ensemble: Suitability mapping based on ensembles of modelling...

ensemble.calibrate.modelsR Documentation

Suitability mapping based on ensembles of modelling algorithms: calibration of models and weights

Description

The basic function ensemble.calibrate.models allows to evaluate different algorithms for (species) suitability modelling, including maximum entropy (MAXENT), boosted regression trees, random forests, generalized linear models (including stepwise selection of explanatory variables), generalized additive models (including stepwise selection of explanatory variables), multivariate adaptive regression splines, regression trees, artificial neural networks, flexible discriminant analysis, support vector machines, the BIOCLIM algorithm, the DOMAIN algorithm and the Mahalanobis algorithm. These sets of functions were developed in parallel with the biomod2 package, especially for inclusion of the maximum entropy algorithm, but also to allow for a more direct integration with the BiodiversityR package, more direct handling of model formulae and greater focus on mapping. Researchers and students of species distribution are strongly encouraged to familiarize themselves with all the options of the BIOMOD and dismo packages.

Usage

ensemble.calibrate.models(x = NULL, p = NULL, 
    a = NULL, an = 1000, excludep = FALSE, target.groups = FALSE,
    k = 0, pt = NULL, at = NULL, SSB.reduce = FALSE, CIRCLES.d = 250000, 
    TrainData = NULL, TestData = NULL, 
    VIF = FALSE, COR = FALSE,
    SINK = FALSE, PLOTS = FALSE, CATCH.OFF = FALSE, 
    threshold.method = "spec_sens", threshold.sensitivity = 0.9, 
    threshold.PresenceAbsence = FALSE,
    evaluations.keep = FALSE, 
    models.list = NULL, models.keep = FALSE, 
    models.save = FALSE, species.name = "Species001",
    ENSEMBLE.tune = FALSE, 
    ENSEMBLE.best = 0, ENSEMBLE.min = 0.7, ENSEMBLE.exponent = 1, 
    ENSEMBLE.weight.min = 0.05,
    input.weights = NULL, 
    MAXENT = 1, MAXNET = 1, MAXLIKE = 1, GBM = 1, GBMSTEP = 1, RF = 1, CF = 1,
    GLM = 1, GLMSTEP = 1, GAM = 1, GAMSTEP = 1, MGCV = 1, MGCVFIX = 0, 
    EARTH = 1, RPART = 1, NNET = 1, FDA = 1, SVM = 1 , SVME = 1, GLMNET = 1,
    BIOCLIM.O = 0, BIOCLIM = 1, DOMAIN = 1, MAHAL = 1, MAHAL01 = 1,
    PROBIT = FALSE,
    Yweights = "BIOMOD", 
    layer.drops = NULL, factors = NULL, dummy.vars = NULL,
    formulae.defaults = TRUE, maxit = 100,
    MAXENT.a = NULL, MAXENT.an = 10000, 
    MAXENT.path = paste(getwd(), "/models/maxent_", species.name,  sep=""), 
    MAXNET.classes = "default", MAXNET.clamp = FALSE, MAXNET.type = "cloglog",
    MAXLIKE.formula = NULL, MAXLIKE.method = "BFGS",
    GBM.formula = NULL, GBM.n.trees = 2001, 
    GBMSTEP.gbm.x = 2:(ncol(TrainData.orig)), GBMSTEP.tree.complexity = 5, 
    GBMSTEP.learning.rate = 0.005, GBMSTEP.bag.fraction = 0.5, 
    GBMSTEP.step.size = 100, 
    RF.formula = NULL, RF.ntree = 751, RF.mtry = floor(sqrt(ncol(TrainData.vars))), 
    CF.formula = NULL, CF.ntree = 751, CF.mtry = floor(sqrt(ncol(TrainData.vars))),
    GLM.formula = NULL, GLM.family = binomial(link = "logit"), 
    GLMSTEP.steps = 1000, STEP.formula = NULL, GLMSTEP.scope = NULL, 
    GLMSTEP.k = 2, 
    GAM.formula = NULL, GAM.family = binomial(link = "logit"), 
    GAMSTEP.steps = 1000, GAMSTEP.scope = NULL, GAMSTEP.pos = 1, 
    MGCV.formula = NULL, MGCV.select = FALSE, 
    MGCVFIX.formula = NULL, 
    EARTH.formula = NULL, 
    EARTH.glm = list(family = binomial(link = "logit"), maxit = maxit), 
    RPART.formula = NULL, RPART.xval = 50, 
    NNET.formula = NULL, NNET.size = 8, NNET.decay = 0.01, 
    FDA.formula = NULL, 
    SVM.formula = NULL, 
    SVME.formula = NULL, 
    GLMNET.nlambda = 100, GLMNET.class = FALSE,
    BIOCLIM.O.fraction = 0.9,
    MAHAL.shape = 1)

ensemble.calibrate.weights(x = NULL, p = NULL, TrainTestData=NULL,
    a = NULL, an = 1000, 
    get.block = FALSE, block.default = TRUE, get.subblocks = FALSE,
    SSB.reduce = FALSE, CIRCLES.d = 100000, 
    excludep = FALSE, target.groups = FALSE,
    k = 4, 
    VIF = FALSE, COR = FALSE,
    SINK = FALSE, PLOTS = FALSE, CATCH.OFF = FALSE, 
    data.keep = FALSE,
    species.name = "Species001",
    threshold.method = "spec_sens", threshold.sensitivity = 0.9, 
    threshold.PresenceAbsence = FALSE,
    ENSEMBLE.tune = FALSE, 
    ENSEMBLE.best = 0, ENSEMBLE.min = 0.7, ENSEMBLE.exponent = 1,
    ENSEMBLE.weight.min = 0.05,
    input.weights = NULL,
    MAXENT = 1, MAXNET = 1, MAXLIKE = 1, GBM = 1, GBMSTEP = 1, RF = 1, CF = 1, 
    GLM = 1, GLMSTEP = 1, GAM = 1, GAMSTEP = 1, MGCV = 1, MGCVFIX = 0, 
    EARTH = 1, RPART = 1, NNET = 1, FDA = 1, SVM = 1 , SVME = 1, GLMNET = 1,
    BIOCLIM.O = 0, BIOCLIM = 1, DOMAIN = 1, MAHAL = 1, MAHAL01 = 1,
    PROBIT = FALSE,
    Yweights = "BIOMOD", 
    layer.drops = NULL, factors = NULL, dummy.vars = NULL,
    formulae.defaults = TRUE, maxit = 100,
    MAXENT.a = NULL, MAXENT.an = 10000, 
    MAXENT.path = paste(getwd(), "/models/maxent_", species.name,  sep=""), 
    MAXNET.classes = "default", MAXNET.clamp = FALSE, MAXNET.type = "cloglog",
    MAXLIKE.formula = NULL, MAXLIKE.method = "BFGS",
    GBM.formula = NULL, GBM.n.trees = 2001, 
    GBMSTEP.gbm.x = 2:(length(var.names)+1), GBMSTEP.tree.complexity = 5, 
    GBMSTEP.learning.rate = 0.005, 
    GBMSTEP.bag.fraction = 0.5, GBMSTEP.step.size = 100, 
    RF.formula = NULL, RF.ntree = 751, RF.mtry = floor(sqrt(length(var.names))), 
    CF.formula = NULL, CF.ntree = 751, CF.mtry = floor(sqrt(length(var.names))),
    GLM.formula = NULL, GLM.family = binomial(link = "logit"), 
    GLMSTEP.steps = 1000, STEP.formula = NULL, GLMSTEP.scope = NULL, GLMSTEP.k = 2, 
    GAM.formula = NULL, GAM.family = binomial(link = "logit"), 
    GAMSTEP.steps = 1000, GAMSTEP.scope = NULL, GAMSTEP.pos = 1, 
    MGCV.formula = NULL, MGCV.select = FALSE, 
    MGCVFIX.formula = NULL, 
    EARTH.formula = NULL, 
    EARTH.glm = list(family = binomial(link = "logit"), maxit = maxit), 
    RPART.formula = NULL, RPART.xval = 50, 
    NNET.formula = NULL, NNET.size = 8, NNET.decay = 0.01, 
    FDA.formula = NULL, 
    SVM.formula = NULL, 
    SVME.formula = NULL, 
    GLMNET.nlambda = 100, GLMNET.class = FALSE,
    BIOCLIM.O.fraction = 0.9,
    MAHAL.shape = 1)

ensemble.calibrate.models.gbm(x = NULL, p = NULL, a = NULL, an = 1000, excludep = FALSE, 
    k = 4, 
    TrainData = NULL,
    VIF = FALSE, COR = FALSE,
    SINK = FALSE, PLOTS = FALSE, 
    species.name = "Species001",
    Yweights = "BIOMOD", 
    layer.drops = NULL, factors = NULL, 
    GBMSTEP.gbm.x = 2:(ncol(TrainData.orig)), 
    complexity = c(3:6), learning = c(0.005, 0.002, 0.001), 
    GBMSTEP.bag.fraction = 0.5, GBMSTEP.step.size = 100)

ensemble.calibrate.models.nnet(x = NULL, p = NULL, a = NULL, an = 1000, excludep = FALSE, 
    k = 4, 
    TrainData = NULL,
    VIF = FALSE, COR = FALSE,
    SINK = FALSE, PLOTS = FALSE, 
    species.name = "Species001",
    Yweights = "BIOMOD", 
    layer.drops = NULL, factors = NULL, 
    formulae.defaults = TRUE, maxit = 100, 
    NNET.formula = NULL, 
    sizes = c(2, 4, 6, 8), decays = c(0.1, 0.05, 0.01, 0.001) )

ensemble.drop1(x = NULL, p = NULL, 
    a = NULL, an = 1000, excludep = FALSE, target.groups = FALSE,
    k = 0, pt = NULL, at = NULL, SSB.reduce = FALSE, CIRCLES.d = 100000, 
    TrainData = NULL, TestData = NULL,
    VIF = FALSE, COR = FALSE,
    SINK = FALSE, 
    species.name = "Species001",
    difference = FALSE, variables.alone = FALSE,
    ENSEMBLE.tune = FALSE,
    ENSEMBLE.best = 0, ENSEMBLE.min = 0.7, ENSEMBLE.exponent = 1,
    input.weights = NULL, 
    MAXENT = 1, MAXNET = 1, MAXLIKE = 1, GBM = 1, GBMSTEP = 0, RF = 1, CF = 1, 
    GLM = 1, GLMSTEP = 1, GAM = 1, GAMSTEP = 1, MGCV = 1, MGCVFIX = 0, 
    EARTH = 1, RPART = 1, NNET = 1, FDA = 1, SVM = 1, SVME = 1, GLMNET = 1,
    BIOCLIM.O = 0, BIOCLIM = 1, DOMAIN = 1, MAHAL = 1, MAHAL01 = 1,
    PROBIT = FALSE,
    Yweights = "BIOMOD", 
    layer.drops = NULL, factors = NULL, dummy.vars = NULL, 
    maxit = 100,
    MAXENT.a = NULL, MAXENT.an = 10000, 
    MAXENT.path = paste(getwd(), "/models/maxent_", species.name,  sep=""), 
    MAXNET.classes = "default", MAXNET.clamp = FALSE, MAXNET.type = "cloglog",
    MAXLIKE.method = "BFGS",
    GBM.n.trees = 2001, 
    GBMSTEP.tree.complexity = 5, GBMSTEP.learning.rate = 0.005, 
    GBMSTEP.bag.fraction = 0.5, GBMSTEP.step.size = 100, 
    RF.ntree = 751, 
    CF.ntree = 751,
    GLM.family = binomial(link = "logit"), 
    GLMSTEP.steps = 1000, GLMSTEP.scope = NULL, GLMSTEP.k = 2, 
    GAM.family = binomial(link = "logit"), 
    GAMSTEP.steps = 1000, GAMSTEP.scope = NULL, GAMSTEP.pos = 1, 
    MGCV.select = FALSE, 
    EARTH.glm = list(family = binomial(link = "logit"), maxit = maxit), 
    RPART.xval = 50, 
    NNET.size = 8, NNET.decay = 0.01, 
    GLMNET.nlambda = 100, GLMNET.class = FALSE,
    BIOCLIM.O.fraction = 0.9,
    MAHAL.shape = 1)

ensemble.weights(weights = c(0.9, 0.8, 0.7, 0.5), 
    best = 0, min.weight = 0, 
    exponent = 1, digits = 6)

ensemble.strategy(TrainData = NULL, TestData = NULL, 
    verbose = FALSE,
    ENSEMBLE.best = c(4:10), ENSEMBLE.min = c(0.7),
    ENSEMBLE.exponent = c(1, 2, 3) )

ensemble.formulae(x, 
    layer.drops = NULL, factors = NULL, dummy.vars = NULL, weights = NULL)

ensemble.threshold(eval, threshold.method = "spec_sens", threshold.sensitivity = 0.9, 
    threshold.PresenceAbsence = FALSE, Pres, Abs) 

ensemble.VIF(x = NULL, a = NULL, an = 10000, 
    VIF.max = 10, keep = NULL,
    layer.drops = NULL, factors = NULL, dummy.vars = NULL)

ensemble.VIF.dataframe(x=NULL, 
    VIF.max=10, keep=NULL,
    car=TRUE, silent=F)

ensemble.pairs(x = NULL, a = NULL, an = 10000)

Arguments

x

RasterStack object (stack) containing all layers that correspond to explanatory variables

p

presence points used for calibrating the suitability models, typically available in 2-column (lon, lat) dataframe; see also prepareData and extract

a

background points used for calibrating the suitability models (except for maxent), typically available in 2-column (lon, lat) dataframe; see also prepareData and extract

an

number of background points for calibration to be selected with randomPoints in case argument a is missing

excludep

parameter that indicates (if TRUE) that presence points will be excluded from the background points; see also randomPoints

target.groups

Parameter that indicates (if TRUE) that the provided background points (argument a) represent presence points from a target group sensu Phillips et al. 2009 (these are species that are all collected or observed using the same methods or equipment). Setting the parameter to TRUE results in selecting the centres of cells of the target groups as background points, while avoiding to select the same cells twice. Via argument excludep, it is possible to filter out cells with presence observations (argument p).

k

If larger than 1, the number of groups to split between calibration (k-1) and evaluation (1) data sets (for example, k = 4 results in 3/4 of presence and background points to be used for calibrating the models, and 1/4 of presence and background points to be used for evaluating the models). For ensemble.calibrate.weights, ensemble.calibrate.models.gbm and ensemble.calibrate.models.nnet, this procedure is repeated k times (k-fold cross-validation). See also kfold.

pt

presence points used for evaluating the suitability models, available in 2-column (lon, lat) dataframe; see also prepareData and extract

at

background points used for evaluating the suitability models, available in 2-column (lon, lat) dataframe; see also prepareData and extract

SSB.reduce

If TRUE, then new background points that will be used for evaluationg the suitability models will be selected (randomPoints) in circular neighbourhoods (created with circles) around presence locations (p and pt). The abbreviation of SSB refers to spatial sorting bias; see also ssb.

CIRCLES.d

Radius in m of circular neighbourhoods (created with circles) around presence locations (p and pt).

TrainData

dataframe with first column 'pb' describing presence (1) and absence (0) and other columns containing explanatory variables; see also prepareData. In case that this dataframe is provided, then locations p and a are not used. For the maximum entropy model (maxent), a different dataframe is used for calibration; see parameter MAXENT.TrainData.

TestData

dataframe with first column 'pb' describing presence (1) and absence (0) and other columns containing explanatory variables; see also prepareData. In case that this dataframe is provided, then locations pt and at are not used. For ensemble.strategy, this dataframe should be a dataframe that contains predictions for various models - such dataframe can be provided by the ensemble.calibrate.models or ensemble.raster functions.

VIF

Estimate the variance inflation factors based on a linear model calibrated on the training data (if TRUE). Only background locations will be used and the response variable 'pb' will be replaced by a random variable. See also vif.

COR

Provide information on the correlation between the numeric explanatory variables (if TRUE). See also cor.

SINK

Append the results to a text file in subfolder 'outputs' (if TRUE). The name of file is based on argument species.name. In case the file already exists, then results are appended. See also sink.

PLOTS

Disabled option of plotting evaluation results(BiodiversityR version 2.9-1) - see examples how to plot results afterwards and also evaluate.

CATCH.OFF

Disable calls to function tryCatch.

threshold.method

Method to calculate the threshold between predicted absence and presence; possibilities include spec_sens (highest sum of the true positive rate and the true negative rate), kappa (highest kappa value), no_omission (highest threshold that corresponds to no omission), prevalence (modeled prevalence is closest to observed prevalence) and equal_sens_spec (equal true positive rate and true negative rate). See threshold. Options specific to the BiodiversityR implementation are: threshold2005.mean, threshold2005.min, threshold2013.mean and threshold2013.min (resulting in calculating the mean or minimum value of recommended threshold values by studies published in 2005 and 2013; see details below).

threshold.sensitivity

Sensitivity value for threshold.method = 'sensitivity'. See threshold.

threshold.PresenceAbsence

If TRUE calculate thresholds with the PresenceAbsence package. See optimal.thresholds.

evaluations.keep

Keep the results of evaluations (if TRUE). See also evaluate.

models.list

list with 'old' model objects such as MAXENT or RF.

models.keep

store the details for each suitability modelling algorithm (if TRUE). (This may be particularly useful when projecting to different possible future climates.)

models.save

Save the list with model details to a file (if TRUE). The filename will be species.name with extension .models; this file will be saved in subfolder of models. When loading this file, model results will be available as ensemble.models.

species.name

Name by which the model details will be saved to a file; see also argument models.save

data.keep

Keep the data for each k-fold cross-validation run (if TRUE).

ENSEMBLE.tune

Determine weights for the ensemble model based on AUC values (if TRUE). See details.

ENSEMBLE.best

The number of individual suitability models to be used in the consensus suitability map (based on a weighted average). In case this parameter is smaller than 1 or larger than the number of positive input weights of individual models, then all individual suitability models with positive input weights are included in the consensus suitability map. In case a vector is provided, ensemble.strategy is called internally to determine weights for the ensemble model.

ENSEMBLE.min

The minimum input weight (typically corresponding to AUC values) for a model to be included in the ensemble. In case a vector is provided, function ensemble.strategy is called internally to determine weights for the ensemble model.

ENSEMBLE.exponent

Exponent applied to AUC values to convert AUC values into weights (for example, an exponent of 2 converts input weights of 0.7, 0.8 and 0.9 into 0.7^2=0.49, 0.8^2=0.64 and 0.9^2=0.81). See details.

ENSEMBLE.weight.min

The minimum output weight for models included in the ensemble, applying to weights that sum to one. Note that ENSEMBLE.min typically refers to input AUC values.

input.weights

array with numeric values for the different modelling algorithms; if NULL then values provided by parameters such as MAXENT and GBM will be used. As an alternative, the output from ensemble.calibrate.weights can be used.

MAXENT

number: if larger than 0, then a maximum entropy model (maxent) will be fitted among ensemble

MAXNET

number: if larger than 0, then a maximum entropy model (maxnet) will be fitted among ensemble

MAXLIKE

number: if larger than 0, then a maxlike model (maxlike) will be fitted among ensemble

GBM

number: if larger than 0, then a boosted regression trees model (gbm) will be fitted among ensemble

GBMSTEP

number: if larger than 0, then a stepwise boosted regression trees model (gbm.step) will be fitted among ensemble

RF

number: if larger than 0, then a random forest model (randomForest) will be fitted among ensemble

CF

number: if larger than 0, then a random forest model (cforest) will be fitted among ensemble

GLM

number: if larger than 0, then a generalized linear model (glm) will be fitted among ensemble

GLMSTEP

number: if larger than 0, then a stepwise generalized linear model (stepAIC) will be fitted among ensemble

GAM

number: if larger than 0, then a generalized additive model (gam) will be fitted among ensemble

GAMSTEP

number: if larger than 0, then a stepwise generalized additive model (step.gam) will be fitted among ensemble

MGCV

number: if larger than 0, then a generalized additive model (gam) will be fitted among ensemble

MGCVFIX

number: if larger than 0, then a generalized additive model with fixed d.f. regression splines (gam) will be fitted among ensemble

EARTH

number: if larger than 0, then a multivariate adaptive regression spline model (earth) will be fitted among ensemble

RPART

number: if larger than 0, then a recursive partioning and regression tree model (rpart) will be fitted among ensemble

NNET

number: if larger than 0, then an artificial neural network model (nnet) will be fitted among ensemble

FDA

number: if larger than 0, then a flexible discriminant analysis model (fda) will be fitted among ensemble

SVM

number: if larger than 0, then a support vector machine model (ksvm) will be fitted among ensemble

SVME

number: if larger than 0, then a support vector machine model (svm) will be fitted among ensemble

GLMNET

number: if larger than 0, then a GLM with lasso or elasticnet regularization (glmnet) will be fitted among ensemble

BIOCLIM.O

number: if larger than 0, then the original BIOCLIM algorithm (ensemble.bioclim) will be fitted among ensemble

BIOCLIM

number: if larger than 0, then the BIOCLIM algorithm (bioclim) will be fitted among ensemble

DOMAIN

number: if larger than 0, then the DOMAIN algorithm (domain) will be fitted among ensemble

MAHAL

number: if larger than 0, then the Mahalanobis algorithm (mahal) will be fitted among ensemble

MAHAL01

number: if larger than 0, then the Mahalanobis algorithm (mahal) will be fitted among ensemble, using a transformation method afterwards whereby output is within the range between 0 and 1 (see details)

PROBIT

If TRUE, then subsequently to the fitting of the individual algorithm (e.g. maximum entropy or GAM) a generalized linear model (glm) with probit link family=binomial(link="probit") will be fitted to transform the predictions, using the previous predictions as explanatory variable. This transformation results in all model predictions to be probability estimates.

Yweights

chooses how cases of presence and background (absence) are weighted; "BIOMOD" results in equal weighting of all presence and all background cases, "equal" results in equal weighting of all cases. The user can supply a vector of weights similar to the number of cases in the calibration data set.

layer.drops

vector that indicates which layers should be removed from RasterStack x. This argument is especially useful for the ensemble.drop1 function. See also addLayer.

factors

vector that indicates which variables are factors; see also prepareData

dummy.vars

vector that indicates which variables are dummy variables (influences formulae suggestions)

formulae.defaults

Suggest formulae for most of the models (if TRUE). See also ensemble.formulae.

maxit

Maximum number of iterations for some of the models. See also glm.control, gam.control, gam.control and nnet.

MAXENT.a

background points used for calibrating the maximum entropy model (maxent), typically available in 2-column (lon, lat) dataframe; see also prepareData and extract.

MAXENT.an

number of background points for calibration to be selected with randomPoints in case argument MAXENT.a is missing

MAXENT.path

path to the directory where output files of the maximum entropy model are stored; see also maxent

MAXNET.classes

continuous feature classes, either "default" or any subset of "lqpht" (linear, quadratic, product, hinge, threshold). Note that the "default" option chooses feature classes based on the number of presence locations as "l" (< 10 locations), "lq" (10 - 14 locations), "lqh" (15 - 79 locations) or "lqph" (> 79 locations). See also maxnet.

MAXNET.clamp

restrict predictors and features to the range seen during model training; see also predict.maxnet

MAXNET.type

type of response required; see also predict.maxnet

MAXLIKE.formula

formula for the maxlike algorithm; see also maxlike

MAXLIKE.method

method for the maxlike algorithm; see also optim

GBM.formula

formula for the boosted regression trees algorithm; see also gbm

GBM.n.trees

total number of trees to fit for the boosted regression trees model; see also gbm

GBMSTEP.gbm.x

indices of column numbers with explanatory variables for stepwise boosted regression trees; see also gbm.step

GBMSTEP.tree.complexity

complexity of individual trees for stepwise boosted regression trees; see also gbm.step

GBMSTEP.learning.rate

weight applied to individual trees for stepwise boosted regression trees; see also gbm.step

GBMSTEP.bag.fraction

proportion of observations used in selecting variables for stepwise boosted regression trees; see also gbm.step

GBMSTEP.step.size

number of trees to add at each cycle for stepwise boosted regression trees (should be small enough to result in a smaller holdout deviance than the initial number of trees [50]); see also gbm.step

RF.formula

formula for random forest algorithm; see also randomForest

RF.ntree

number of trees to grow for random forest algorithm; see also randomForest

RF.mtry

number of variables randomly sampled as candidates at each split for random forest algorithm; see also randomForest

CF.formula

formula for random forest algorithm; see also cforest

CF.ntree

number of trees to grow in a forest; see also cforest_control

CF.mtry

number of input variables randomly sampled as candidates at each node for random forest like algorithms; see also cforest_control

GLM.formula

formula for the generalized linear model; see also glm

GLM.family

description of the error distribution and link function for the generalized linear model; see also glm

GLMSTEP.steps

maximum number of steps to be considered for stepwise generalized linear model; see also stepAIC

STEP.formula

formula for the "starting model" to be considered for stepwise generalized linear model; see also stepAIC

GLMSTEP.scope

range of models examined in the stepwise search; see also stepAIC

GLMSTEP.k

multiple of the number of degrees of freedom used for the penalty (only k = 2 gives the genuine AIC); see also stepAIC

GAM.formula

formula for the generalized additive model; see also gam

GAM.family

description of the error distribution and link function for the generalized additive model; see also gam

GAMSTEP.steps

maximum number of steps to be considered in the stepwise generalized additive model; see also step.gam

GAMSTEP.scope

range of models examined in the step-wise search n the stepwise generalized additive model; see also step.gam

GAMSTEP.pos

parameter expected to be set to 1 to allow for fitting of the stepwise generalized additive model

MGCV.formula

formula for the generalized additive model; see also gam

MGCV.select

if TRUE, then the smoothing parameter estimation that is part of fitting can completely remove terms from the model; see also gam

MGCVFIX.formula

formula for the generalized additive model with fixed d.f. regression splines; see also gam (the default formulae sets "s(..., fx = TRUE, ...)"; see also s)

EARTH.formula

formula for the multivariate adaptive regression spline model; see also earth

EARTH.glm

list of arguments to pass on to glm; see also earth

RPART.formula

formula for the recursive partioning and regression tree model; see also rpart

RPART.xval

number of cross-validations for the recursive partioning and regression tree model; see also rpart.control

NNET.formula

formula for the artificial neural network model; see also nnet

NNET.size

number of units in the hidden layer for the artificial neural network model; see also nnet

NNET.decay

parameter of weight decay for the artificial neural network model; see also nnet

FDA.formula

formula for the flexible discriminant analysis model; see also fda

SVM.formula

formula for the support vector machine model; see also ksvm

SVME.formula

formula for the support vector machine model; see also svm

GLMNET.nlambda

The number of lambda values; see also glmnet

GLMNET.class

Use the predicted class to calculate the mean predictions of GLMNET; see predict.glmnet

BIOCLIM.O.fraction

Fraction of range representing the optimal limits, default value of 0.9 as in the original BIOCLIM software (ensemble.bioclim).

MAHAL.shape

parameter that influences the transformation of output values of mahal. See details section.

TrainTestData

dataframe with first column 'pb' describing presence (1) and absence (0) and other columns containing explanatory variables; see also prepareData. In case that this dataframe is provided, then locations p and a are not used. This data set will also be used for the maximum entropy and maximum likelihood models.

get.block

if TRUE, instead of creating k-fold cross-validation subsets randomly (kfold), create 4 subsets of presence and background locations with get.block.

block.default

if FALSE, instead of making the first division of presence point locations along the y-coordinates (latitude) as in get.block, make the first division along the x-coordinates (longitude).

get.subblocks

if TRUE, then 4 subsets of presence and background locations are generated in a checkerboard configuration by applying get.block to each of the 4 blocks generated by get.block in a first step.

complexity

vector with values of complexity of individual trees (tree.complexity) for boosted regression trees; see also gbm.step

learning

vector with values of weights applied to individual trees (learning.rate) for boosted regression trees; see also gbm.step

sizes

vector with values of number of units in the hidden layer for the artificial neural network model; see also nnet

decays

vector with values of weight decay for the artificial neural network model; see also nnet

difference

if TRUE, then AUC values of the models with all variables are subtracted from the models where one explanatory variable was excluded. After subtraction, positive values indicate that the model without the explanatory variable has a higher AUC than the model with all variables.

variables.alone

if TRUE, then models are also fitted using each explanatory variable as single explanatory variable

weights

input weights for the ensemble.weights function

best

The number of final weights. In case this parameter is smaller than 1 or larger than the number of positive input weights of individual models, then this parameter is ignored.

min.weight

The minimum input weight to be included in the output.

exponent

Exponent applied to AUC values to convert AUC values into weights (for example, an exponent of 2 converts input weights of 0.7, 0.8 and 0.9 into 0.7^2=0.49, 0.8^2=0.64 and 0.9^2=0.81). See details.

digits

Number of number of decimal places in the output weights; see also round.

verbose

If TRUE, then provide intermediate results for ensemble.strategy)

eval

ModelEvaluation object obtained by evaluate

Pres

Suitabilities (probabilities) at presence locations

Abs

Suitabilities (probabilities) at background locations

VIF.max

Maximum Variance Inflation Factor of the selected subset of variables. In case that at least one of the variables has VIF larger than VIF.max, then the variable with the highest VIF will be removed in the next step.

keep

character vector with names of the variables to be kept.

car

Also provide results from vif.

silent

Limit textual output.

Details

The basic function ensemble.calibrate.models first calibrates individual suitability models based on presence locations p and background locations a, then evaluates these suitability models based on presence locations pt and background locations at. While calibrating and testing individual models, results obtained via the evaluate function can be saved (evaluations.keep).

As an alternative to providing presence locations p, models can be calibrated with data provided in TrainData. In case that both p and TrainData are provided, then models will be calibrated with TrainData.

Calibration of the maximum entropy (MAXENT) algorithm is not based on background locations a, but based on background locations MAXENT.a instead. However, to compare evaluations with evaluations of other algorithms, during evaluations of the MAXENT algorithm, presence locations p and background locations a are used (and not background locations MAXENT.a).

Output from the GLMNET algorithm is calculated as the mean of the output from predict.glmnet. With option GLMNET.class = TRUE, the mean output is the mean prediction of class 1. With option GLMNET.class = FALSE, the mean output is the mean of the responses.

As the Mahalanobis function (mahal) does not always provide values within the range of 0 - 1, the output values are rescaled with option MAHAL01 by first subtracting the value of 1 - MAHAL.shape from each prediction, followed by calculating the absolute value, followed by calculating the reciprocal value and finally multiplying this reciprocal value with MAHAL.shape. As this rescaling method does not estimate probabilities, inclusion in the calculation of a (weighted) average of ensemble probabilities may be problematic and the PROBIT transformation may help here (the same applies to other distance-based methods).

With parameter ENSEMBLE.best, the subset of best models (evaluated by the individual AUC values) can be selected and only those models will be used for calculating the ensemble model (in other words, weights for models not included in the ensemble will be set to zero). It is possible to further increase the contribution to the ensemble model for models with higher AUC values through parameter ENSEMBLE.exponent. With ENSEMBLE.exponent = 2, AUC values of 0.7, 0.8 and 0.9 are converted into weights of 0.7^2=0.49, 0.8^2=0.64 and 0.9^2=0.81). With ENSEMBLE.exponent = 4, AUC values of 0.7, 0.8 and 0.9 are converted into weights of 0.7^4=0.2401, 0.8^4=0.4096 and 0.9^2=0.6561).

ENSEMBLE.tune will result in an internal procedure whereby the best selection of parameter values for ENSEMBLE.min, ENSEMBLE.best or ENSEMBLE.exponent can be identified. Through a factorial procedure, the ensemble model with best AUC for a specific combination of parameter values is identified. The procedure also provides the weights that correspond to the best ensemble. In case that ENSEMBLE.tune is set to FALSE, then the ensemble is calculated based on the input weights.

Function ensemble.calibrate.weights splits the presence and background locations in a user-defined (k) number of subsets (i.e. k-fold cross-validation), then sequentially calibrates individual suitability models with (k-1) combined subsets and evaluates those with the remaining one subset, whereby each subset is used once for evaluation in the user-defined number (k) of runs. For example, k = 4 results in splitting the locations in 4 subsets, then using one of these subsets in turn for evaluations (see also kfold). Note that for the maximum entropy (MAXENT) algorithm, the same background data will be used in each cross-validation run (this is based on the assumption that a large number (~10000) of background locations are used).

Among the output from function ensemble.calibrate.weights are suggested weights for an ensemble model (output.weights and output.weights.AUC), and information on the respective AUC values of the ensemble model with the suggested weights for each of the (k) subsets. Suggested weights output.weights are calculated as the average of the weights of the different algorithms (submodels) of the k ensembles. Suggested weights output.weights.AUC are calculated as the average of the AUC of the different algorithms of the for the k runs.

Function ensemble.calibrate.models.gbm allows to test various combinations of parameters tree.complexity and learning.rate for the gbm.step model.

Function ensemble.calibrate.models.nnet allows to test various combinations of parameters size and decay for the nnet model.

Function ensemble.drop1 allows to test the effects of leaving out each of the explanatory variables, and comparing these results with the "full" model. Note that option of difference = TRUE may result in positive values, indicating that the model without the explanatory variable having larger AUC than the "full" model. A procedure is included to estimate the deviance of a model based on the fitted values, using -2 * (sum(x*log(x)) + sum((1-x)*log(1-x))) where x is a vector of the fitted values for a respective model. (It was checked that this procedure results in similar deviance estimates for the null and 'full' models for glm, but note that it is not certain whether deviance can be calculated in a similar way for other submodels.)

Function ensemble.formulae provides suggestions for formulae that can be used for ensemble.calibrate.models and ensemble.raster. This function is always used internally by the ensemble.drop1 function.

The ensemble.weights function is used internally by the ensemble.calibrate.models and ensemble.raster functions, using the input weights for the different suitability modelling algorithms. Ties between input weights result in the same output weights.

The ensemble.strategy function is used internally by the ensemble.calibrate.models function, using the train and test data sets with predictions of the different suitability modelling algorithms and different combinations of parameters ENSEMBLE.best, ENSEMBLE.min and ENSEMBLE.exponent. The final ensemble model is based on the parameters that generate the best AUC.

The ensemble.threshold function is used internally by the ensemble.calibrate.models, ensemble.mean and ensemble.plot functions. threshold2005.mean results in calculating the mean value of threshold methods that resulted in better results (calculated by optimal.thresholds with methods of ObsPrev, MeanProb, MaxSens+Spec, Sens=Spec and MinROCdist) in a study by Liu et al. (Ecography 28: 385-393. 2005). threshold2005.min results in calculating the mean value of threshold methods that resulted in better results (calculated by optimal.thresholds with methods of ObsPrev, MeanProb and MaxSens+Spec) in a study by Liu et al. (Ecography 28: 385-393. 2005). threshold2013.mean results in calculating the mean value of threshold methods that resulted in better results (calculated by optimal.thresholds with methods of ObsPrev, MeanProb, MaxSens+Spec, Sens=Spec and MinROCdist) in a study by Liu et al. (J. Biogeogr. 40: 778-789. 2013). threshold2013.min results in calculating the minimum value of threshold methods that resulted in better results (calculated by optimal.thresholds with methods of ObsPrev, MeanProb, MaxSens+Spec, Sens=Spec and MinROCdist) in a study by Liu et al. (J. Biogeogr. 40: 778-789. 2013).

Function ensemble.VIF implements a stepwise procedure whereby the explanatory variable with highest Variance Inflation Factor is removed from the list of variables. The procedure ends when no variable has VIF larger than parameter VIF.max.

Function ensemble.pairs provides a matrix of scatterplots similar to the example of pairs for version 3.4.3 of that package.

Value

Function ensemble.calibrate.models (potentially) returns a list with results from evaluations (via evaluate) of calibration and test runs of individual suitability models.

Function ensemble.calibrate.weights returns a matrix with, for each individual suitability model, the AUC of each run and the average AUC over the runs. Models are sorted by the average AUC. The average AUC for each model can be used as input weights for the ensemble.raster function.

Functions ensemble.calibrate.models.gbm and ensemble.calibrate.models.nnet return a matrix with, for each combination of model parameters, the AUC of each run and the average AUC. Models are sorted by the average AUC.

Author(s)

Roeland Kindt (World Agroforestry Centre)

References

Kindt R. 2018. Ensemble species distribution modelling with transformed suitability values. Environmental Modelling & Software 100: 136-145. \Sexpr[results=rd]{tools:::Rd_expr_doi("10.1016/j.envsoft.2017.11.009")}

Buisson L, Thuiller W, Casajus N, Lek S and Grenouillet G. 2010. Uncertainty in ensemble forecasting of species distribution. Global Change Biology 16: 1145-1157

Liu C, Berry PM, Dawson TP and Pearson RC. 2005. Selecting thresholds of occurrence in the prediction of species distributions. Ecography 28: 385-393

Liu C, White M and Newell G. 2013. Selecting thresholds for the prediction of species occurrence with presence-only data. Journal of Biogeography 40: 778-789

Phillips SJ, Dudik M, Elith J et al. 2009. Sample selection bias and presence-only distribution models: implications for background and pseudo-absence data. Ecological Applications 19: 181-197.

See Also

ensemble.raster, ensemble.batch

Examples

## Not run: 
# based on examples in the dismo package

# get predictor variables
library(dismo)
predictor.files <- list.files(path=paste(system.file(package="dismo"), '/ex', sep=''),
    pattern='grd', full.names=TRUE)
predictors <- stack(predictor.files)
# subset based on Variance Inflation Factors
predictors <- subset(predictors, subset=c("bio5", "bio6", 
    "bio16", "bio17", "biome"))
predictors
predictors@title <- "predictors"

# presence points
presence_file <- paste(system.file(package="dismo"), '/ex/bradypus.csv', sep='')
pres <- read.table(presence_file, header=TRUE, sep=',')[,-1]

# the kfold function randomly assigns data to groups; 
# groups are used as calibration (1/4) and training (3/4) data
groupp <- kfold(pres, 4)
pres_train <- pres[groupp !=  1, ]
pres_test <- pres[groupp ==  1, ]

# choose background points
background <- randomPoints(predictors, n=1000, extf=1.00)
colnames(background)=c('lon', 'lat')
groupa <- kfold(background, 4)
backg_train <- background[groupa != 1, ]
backg_test <- background[groupa == 1, ]

# formulae for random forest and generalized linear model
# compare with: ensemble.formulae(predictors, factors=c("biome"))

rfformula <- as.formula(pb ~ bio5+bio6+bio16+bio17)

glmformula <- as.formula(pb ~ bio5 + I(bio5^2) + I(bio5^3) + 
    bio6 + I(bio6^2) + I(bio6^3) + bio16 + I(bio16^2) + I(bio16^3) + 
    bio17 + I(bio17^2) + I(bio17^3) )

# fit four ensemble models (RF, GLM, BIOCLIM, DOMAIN)
# factors removed for BIOCLIM, DOMAIN, MAHAL
ensemble.nofactors <- ensemble.calibrate.models(x=predictors, p=pres_train, a=backg_train, 
    pt=pres_test, at=backg_test,
    species.name="Bradypus",
    ENSEMBLE.tune=TRUE,
    ENSEMBLE.min = 0.65,
    MAXENT=0, MAXNET=0, MAXLIKE=0, GBM=0, GBMSTEP=0, RF=1, CF=0, 
    GLM=1, GLMSTEP=0, GAM=0, GAMSTEP=0, MGCV=0, MGCVFIX=0, 
    EARTH=0, RPART=0, NNET=0, FDA=0, SVM=0, SVME=0, GLMNET=0,
    BIOCLIM.O=0, BIOCLIM=1, DOMAIN=1, MAHAL=0, MAHAL01=0,
    Yweights="BIOMOD",
    factors="biome",
    evaluations.keep=TRUE, models.keep=TRUE,
    RF.formula=rfformula,
    GLM.formula=glmformula)

# with option models.keep, all model objects are saved in ensemble object
# the same slots can be used to replace model objects with new calibrations
ensemble.nofactors$models$RF
summary(ensemble.nofactors$models$GLM)
ensemble.nofactors$models$BIOCLIM
ensemble.nofactors$models$DOMAIN
ensemble.nofactors$models$formulae

# evaluations are kept in different slot
attributes(ensemble.nofactors$evaluations)
plot(ensemble.nofactors$evaluations$RF.T, "ROC")

# fit four ensemble models (RF, GLM, BIOCLIM, DOMAIN) using default formulae
# variable 'biome' is not included as explanatory variable
# results are provided in a file in the 'outputs' subfolder of the working
# directory
ensemble.nofactors <- ensemble.calibrate.models(x=predictors,
    p=pres_train, a=backg_train, 
    pt=pres_test, at=backg_test,
    layer.drops="biome",
    species.name="Bradypus",
    ENSEMBLE.tune=TRUE,
    ENSEMBLE.min=0.65,
    SINK=TRUE,
    MAXENT=0, MAXNET=0, MAXLIKE=0, GBM=0, GBMSTEP=0, RF=1, CF=0, 
    GLM=1, GLMSTEP=0, GAM=0, 
    GAMSTEP=0, MGCV=0, MGCVFIX=0, EARTH=0, RPART=0, NNET=0, FDA=0, 
    SVM=0, SVME=0, GLMNET=0,
    BIOCLIM.O=0, BIOCLIM=1, DOMAIN=1, MAHAL=0, MAHAL01=0,
    Yweights="BIOMOD", 
    evaluations.keep=TRUE,
    formulae.defaults=TRUE)    

# after fitting the individual algorithms (submodels),
# transform predictions with a probit link.
ensemble.nofactors <- ensemble.calibrate.models(x=predictors,
    p=pres_train, a=backg_train, 
    pt=pres_test, at=backg_test,
    layer.drops="biome",
    species.name="Bradypus",
    SINK=TRUE,
    ENSEMBLE.tune=TRUE,
    ENSEMBLE.min=0.65,
    MAXENT=0, MAXNET=0, MAXLIKE=0, GBM=0, GBMSTEP=0, RF=1, CF=0,
    GLM=1, GLMSTEP=0, GAM=0, GAMSTEP=0, MGCV=0, MGCVFIX=0, 
    EARTH=0, RPART=0, NNET=0, FDA=0, SVM=0, SVME=0, GLMNET=0,
    BIOCLIM.O=0, BIOCLIM=1, DOMAIN=1, MAHAL=0, MAHAL01=0,
    PROBIT=TRUE,
    Yweights="BIOMOD", factors="biome",
    evaluations.keep=TRUE,
    formulae.defaults=TRUE)    

# Instead of providing presence and background locations, provide data.frames.
# Because 'biome' is a factor, RasterStack needs to be provided
# to check for levels in the Training and Testing data set.
TrainData1 <- prepareData(x=predictors, p=pres_train, b=backg_train, 
    factors=c("biome"), xy=FALSE)
TestData1 <- prepareData(x=predictors, p=pres_test, b=backg_test, 
    factors=c("biome"), xy=FALSE)
ensemble.factors1 <- ensemble.calibrate.models(x=predictors, 
    TrainData=TrainData1, TestData=TestData1,
    p=pres_train, a=backg_train, 
    pt=pres_test, at=backg_test,
    species.name="Bradypus",
    SINK=TRUE,
    ENSEMBLE.tune=TRUE,
    ENSEMBLE.min=0.65,
    MAXENT=0, MAXNET=1, MAXLIKE=1, GBM=1, GBMSTEP=0, RF=1, CF=1,
    GLM=1, GLMSTEP=1, GAM=1, GAMSTEP=1, MGCV=1, MGCVFIX=0, 
    EARTH=1, RPART=1, NNET=1, FDA=1, SVM=1, SVME=1, GLMNET=1,
    BIOCLIM.O=1, BIOCLIM=1, DOMAIN=1, MAHAL=0, MAHAL01=1,
    Yweights="BIOMOD", factors="biome",
    evaluations.keep=TRUE)

# compare different methods of calculating ensembles
ensemble.factors2 <- ensemble.calibrate.models(x=predictors, 
    TrainData=TrainData1, TestData=TestData1,
    p=pres_train, a=backg_train, 
    pt=pres_test, at=backg_test,
    species.name="Bradypus",
    SINK=TRUE,
    ENSEMBLE.tune=TRUE,
    MAXENT=0, MAXNET=1, MAXLIKE=1, GBM=1, GBMSTEP=0, RF=1, CF=1,
    GLM=1, GLMSTEP=1, GAM=1, GAMSTEP=1, MGCV=1, MGCVFIX=1, 
    EARTH=1, RPART=1, NNET=1, FDA=1, SVM=1, SVME=1, GLMNET=1,
    BIOCLIM.O=1, BIOCLIM=1, DOMAIN=1, MAHAL=0, MAHAL01=1,
    ENSEMBLE.best=c(4:10), ENSEMBLE.exponent=c(1, 2, 3),
    Yweights="BIOMOD", factors="biome",
    evaluations.keep=TRUE)

# test performance of different suitability models
# data are split in 4 subsets, each used once for evaluation
ensemble.nofactors2 <- ensemble.calibrate.weights(x=predictors, 
    p=pres, a=background, k=4, 
    species.name="Bradypus",
    SINK=TRUE, PROBIT=TRUE,
    MAXENT=0, MAXNET=1, MAXLIKE=1, GBM=1, GBMSTEP=0, RF=1, CF=1,
    GLM=1, GLMSTEP=1, GAM=1, GAMSTEP=1, MGCV=1, MGCVFIX=1, 
    EARTH=1, RPART=1, NNET=1, FDA=1, SVM=1, SVME=1, GLMNET=1,
    BIOCLIM.O=1, BIOCLIM=1, DOMAIN=1, MAHAL=0, MAHAL01=1,
    ENSEMBLE.tune=TRUE,
    ENSEMBLE.best=0, ENSEMBLE.exponent=c(1, 2, 3),
    ENSEMBLE.min=0.7,
    Yweights="BIOMOD", 
    formulae.defaults=TRUE)
ensemble.nofactors2$AUC.table
ensemble.nofactors2$eval.table.all

# test the result of leaving out one of the variables from the model
# note that positive differences indicate that the model without the variable 
# has higher AUC than the full model
ensemble.variables <- ensemble.drop1(x=predictors, 
    p=pres, a=background, k=4,
    species.name="Bradypus",
    SINK=TRUE,
    difference=TRUE,
    VIF=TRUE, PROBIT=TRUE,
    MAXENT=0, MAXNET=1, MAXLIKE=1, GBM=1, GBMSTEP=0, RF=1, CF=1,
    GLM=1, GLMSTEP=1, GAM=1, GAMSTEP=1, MGCV=1, MGCVFIX=1, 
    EARTH=1, RPART=1, NNET=1, FDA=1, SVM=1, SVME=1, GLMNET=1,
    BIOCLIM.O=1, BIOCLIM=1, DOMAIN=1, MAHAL=0, MAHAL01=1,
    ENSEMBLE.tune=TRUE,
    ENSEMBLE.best=0, ENSEMBLE.exponent=c(1, 2, 3),
    ENSEMBLE.min=0.7,
    Yweights="BIOMOD", factors="biome")
ensemble.variables

# use function ensemble.VIF to select a subset of variables
# factor variables are not handled well by the function
# and therefore factors are removed
# however, one can check for factors with car::vif through
# the ensemble.calibrate.models function
# VIF.analysis$var.drops can be used as input for ensemble.calibrate.models or
# ensemble.calibrate.weights

predictors <- stack(predictor.files)
predictors <- subset(predictors, subset=c("bio1", "bio5", "bio6", "bio8", 
    "bio12", "bio16", "bio17", "biome"))

ensemble.pairs(predictors)

VIF.analysis <- ensemble.VIF(predictors, factors="biome")
VIF.analysis
# alternative solution where bio1 and bio12 are kept
VIF.analysis <- ensemble.VIF(predictors, factors="biome", 
    keep=c("bio1", "bio12"))
VIF.analysis

## End(Not run)

BiodiversityR documentation built on June 22, 2024, 11:57 a.m.