SDMSelect: SDMSelect: A package for cross-validation model selection and...

In statnmap/SDMSelect: Cross-validation model selection and species distribution mapping

Description

The SDMSelect package is used with four main functions: Prepare_dataset, findBestModel, ModelResults and Map_predict.

Vignettes

You can test the package through its vignettes. Vignettes are shown with:

• vignette("Covar_Selection", package = "SDMSelect")

• vignette("SDM_Selection", package = "SDMSelect")

You can also find the path of the complete vignettes to be run on your computer with:

• system.file("Covar_Selection", "Covar_Selection.Rmd", package = "SDMSelect")

• system.file("SDM_Selection", "SDM_Selection.Rmd", package = "SDMSelect")

Data preparation

• Prepare_covarStack: This creates a RasterStack from raster files paths. Rasters are reprojected if needed to match the reference raster to allow for stacking.

• CovarExtract: Extract covariates information for a SpatialointsDataFrame in the original raster files, whatever the projection.

• Prepare_dataset: Create the dataset for modelling containing only variables of interest. Data column is names "dataY" for modelling. This function can also resample a spatial dataset in a regular grid to reduce spatial auto-correlation.

• spatialcor_dist: Tests for spatial auto-correlation and proposes values of grid resolution to resample dataset in a regular grid.

• RefRasterize: Create a regular grid from a spatial dataset with a defined grid resolution. This can then be used with Prepare_dataset.

• Param_corr: Test for Spearman's rank correlation between covariates. Covariates couples with correlation above a specific threshold will not be tested in the same model during the cross-validation procedure.

Forward stepwise cross-validation procedure

• modelselect_opt: List of all options that are used for the model selection procedure. All are default values that can be modified according to a specific case study.

• findBestModel: Forward stepwise cross-validation procedure. A forward stepwise cross-validation procedure is run for each model type independently, but on the same subsets of data. All outputs are saved in a directory for further analyses.

• ModelOrder: Compare all models from all model types together to find the best models among all. Some figures shows differences in predictive power of the models. The list of best models with a statistically equivalent predictive power is saved for all models together, but also for each model type.

Model results analysis

• ModelResults: This outputs different figures and tables for the analysis of the model specified (typically the best one). Analysis of variance and gain in predictive power with regards to cross-validation, residual analysis, marginal effect of selected covariates, comparison of predictions with observations.

Mapping predictions

• Map_predict: This provides multiple map predictions for the model specified (typically the best one). Map of average prediction, minimum and maximum predictions (quantiles from parameters uncertainty), Inter-quartile range (better than standard deviation when uncertainty of predictions is not gaussian in the scale of the data; e.g. LogNormal or logit). For presence-absence models, there is also probability that the prediction is over the best threshold value (separating presence from absences). The intuitive threshold of 0.5 is not always the best if the dataset is not well balanced between presence and absence, with regards to covariates selected. Covariates masks are also calculated provided that predictions should not be made out of the range of the data.

statnmap/SDMSelect documentation built on April 1, 2021, 2:01 p.m.