Using H2oBinomialModel to predict outcomes in a raster stack - r

Question

I use machine learning algorithms for species distribution modeling. This involves fitting a model in R and applying the model to a raster stack of environmental predictors using the predict() function. This is relatively straightforward for functions such as randomForest, glm, and maxent. However, I would like to use the best model selected using the h2o.automl function in the package h2o to predict occurrence probabilities across a stack of environmental predictors. Now, it is possible to convert raster stacks into data frames and then use these for h2o prediction. However, this tends to eat up a lot of RAM for large raster stacks. Below is an example using some publicly available low-resolution data.
library(sp)
library(raster)
library(maptools)
library(rgdal)
library(dismo)
library(h2o)
library(sp)
library(h2o)
bioclim.data <- getData(name = "worldclim",
var = "bio",
res = 2.5)
obs.data <- read.csv("https://raw.githubusercontent.com/jcoliver/learn-r/gh-pages/data/Carnegiea-gigantea-GBIF.csv")
obs.data <- obs.data[!is.na(obs.data$latitude), ]
obs.data <- obs.data[, c("longitude", "latitude")]
obs.sp<-SpatialPoints(obs.data, proj4string = CRS("+proj=longlat +datum=WGS84 +no_defs"))
# Determine geographic extent of our data
max.lat <- ceiling(max(obs.data$latitude))
min.lat <- floor(min(obs.data$latitude))
max.lon <- ceiling(max(obs.data$longitude))
min.lon <- floor(min(obs.data$longitude))
geographic.extent <- extent(x = c(min.lon, max.lon, min.lat, max.lat))
bioclim.data <- crop(x = bioclim.data, y = geographic.extent)
background<-randomPoints(mask = bioclim.data,
n = nrow(obs.data),
ext = geographic.extent,
extf = 1.25)
colnames(background)<-c("longitude", "latitude")
background.sp<-SpatialPoints(background, proj4string = CRS("+proj=longlat +datum=WGS84 +no_defs"))
presence<-data.frame(Presence = rep(1, times = nrow(obs.data)), extract(bioclim.data, obs.sp))
absence<-data.frame(Presence = rep(0, times = nrow(obs.data)), extract(bioclim.data, background.sp))
df<-rbind(presence, absence)
df$Presence<-as.factor(df$Presence)
h2o.init(max_mem_size = "4g")
set.seed(1234)
df_h2o <- as.h2o(df)
splits <- h2o.splitFrame(df_h2o, c(0.7, 0.15), seed = 1234)
train <- h2o.assign(splits[[1]], "train")
valid <- h2o.assign(splits[[2]], "valid")
test <- h2o.assign(splits[[3]], "test")
y <- "Presence"
x <- setdiff(names(df_h2o), y)
aml <- h2o.automl(x = x,
y= y,
training_frame = train,
leaderboard_frame = valid,
max_runtime_secs = 60) #default 1 hour 3600 secs, more time, more accurate
pred<-as.data.frame(predict(aml, test)) # this works. It reports predicted probabilities of presence vs. absence.
pred_rstr<-predict(aml, bioclim.data) # this doesn't work. Must be an H2oFrame
biodf<-as.data.frame(bioclim.data) # this forces the raster stack into a dataframe. This works, but only for low resolution rasters.
bio_h2o<-as.h2o(biodf)
pred2<-as.data.frame(predict(aml, bio_h2o))
pred_rstr<-bioclim.data[[1]] # pull one of the rasters from the stack to serve as a template
values(pred_rstr)<-as.vector(pred2$p1) # set raster values to predicted probabilities
plot(pred_rstr)