This is the core function of the package in which a set of environmental variables is reduced in a stepwise fashion in order to avoid overfitting the model to the occurrence records. This can be done for a range of regularization multipliers. The best performing model, based on AICc values (Akaike, 1974) or AUC.Test values (Fielding and Bell, 1997), identifies then the most-important uncorrelated environmental variables along with the optimal regularization multiplier.
VariableSelection(maxent, outdir, gridfolder, occurrencelocations,
backgroundlocations, additionalargs, contributionthreshold,
correlationthreshold, betamultiplier)String specifying the filepath to the maxent.jar file (download from here: https://www.cs.princeton.edu/~schapire/maxent/). The package was tested with maxent.jar version 3.3.3k.
String specifying the path to the output directory to which all the result files will be written.Please don't put important files in this folder as all files but the output files of the VariableSelection function will be deleted from this folder.
String specifying the path to the directory that holds all the ASCII grids (in ESRI's .asc format) of environmental variables. All variables must have the same extent and resolution.
String specifying the filepath to the csv file with occurrence records. Please find the exact specifications of the SWD file format in the details section below.
String specifying the filepath to the csv file with background/pseudoabsence data. Please find the exact specifications of the SWD file format in the details section below.
String specifying additional maxent arguments. Please see in the details section below.
Vector of beta (regularization
multipliers) (positive numerical values). The smaller this value, the
more closely will the projected distribution fit to the training data
set. Overfitted models are poorly transferable to novel environments
and, thus, not appropriate to project distribution changes under
environmental change. The model performance will be compared between
models created with the beta values given in this betamultiplier
vector. Thus, providing a range of beta values from 1 (the default in
Maxent) to 15 or so, will help you to spot the optimal beta multiplier
for your specific model.
Numerical value (between 0 and 1) that sets the threshold of Pearson's correlation coefficient above which environmental variables are regarded to be correlated (based on values at all background locations). Of the correlated variables, only the variable with the highest contribution score will be kept, all other correlated variables will be excluded from the Maxent model. Correlated variables should be removed because they may reflect the same environmental conditions, and can lead to overly complex or overpredicted models. Also, models comiled with correlated variables might give wrong predictions in scenarios where the correlations between the variables differ.
Numerical value (between 0 and 100) that sets the threshold of model contributions below which environmental variables are excluded from the Maxent model. Model contributions reflect the importance of environmental variables in limiting the distribution of the target species.
The following result files are saved in the directory specified with the outdir
argument.
A table listing the performance indicators of all created Maxent models
ModelUnique model number
betamultiplierMaxent's regularization multiplier
variablesNumber of environmental variables
samplesNumber of occurrence sites
parametersNumber of parameters estimated from the model
loglikelihoodlog likelihood value
AICAkaike Information Criterion
AICcsample size corrected AIC
BICBayesian information criterion
AUC.TestArea under the receiver operating characteristic fro test data
AUC.TrainArea under the receiver operating characteristic fro training data
AUC.DiffDifference between AUC.Test and AUC.Train
A figure showing the AICc values of all models, which are ordered along the x-axis based on the applied beta-multiplier. The number of environmental variables included in each model is coded by dot color and size. The model with highest AUC.Test value is marked in red.
A figure showing the AICc values of all models, which are ordered along the x-axis based on the applied beta-multiplier. The number of environmental variables included in each model is coded by dot color and size. The model with highest minimum AICc value is marked in red.
A figure showing the AUC.Test values of all models, which are ordered along the x-axis based on the applied beta-multiplier. The number of environmental variables included in each model is coded by dot color and size. The model with highest AUC.Test value is marked in red.
A figure showing the AUC.Test values of all models, which are ordered along the x-axis based on the applied beta-multiplier. The number of environmental variables included in each model is coded by dot color and size. The model with highest minimum AICc value is marked in red.
Subset of the table ModelPerformance.txt, which shows only the
model with the highest AUC.Test value.
Subset of the table
ModelPerformance.txt, which shows only the model with the
lowest AICc value.
Table listing model contributions for
and correlations between each of the environmental variables for all
created Maxent models. The numbers of the models refer to the unique
model numbers in the table ModelPerformance.txt (see above).
The following entries describe the content row-wise (not column-wise)
TestEither 'Contributions' or 'Correlation. Informs if the numbers for each of the environmental variables refers to model contribution coefficients or to correlation coefficients.
ModelThe unique model number (the same unique model
number as in ModelPerformance.txt.)
betamultiplierThe (regularization multipliers) used to compile the respective model.
X'X' stands here for the name of an environmental
variable. The Test row above informs whether the values in this row refer to the model
contribution of this environmental variable or to its coefficient
of correlation with another environmental variable. The variable
to which it is compared is recognizable by a correlation
coefficient of 1. If this environmental variable was excluded from
the model, the value in this row is 'NA', which stands for 'Not
Available'.'
Subset of
VariableSelectionProcess.txt that shows only those models which
lead directly to the model with the highest AUC.Test value.
Subset of
VariableSelectionProcess.txt that shows only those models which
lead directly to the model with the lowest AICc value.
Depending on the number of environmental variables and the range of different betamultipliers you want to test, variable selection can take several hours so that you might want to run the analysis over night.
For further details on the model selection process and the variable settings, please have a look at the vignette that comes with this package.
Akaike H (1974) A new look at the statistical model identification IEEE Transactions on Automatic Control 19:6 716--723.
Fielding AH and Bell JF (1997) A review of methods for the assessment of prediction errors in conservation presence/absence models Environmental Conservation 24:1 38--49.
# NOT RUN {
# Please find a workflow tutorial in the vignette of this package. It
# will guide you through the settings and usage of the
# 'VariableSelection' function, the core function of this package.
# }
# NOT RUN {
VariableSelection(
maxent="C:/.../maxent.jar",
outdir="OutputDirectory",
gridfolder="BioORACLEVariables",
occurrencelocations=system.file("extdata", "Occurrencedata.csv", package="MaxentVariableSelection"),
backgroundlocations=system.file("extdata", "Backgrounddata.csv", package="MaxentVariableSelection"),
additionalargs="nolinear noquadratic noproduct nothreshold noautofeature",
contributionthreshold=5,
correlationthreshold=0.9,
betamultiplier=seq(2,6,0.5)
)
# }
Run the code above in your browser using DataLab