Dear List,
I've been working on a way to make knn forest maps from satellite imagery and
I'm using both the raster package (to prepare my data) and the yaImpute package
(to do the knn). The problem is that the yaImpute package requires data frames
as inputs and I have rasters. So I have to do:
Raster -> data.frame -> knn -> data.frame -> raster
Which is not efficient. I would like to be able to do directly:
Raster -> knn -> raster
Using the calc() fonction from the raster. I've made a function that I thought
would work however it doesn't, it gives me the error: "cannot use this
function". Any idea if my task is doable?
Here is a code example:
####
library(raster);library(yaImpute)
### preparing the raster used as input:
r <- raster(nrow=50, ncol=50, xmn=0, xmx=50, ymn=0, ymx=50, crs=NA)
r.x1 <- r.x2 <- r.x3 <- r.x4 <- r.y1 <- r.y2 <- r
set.seed(123)
values(r.x1) <- rnorm(2500,10,1)
values(r.x2) <- rnorm(2500,13,1.1)
values(r.x3) <- rnorm(2500,20,5)
values(r.x4) <- rnorm(2500,.5,.05)
values(r.y1) <- rnorm(2500,10,1)
values(r.y2) <- rnorm(2500,50,5)
r.x.st <- stack(r.x1,r.x2,r.x3,r.x4)
names(r.x.st) <- c("varX1", "varX2","varX3","varX4")
r.y.st <- stack(r.y1,r.y2)
names(r.y.st) <- c("varY1", "varY2")
###
### Making the table to use in yaImpute as reference observations
tab.x <- sampleRegular(r.x.st, 100, cells=T) # choose the reference
observations
rownames(tab.x) <- tab.x[,1]
tab.x <- tab.x[,-1]
tab.y <- sampleRegular(r.y.st, 100, cells=T) # choose the reference
observations
rownames(tab.y) <- tab.y[,1]
tab.y <- tab.y[,-1]
###
### Praparing the knn (it's kind of a pre-knn that uses only the references
set.
### The object that it produces here is latter used on the target observations
res.yai <- yai(x=tab.x, y=tab.y, k=4, method="euclidean")
###
### My function to get the prediction for a target observation
prod.knn.raster <-function(vec.bandes, info.yai, na.rm=T){
dat <- rbind(vec.bandes,vec.bandes) # yaImpute needs a
matrix as input (of usually at least 2 observations or it doesn't work)
rownames(dat) <-
c(paste0("o",round(runif(1,1,1000000000000000))),paste0("o",round(runif(1,1,1000000000000000)),"b"))
# it need a UNIQUE id for the pixel, there is maybe a way to use the cell
number but I don't know it.
res.knn.cible<-newtargets(info.yai,dat) # find the k
nearest neighbours associated with the pixel of the raster
res.imp <- impute.yai(res.knn.cible,observed=F) # impute the values
of the response variable for the pixel
res <- res.imp [1,] # keep one of the
observations
res # returns a vector
of same length as the number of Y variables
}
### If I try my fonction directly it works
prod.knn.raster(vec.bandes=r.x.st[1], info.yai=res.yai, na.rm=T)
varY1 varY2
o87892296724022 11.17959 48.55119
### but if I put it in calc it doesn't:
qwqw <- calc(r.x.st, prod.knn.raster,info.yai=res.yai)
Error in .calcTest(x[1:5], fun, na.rm, forcefun, forceapply) :
cannot use this function
####
I have a feeling the problem is linked with the fact I have to get the res.yai
object into the function...
Anyway, thanks a lot for the potential help!
Bastien Ferland-Raymond, M.Sc. Stat., M.Sc. Biol.
Division des orientations et projets spéciaux
Direction des inventaires forestiers
Ministère des Ressources naturelles
880, chemin Sainte-Foy, 5e étage
Québec (Québec) G1S 4X4
_______________________________________________
R-sig-Geo mailing list
R-sig-Geo@r-project.org
https://stat.ethz.ch/mailman/listinfo/r-sig-geo
