rangeSVM {maskRangeR}

Classify species ranges based on occurrence coordinates and SDM scores.

Description

rangeSVM() returns a tuned support vector machine (SVM) model that predicts species identity based on predictors that are solely spatial, based on occurrence coordinates, or a combination of spatial and environmental, based on both occurrence coordinates and environmental suitability values. Suitability values can be predicted with species distribution models (SDMs; a.k.a. ecological niche models).

Usage

rangeSVM(xy1, xy2, ..., sdm = NULL, nrep = 100, weight = FALSE, mc.cores = 1)

Arguments

xy1

Matrix or data frame of occurrence coordinates for species 1.

xy2

Matrix or data frame of occurrence coordinates for species 2.

...

Other matrices or data frames of occurrence coordinates for additional species.

sdm

Raster or RasterStack representing environmental suitability (can be predictions from SDMs). These must have the same extent as both species' occurrence points. Default is NULL.

nrep

Numeric for number of SVM tuning iterations. Default is 100.

weight

Boolean. If TRUE, species with fewer occurrence records are weighted higher in the SVM. Default is FALSE.

mc.cores

Number of cores to use for parallel processing. Default is 1.

Details

The tuning operation uses tune.svm() from the e1071 package, which performs 10-fold cross validation and selects the best model based on classification error. Ranges of the cost and gamma parameters are explored in the tuning exercise. The tuning function is iterated nrep times, and the parameter combination used most frequently across all iterations is used to build a final SVM model.

When sdm = NULL, the SVM is purely spatial, based only on the occurrence coordinates of each species. Otherwise, the SVM is fit with both a spatial predictor and any additional ones added as rasters. These extra predictors can be based on predictions from a species distribution model (SDM; a.k.a. ecological niche model), and in this case would represent environmental or climatic suitability, depending on the variables used in the SDM.

Value

The tuned SVM model.

Examples

r1.sdm <- raster::raster(raster::extent(c(-72, -64, 41, 50)), res = c(0.008333333, 0.008333333))
raster::values(r1.sdm) <- (1:raster::ncell(r1.sdm))^2
r2.sdm <- raster::raster(raster::extent(c(-72, -64, 41, 50)), res = c(0.008333333, 0.008333333))
raster::values(r2.sdm) <- (raster::ncell(r2.sdm):1)^2
r3.sdm <- raster::raster(raster::extent(c(-72, -64, 41, 50)), res = c(0.008333333, 0.008333333))
r3.sdm [1] <- 10
r3.sdm <- raster::distance(r3.sdm)
sp1.xy <- data.frame(dismo::randomPoints(r1.sdm, 15, prob = TRUE))
colnames(sp1.xy) <- c("longitude", "latitude")
sp2.xy <- data.frame(dismo::randomPoints(r2.sdm, 15, prob = TRUE))
colnames(sp2.xy) <- c("longitude", "latitude")
sp3.xy <- data.frame(dismo::randomPoints(r3.sdm, 15, prob = TRUE))
colnames(sp3.xy) <- c("longitude", "latitude")
# Spatial SVMs (this can take about a minute to run)
svm.SP <- rangeSVM(sp1.xy, sp2.xy, sp3.xy, nrep=5) # more reps are recommended