NAs in rasters and randomForest::predict() - r

Question

New here, please let me know if you need more info.
My goal: I am using Rehfeldt climate data and eBird presence/absence data to produce niche models using Random Forest models.
My problem: I want to predict niche models for the entirety of North America. The Rehfeldt climate rasters have data values for every cell on the continent, but these are surrounded by NAs in the "ocean cells". See the plot here, where I have colored the NAs dark green. randomForest::predict() does not run if the independent dataset contains NAs. Thus, I want to crop my climate rasters (or set a working extent?) so that the predict() function only operates over the cells which contain data.
Troubleshooting:
I've run the Random Forest model using a smaller extent which does not include the "NA oceans" of the rasters and the model runs just fine. So, I know the NAs are the problem. However, I don't want to predict my niche models for just a rectangular chunk of North America.
I used flowla's approach here for cropping and masking rasters using a polygon shapefile for North America. I hoped that this would remove the NAs but it doesn't. Is there something similar I can do to remove the NAs?
I've done some reading but can't figure out a way to adjust the Random Forest code itself so that predict() ignores NAs. This post looks relevant but I'm not sure whether it helps in my case.
Data
My rasters, the input presence/absence text file, and code for additional functions are here. Use with the main code below for a reproducible example.
Code
require(sp)
require(rgdal)
require(raster)
library(maptools)
library(mapproj)
library(dismo)
library(maps)
library(proj4)
data(stateMapEnv)
# This source code has all of the functions necessary for running the Random Forest models, as well as the code for the function detecting multi-collinearity
source("Functions.R")
# Read in Rehfeldt climate rasters
# these rasters were converted to .img and given WGS 84 projection in ArcGIS
d100 <- raster("d100.img")
dd0 <- raster("dd0.img")
dd5 <- raster("dd5.img")
fday <- raster("fday.img")
ffp <- raster("ffp.img")
gsdd5 <- raster("gsdd5.img")
gsp <- raster("gsp.img")
map <- raster("map.img")
mat <- raster("mat_tenths.img")
mmax <- raster("mmax_tenths.img")
mmin <- raster("mmin_tenths.img")
mmindd0 <- raster("mmindd0.img")
mtcm <- raster("mtcm_tenths.img")
mtwm <- raster("mtwm_tenths.img")
sday <- raster("sday.img")
smrpb <- raster("smrpb.img")
# add separate raster files into one big raster, with each file being a different layer.
rehfeldt <- addLayer(d100, dd0, dd5, fday, ffp, gsdd5, gsp, map, mat, mmax, mmin, mmindd0, mtcm, mtwm, sday, smrpb)
# plot some rasters to make sure everything worked
plot(d100)
plot(rehfeldt)
# read in presence/absence data
LAZB.INBUtemp <- read.table("LAZB.INBU.txt", header=T, sep = "\t")
colnames(LAZB.INBUtemp) <- c("Lat", "Long", "LAZB", "INBU")
LAZB.INBUtemp <- LAZB.INBUtemp[c(2,1,3,4)]
LAZB.INBU <- LAZB.INBUtemp
latpr <- (LAZB.INBU$Lat)
lonpr <- (LAZB.INBU$Long)
sites <- SpatialPoints(cbind(lonpr, latpr))
LAZB.INBU.spatial <- SpatialPointsDataFrame(sites, LAZB.INBU, match.ID=TRUE)
# The below function extracts raster values for each of the different layers for each of the eBird locations
pred <- raster::extract(rehfeldt, LAZB.INBU.spatial)
LAZB.INBU.spatial#data = data.frame(LAZB.INBU.spatial#data, pred)
LAZB.INBU.spatial#data <- na.omit(LAZB.INBU.spatial#data)
# ITERATIVE TEST FOR MULTI-COLINEARITY
# Determines which variables show multicolinearity
cl <- MultiColinear(LAZB.INBU.spatial#data[,7:ncol(LAZB.INBU.spatial#data)], p=0.05)
xdata <- LAZB.INBU.spatial#data[,7:ncol(LAZB.INBU.spatial#data)]
for(l in cl) {
cl.test <- xdata[,-which(names(xdata)==l)]
print(paste("REMOVE VARIABLE", l, sep=": "))
MultiColinear(cl.test, p=0.05)
}
# REMOVE MULTI-COLINEAR VARIABLES
for(l in cl) { LAZB.INBU.spatial#data <- LAZB.INBU.spatial#data[,-which(names(LAZB.INBU.spatial#data)==l)] }
################################################################################################
# FOR LAZB
# RANDOM FOREST MODEL AND RASTER PREDICTION
require(randomForest)
# NUMBER OF BOOTSTRAP REPLICATES
b=1001
# CREATE X,Y DATA
# use column 3 for LAZB and 4 for INBU
ydata <- as.factor(LAZB.INBU.spatial#data[,3])
xdata <- LAZB.INBU.spatial#data[,7:ncol(LAZB.INBU.spatial#data)]
# PERCENT OF PRESENCE OBSERVATIONS
( dim(LAZB.INBU.spatial[LAZB.INBU.spatial$LAZB == 1, ])[1] / dim(LAZB.INBU.spatial)[1] ) * 100
# RUN RANDOM FORESTS MODEL SELECTION FUNCTION
# This model is using the model improvement ratio to select a final model.
pdf(file = "LAZB Random Forest Model Rehfeldt.pdf")
( rf.model <- rf.modelSel(x=xdata, y=ydata, imp.scale="mir", ntree=b) )
dev.off()
# RUN RANDOM FORESTS CLASS BALANCE BASED ON SELECTED VARIABLES
# This code would help in the case of imbalanced sample
mdata <- data.frame(y=ydata, xdata[,rf.model$SELVARS])
rf.BalModel <- rfClassBalance(mdata[,1], mdata[,2:ncol(mdata)], "y", ntree=b)
# CREATE NEW XDATA BASED ON SELECTED MODEL AND RUN FINAL RF MODEL
sel.vars <- rf.model$PARAMETERS[[3]]
rf.data <- data.frame(y=ydata, xdata[,sel.vars])
write.table(rf.data, "rf.data.txt", sep = ",", row.names = F)
# This the code given to me; takes forever to run for my dataset (I haven't tried to let it finish)
# ( rf.final <- randomForest(y ~ ., data=rf.data, ntree=b, importance=TRUE, norm.votes=TRUE, proximity=TRUE) )
# I use this form because it's a lot faster
( rf.final <- randomForest(x = rf.data[2:6], y = rf.data$y, ntree=1000, importance=TRUE, norm.votes=TRUE, proximity=F) )
################################################################################################
# MODEL VALIDATION
# PREDICT TO VALIDATION DATA
# Determines the percent correctly classified
rf.pred <- predict(rf.final, rf.data[,2:ncol(rf.data)], type="response")
rf.prob <- as.data.frame(predict(rf.final, rf.data[,2:ncol(rf.data)], type="prob"))
ObsPred <- data.frame(cbind(Observed=as.numeric(as.character(ydata)),
PRED=as.numeric(as.character(rf.pred)), Prob1=rf.prob[,2],
Prob0=rf.prob[,1]) )
op <- (ObsPred$Observed == ObsPred$PRED)
( pcc <- (length(op[op == "TRUE"]) / length(op))*100 )
# PREDICT MODEL PROBABILITIES RASTER
# The first line of code says what directory I'm working, and then what folder in that directory has the raster files that I'm using to predict the range
# The second line defines the x variable, wich is my final Random Forest model
rpath=paste('~/YOURPATH', "example", sep="/")
xvars <- stack(paste(rpath, paste(rownames(rf.final$importance), "img", sep="."), sep="/"))
tr <- blockSize(xvars)
s <- writeStart(xvars[[1]], filename=paste('~/YOURPATH', "prob_LAZB_Rehfeldt.img", sep="/"), overwrite=TRUE)
for (i in 1:tr$n) {
v <- getValuesBlock(xvars, row=tr$row[i], nrows=tr$nrows[i])
v <- as.data.frame(v)
rf.pred <- predict(rf.final, v, type="prob")[,2]
writeValues(s, rf.pred, tr$row[i])
}
s <- writeStop(s)
prob_LAZB <- raster("prob_LAZB_Rehfeldt.img")
# Write range prediction raster to .pdf
pdf(file="LAZB_range_pred.pdf")
plot(prob_LAZB)
map("state", add = TRUE)
dev.off()
Thanks!!

Answer 1

Did you try setting 'na.action` in your call to RF? The option is clearly labelled in the randomForest R manual. Your call to RF would look like this:
rf.final <- randomForest(x = rf.data[2:6], y = rf.data$y, ntree=1000, importance=TRUE, norm.votes=TRUE, proximity=F, na.action = omit)
This will tell RF to omit rows where NA exists, thereby throwing out those observations. This is not necessarily the best approach, but it might be handy in your situation.
Option 2: rfImpute or na.roughfix: This will fill in your NAs so that you can go ahead with your prediction. Watch out as this can give you spurious predictions wherever the NAs are being imputed/"fixed".
Option 3: Start with Option 2, and after you get your prediction, bring your raster into your GIS/Image processing software of choice, and mask out the areas you don't want. In your case, masking out water bodies would be pretty simple.