terra (version 0.4-3)

predict: Spatial model predictions

Description

Make a SpatRaster object with predictions from a fitted model object (for example, obtained with glm or randomForest). The first argument is a SpatRaster object with the predictor variables. The names in the Raster object should exactly match those expected by the model. Any regression like model for which a predict method has been implemented (or can be implemented) can be used.

This approach of using model predictions is commonly used in remote sensing (for the classification of satellite images) and in ecology, for species distribution modeling.

Usage

# S4 method for SpatRaster
predict(object, model, fun=predict, ..., factors=NULL, 
    const=NULL, na.rm=FALSE, index=NULL, filename="", overwrite=FALSE, wopt=list())

Arguments

object

SpatRaster

model

fitted model of any class that has a "predict" method (or for which you can supply a similar method as fun argument. E.g. glm, gam, or randomForest

fun

function. Default value is predict, but can be replaced with e.g. predict.se (depending on the type of model), or your own custom function

...

additional arguments for fun

const

data.frame. Can be used to add a constant value as a predictor variable so that you do not need to make a SpatRaster layer for it

factors

list with levels for factor variables. The list elements should be named with names that correspond to names in object such that they can be matched. This argument may be omitted for standard models such as "glm" as the predict function will extract the levels from the model object, but it is necessary in some other cases (e.g. cforest models from the party package)

na.rm

logical. If TRUE, cells with NA values in the predictors are removed from the computation. This option prevents errors with models that cannot handle NA values. In most other cases this will not affect the output. An exception is when predicting with a model that returns predicted values even if some (or all!) variables are NA

index

integer. To select subset of output variables

filename

character. Output filename. Optional

overwrite

logical. If TRUE, filename is overwritten

wopt

list. Options for writing files as in writeRaster

Value

SpatRaster

Examples

# NOT RUN {
logo <- rast(system.file("exdata/logo.tif", package="terra"))   
names(logo) <- c("red", "green", "blue")
p <- matrix(c(48, 48, 48, 53, 50, 46, 54, 70, 84, 85, 74, 84, 95, 85, 
   66, 42, 26, 4, 19, 17, 7, 14, 26, 29, 39, 45, 51, 56, 46, 38, 31, 
   22, 34, 60, 70, 73, 63, 46, 43, 28), ncol=2)

a <- matrix(c(22, 33, 64, 85, 92, 94, 59, 27, 30, 64, 60, 33, 31, 9,
   99, 67, 15, 5, 4, 30, 8, 37, 42, 27, 19, 69, 60, 73, 3, 5, 21,
   37, 52, 70, 74, 9, 13, 4, 17, 47), ncol=2)

xy <- rbind(cbind(1, p), cbind(0, a))

# extract predictor values for points
e <- extract(logo, xy[,2:3])

# combine with response 
v <- data.frame(cbind(pa=xy[,1], e))

#build a model, here with glm 
model <- glm(formula=pa~., data=v)

#predict to a raster
r1 <- predict(logo, model)

plot(r1)
points(p, bg='blue', pch=21)
points(a, bg='red', pch=21)

# logistic regression
model <- glm(formula=pa~., data=v, family="binomial")
r1log <- predict(logo, model, type="response")

# use a modified function to get the probability and standard error
# from the glm model. The values returned by "predict" are in a list,
# and this list needs to be transformed to a matrix

predfun <- function(model, data) {
  v <- predict(model, data, se.fit=TRUE)
  cbind(p=as.vector(v$fit), se=as.vector(v$se.fit))
}

r2 <- predict(logo, model, fun=predfun)

# principal components of a SpatRaster
# here using sampling to simulate an object too large
# to feed all its values to prcomp

sr <- values(sampleRegular(logo, 100))
pca <- prcomp(sr)

x <- predict(logo, pca)
plot(x)

# }
# NOT RUN {
library(pls)
model <- plsr(formula=pa~., data=v)
# this returns an array:
predict(model, v[1:5,])
# write a function to turn that into a matrix
pfun <- function(x, data) {
   y <- predict(x, data)
   d <- dim(y)
   dim(y) <- c(prod(d[1:2]), d[3])
   y
}

pp <- predict(logo, model, fun=pfun)

# Random Forest

library(randomForest)
rfmod <- randomForest(pa ~., data=v)

## note the additional argument "type='response'" that is 
## passed to predict.randomForest
r3 <- predict(logo, rfmod, type='response')

## get class membership probabilities
vv <- v
vv$pa <- as.factor(vv$pa)
rfmod2 <- randomForest(pa ~., data=vv)
r4 <- predict(logo, rfmod2, type='prob')

raster::spplot(r4)


# cforest (other Random Forest implementation) example with factors argument

library(party)
m <- cforest(pa~., control=cforest_unbiased(mtry=3), data=v)
# the second argument in party:::predict.RandomForest
# is "OOB", and not "newdata" or similar. We need to write a wrapper
# predict function to deal with this 	
predfun <- function(m, d, ...) predict(m, newdata=d, ...)

pc <- predict(logo, m, OOB=TRUE, fun=predfun)

# knn, using "app" instead of "predict"
library(class)
cl <- factor(c(rep(1, nrow(p)), rep(0, nrow(a))))
train <- extract(logo, rbind(p, a))
k <- app(logo, function(x) as.integer(as.character(knn(train, x, cl))))

## End(Not run)
# }