Train

Train a model using one of the following methods: Artificial Neural Networks, Boosted Regression Trees, Maxent, Maxnet or Random Forest.

train(method, data, folds = NULL, verbose = TRUE, ...)

Arguments

method	character or character vector. Method used to train the model, possible values are "ANN", "BRT", "Maxent", "Maxnet" or "RF", see details.
data	SWD object with presence and absence/background locations.
folds	list. Output of the function randomFolds or folds object created with other packages, see details, default is NULL.
verbose	logical, if TRUE shows a progress bar during cross validation, default is TRUE.
...	Arguments passed to the relative method, see details.

Value

An SDMmodel or SDMmodelCV or a list of model objects.

Details

• For the ANN method possible arguments are (for more details see nnet):

  • size: integer. Number of the units in the hidden layer.

  • decay numeric. Weight decay, default is 0.

  • rang numeric. Initial random weights, default is 0.7.

  • maxit integer. Maximum number of iterations, default is 100.

• For the BRT method possible arguments are (for more details see gbm):

  • distribution: character. Name of the distribution to use, default is "bernoulli".

  • n.trees: integer. Maximum number of tree to grow, default is 100.

  • interaction.depth: integer. Maximum depth of each tree, default is 1.

  • shrinkage: numeric. The shrinkage parameter, default is 0.1.

  • bag.fraction: numeric. Random fraction of data used in the tree expansion, default is 0.5.

• For the RF method the model is trained as classification. Possible arguments are (for more details see randomForest):

  • mtry: integer. Number of variable randomly sampled at each split, default is floor(sqrt(number of variables)).

  • ntree: integer. Number of tree to grow, default is 500.

  • nodesize: integer. Minimum size of terminal nodes, default is 1.

• Maxent models are trained using the arguments "removeduplicates=false" and "addsamplestobackground=false". Use the function thinData to remove duplicates and the function addSamplesToBg to add presence locations to background locations. For the Maxent method, possible arguments are:

  • reg: numeric. The value of the regularization multiplier, default is 1.

  • fc: character. The value of the feature classes, possible values are combinations of "l", "q", "p", "h" and "t", default is "lqph".

  • iter: numeric. Number of iterations used by the MaxEnt algorithm, default is 500.

• Maxnet models are trained using the argument "addsamplestobackground = FALSE", use the function addSamplesToBg to add presence locations to background locations. For the Maxnet method, possible arguments are (for more details see maxnet):

  • reg: numeric. The value of the regularization intensity, default is 1.

  • fc: character. The value of the feature classes, possible values are combinations of "l", "q", "p", "h" and "t", default is "lqph".

The folds argument accepts also objects created with other packages: ENMeval or blockCV. In this case the function converts internally the folds into a format valid for SDMtune.

When multiple methods are given as method argument, the function returns a named list of model object, with the name corresponding to the used method, see examples.

References

Venables, W. N. & Ripley, B. D. (2002) Modern Applied Statistics with S. Fourth Edition. Springer, New York. ISBN 0-387-95457-0.

Brandon Greenwell, Bradley Boehmke, Jay Cunningham and GBM Developers (2019). gbm: Generalized Boosted Regression Models. https://CRAN.R-project.org/package=gbm.

A. Liaw and M. Wiener (2002). Classification and Regression by randomForest. R News 2(3), 18--22.

Hijmans, Robert J., Steven Phillips, John Leathwick, and Jane Elith. 2017. dismo: Species Distribution Modeling. https://cran.r-project.org/package=dismo.

Steven Phillips (2017). maxnet: Fitting 'Maxent' Species Distribution Models with 'glmnet'. https://CRAN.R-project.org/package=maxnet.

Muscarella, R., Galante, P.J., Soley-Guardia, M., Boria, R.A., Kass, J., Uriarte, M. and R.P. Anderson (2014). ENMeval: An R package for conducting spatially independent evaluations and estimating optimal model complexity for ecological niche models. Methods in Ecology and Evolution.

Roozbeh Valavi, Jane Elith, José Lahoz-Monfort and Gurutzeta Guillera-Arroita (2018). blockCV: Spatial and environmental blocking for k-fold cross-validation. https://github.com/rvalavi/blockCV.

See also

randomFolds.

Author

Sergio Vignali

Examples

# \donttest{
# Acquire environmental variables
files <- list.files(path = file.path(system.file(package = "dismo"), "ex"),
                    pattern = "grd", full.names = TRUE)
predictors <- raster::stack(files)

# Prepare presence and background locations
p_coords <- virtualSp$presence
bg_coords <- virtualSp$background

# Create SWD object
data <- prepareSWD(species = "Virtual species", p = p_coords, a = bg_coords,
                   env = predictors, categorical = "biome")
#> Extracting predictor information for presence locations...
#> Extracting predictor information for absence/background locations...

## Train a Maxent model
# The next line checks if Maxent is correctly configured but you don't need
# to run it in your script
if (dismo::maxent(silent = TRUE)) {
model <- train(method = "Maxent", data = data, fc = "l", reg = 1.5,
               iter = 700)

# Add samples to background. This should be done preparing the data before
# training the model without using
data <- addSamplesToBg(data)
model <- train("Maxent", data = data)
}
#> This is MaxEnt version 3.4.1 
#> This is MaxEnt version 3.4.1 

## Train a Maxnet model
model <- train(method = "Maxnet", data = data, fc = "lq", reg = 1.5)

## Cross Validation
# Create 4 random folds splitting only the presence data
folds <- randomFolds(data, k = 4, only_presence = TRUE)
model <- train(method = "Maxnet", data = data, fc = "l", reg = 0.8,
               folds = folds)

if (FALSE) {
# Run only if you have the package ENMeval installed
## Block partition using the ENMeval package
require(ENMeval)
block_folds <- get.block(occ = data@coords[data@pa == 1, ],
                         bg.coords = data@coords[data@pa == 0, ])
model <- train(method = "Maxnet", data = data, fc = "l", reg = 0.8,
               folds = block_folds)

## Checkerboard1 partition using the ENMeval package
cb_folds <- get.checkerboard1(occ = data@coords[data@pa == 1, ],
                              env = predictors,
                              bg.coords = data@coords[data@pa == 0, ],
                              aggregation.factor = 4)
model <- train(method = "Maxnet", data = data, fc = "l", reg = 0.8,
               folds = cb_folds)

## Environmental block using the blockCV package
# Run only if you have the package blockCV
require(blockCV)
# Create spatial points data frame
library(raster)
sp_df <- SpatialPointsDataFrame(data@coords, data = as.data.frame(data@pa),
                                proj4string = crs(predictors))
e_folds <- envBlock(rasterLayer = predictors,
                    speciesData = sp_df,
                    species = "data@pa",
                    k = 4,
                    standardization = "standard",
                    rasterBlock = FALSE)
model <- train(method = "Maxnet", data = data, fc = "l", reg = 0.8,
               folds = e_folds)
}

## Train presence absence models
# Prepare presence and absence locations
p_coords <- virtualSp$presence
a_coords <- virtualSp$absence
# Create SWD object
data <- prepareSWD(species = "Virtual species", p = p_coords, a = a_coords,
                   env = predictors[[1:5]])
#> Extracting predictor information for presence locations...
#> Extracting predictor information for absence/background locations...

## Train an Artificial Neural Network model
model <- train("ANN", data = data, size = 10)

## Train a Random Forest model
model <- train("RF", data = data, ntree = 300)

## Train a Boosted Regression Tree model
model <- train("BRT", data = data, n.trees = 300, shrinkage = 0.001)

## Multiple methods trained together with default arguments
output <- train(method = c("ANN", "BRT", "RF"), data = data, size = 10)
output$ANN
#> Object of class SDMmodel 
#> Method: ANN 
#> 
#> Species: Virtual species 
#> Presence locations: 400 
#> Absence locations: 300 
#> 
#> Model configurations:
#> --------------------
#> size: 10
#> decay: 0
#> rang: 0.7
#> maxit: 100
#> 
#> Variables:
#> ---------
#> Continuous: bio1 bio12 bio16 bio17 bio5 
#> Categorical: NA
output$BRT
#> Object of class SDMmodel 
#> Method: BRT 
#> 
#> Species: Virtual species 
#> Presence locations: 400 
#> Absence locations: 300 
#> 
#> Model configurations:
#> --------------------
#> distribution: bernoulli
#> n.trees: 100
#> interaction.depth: 1
#> shrinkage: 0.1
#> bag.fraction: 0.5
#> 
#> Variables:
#> ---------
#> Continuous: bio1 bio12 bio16 bio17 bio5 
#> Categorical: NA
output$RF
#> Object of class SDMmodel 
#> Method: RF 
#> 
#> Species: Virtual species 
#> Presence locations: 400 
#> Absence locations: 300 
#> 
#> Model configurations:
#> --------------------
#> mtry: 2
#> ntree: 500
#> nodesize: 1
#> 
#> Variables:
#> ---------
#> Continuous: bio1 bio12 bio16 bio17 bio5 
#> Categorical: NA

## Multiple methods trained together passing extra arguments
output <- train(method = c("ANN", "BRT", "RF"), data = data, size = 10,
                ntree = 300, n.trees = 300, shrinkage = 0.001)
output
#> $ANN
#> Object of class SDMmodel 
#> Method: ANN 
#> 
#> Species: Virtual species 
#> Presence locations: 400 
#> Absence locations: 300 
#> 
#> Model configurations:
#> --------------------
#> size: 10
#> decay: 0
#> rang: 0.7
#> maxit: 100
#> 
#> Variables:
#> ---------
#> Continuous: bio1 bio12 bio16 bio17 bio5 
#> Categorical: NA
#> $BRT
#> Object of class SDMmodel 
#> Method: BRT 
#> 
#> Species: Virtual species 
#> Presence locations: 400 
#> Absence locations: 300 
#> 
#> Model configurations:
#> --------------------
#> distribution: bernoulli
#> n.trees: 300
#> interaction.depth: 1
#> shrinkage: 0.001
#> bag.fraction: 0.5
#> 
#> Variables:
#> ---------
#> Continuous: bio1 bio12 bio16 bio17 bio5 
#> Categorical: NA
#> $RF
#> Object of class SDMmodel 
#> Method: RF 
#> 
#> Species: Virtual species 
#> Presence locations: 400 
#> Absence locations: 300 
#> 
#> Model configurations:
#> --------------------
#> mtry: 2
#> ntree: 300
#> nodesize: 1
#> 
#> Variables:
#> ---------
#> Continuous: bio1 bio12 bio16 bio17 bio5 
#> Categorical: NA
# }