable so that you can have a simple workflow that you know will work whether there are factors or not.

Your idea of having the "raw values" as the R factor indices as well is good, in principle, but I am not sure if that is possible. I think they have to be from 1 to n. That would require adding, perhaps very many, empty levels if the raw values start at, say, 10,000.

[rhijmans]

I have added argument raw=FALSE. When set to TRUE a matrix with the raw values is returned

library(terra)
#terra 1.6.10
set.seed(0)
r <- rast(nrows=10, ncols=10)
values(r) <- sample(3, ncell(r), replace=TRUE) + 10
levels(r) <- data.frame(id=11:13, cover=c("forest", "water", "urban"), letters=letters[1:3], value=10:12)
 
extract(r, cbind(0,0))
#  cover
#1 water

extract(r, cbind(0,0), raw=TRUE)
#     cover
#[1,]    12

[tomroh]

Thanks!

[tomroh]

Your idea of having the "raw values" as the R factor indices as well is good, in principle, but I am not sure if that is possible. I think they have to be from 1 to n. That would require adding, perhaps very many, empty levels if the raw values start at, say, 10,000.

Yes, I thought this might be case with using factor. That would add too much overhead for large values. A formal datatype that behaves as integer coded strings allowing non-consecutive integer levels doesn't exist currently in R. This would be useful in a variety of applications in R.