In sgvignali/SDMtune: Species Distribution Model Selection

knitr::opts_chunk$set(collapse = TRUE,
                      comment = "#>",
                      fig.align = "center")
options(knitr.table.format = "html")

library(SDMtune)
files <- list.files(path = file.path(system.file(package = "dismo"), "ex"), 
                    pattern = "grd", 
                    full.names = TRUE)
predictors <- terra::rast(files)

Intro

All the previous articles are based on presence only methods, in this article you will learn how to train a presence absence model. The following examples are based on the Artificial Neural Networks method [@Venables2002], but you can adapt the code for any of the other supported methods.

Prepare the data for the analysis

We use the first 8 environmental variables and the same virtualSp() dataset selecting the absence instead of the background locations.

p_coords <- virtualSp$presence
a_coords <- virtualSp$absence

data <- prepareSWD(species = "Virtual species", 
                   p = p_coords, 
                   a = a_coords, 
                   env = predictors[[1:8]])

data

p_coords <- virtualSp$presence
a_coords <- virtualSp$absence

data <- prepareSWD(species = "Virtual species", 
                   p = p_coords, 
                   a = a_coords, 
                   env = predictors[[1:8]],
                   verbose = FALSE)

data

There are r sum(data@pa == 1) presence and r sum(data@pa == 0) absence locations.

For the model evaluation we will create a training and testing datasets, holding apart 20% of the data:

library(zeallot)
c(train, test) %<-% trainValTest(data, 
                                 test = 0.2, 
                                 seed = 25)

At this point we have r nrow(train@data) training and r nrow(test@data) testing locations. We create a 4-folds partition to run cross validation:

folds <- randomFolds(train, 
                     k = 4, 
                     seed = 25)

Train the model

We first train the model with default settings and 10 neurons:

set.seed(25)
model <- train("ANN", 
               data = train, 
               size = 10, 
               folds = folds)
model

Let's check the training and testing AUC:

auc(model)
auc(model, test = TRUE)

Tune model hyperparameters

To check which hyperparameters can be tuned we use the function getTunableArgs() function:

getTunableArgs(model)

We use the function optimizeModel() to tune the hyperparameters:

h <- list(size = 10:50, 
          decay = c(0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5),
          maxit = c(50, 100, 300, 500))

om <- optimizeModel(model, 
                    hypers = h, 
                    metric = "auc", 
                    seed = 25)

h <- list(size = 10:50, 
          decay = c(0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5),
          maxit = c(50, 100, 300, 500))

om <- optimizeModel(model, 
                    hypers = h, 
                    metric = "auc", 
                    seed = 25,
                    interactive = FALSE,
                    progress = FALSE)

The best model is:

best_model <- om@models[[1]]
om@results[1, ]

best_model <- om@models[[1]]
kableExtra::kable(om@results[1, ]) |> 
  kableExtra::kable_styling(bootstrap_options = c("striped", "hover"),
                            position = "center",
                            full_width = FALSE)

The validation AUC increased from r auc(model, test = TRUE) of the default models to r om@results[1, 6] of the optimized one.

Evaluate the final model

We now train a model with the same configuration as found by the function optimizeModel() without cross validation (i.e. using all presence and background locations) and we evaluate it using the held apart testing dataset:

set.seed(25)
final_model <- combineCV(om@models[[1]])

plotROC(final_model, 
        test = test)

Conclusion

In this tutorial you have learned a general way to train, evaluate and tune model using Artificial Neural Network, but you can apply the same workflow to other methods.