README

The goal of PointedSDMs is to simplify the construction of integrated species distribution models (ISDMs) for large collections of heterogeneous data. It does so by building wrapper functions around inlabru, which uses the INLA methodology to estimate a class of latent Gaussian models.

Installation

# install.packages("devtools")
devtools::install_github("PhilipMostert/PointedSDMs")

Package functionality

PointedSDMs includes a selection of functions used to streamline the construction of ISDMs as well and perform model cross-validation. The core functions of the package are:

Function name	Function description
`intModel()`	Initialize and specify the components used in the integrated model.
`blockedCV()`	Perform spatial blocked cross-validation.
`runModel()`	Estimate and preform inference on the integrated model.
`datasetOut()`	Perform dataset-out cross-validation, which calculates the impact individual datasets have on the full model.

The function intModel() produces an R6 object, and as a result there are various slot functions available to further specify the components of the model. These slot functions include:

Example

`intModel()` slot function	Function description
`.$plot()`	Used to create a plot of the available data. The output of this function is an object of class `gg`.
`.$addBias()`	Add an additional spatial field to a dataset to account for sampling bias in unstructured datasets.
`.$updateFormula()`	Used to update a formula for a process. The idea is to start specify the full model with `intModel()`, and then thin components per dataset with this function.
`.$updateComponents()`	Change or add new components used by inlabru in the integrated model.
`.$priorsFixed()`	Change the specification of the prior distribution for the fixed effects in the model.
`.$specifySpatial()`	Specify the spatial field in the model using penalizing complexity (PC) priors.
`.$spatialBlock()`	Used to specify how the points are spatially blocked. Spatial cross-validation is subsequently performed using `blockedCV()`.

This is a basic example which shows you how to specify and run an integrated model, using three disparate datasets containing locations of the solitary tinamou (Tinamus solitarius).

library(PointedSDMs)
library(ggplot2)
library(raster)

#Load data in

data("SolitaryTinamou")

projection <- CRS("+proj=longlat +ellps=WGS84")

species <- SolitaryTinamou$datasets

Forest <- SolitaryTinamou$covariates$Forest

crs(Forest) <- projection

mesh <- SolitaryTinamou$mesh
mesh$crs <- projection

Setting up the model is done easily with intModel(), where we specify the required components of the model:

#Specify model -- here we run a model with one spatial covariate and a shared spatial field

model <- intModel(species, spatialCovariates = Forest, Coordinates = c('X', 'Y'),
                 Projection = projection, Mesh = mesh, responsePA = 'Present')

We can also make a quick plot of where the species are located using `.$plot()`:

region <- SolitaryTinamou$region

model$plot(Boundary = FALSE) + gg(region)

#Run the integrated model

modelRun <- runModel(model, options = list(control.inla = list(int.strategy = 'eb')))
summary(modelRun)
#> Summary of 'bruSDM' object:
#> 
#> inlabru version: 2.5.2
#> INLA version: 22.05.18-2
#> 
#> Types of data modelled:
#>                                     
#> eBird                   Present only
#> Parks                Present absence
#> Gbif                    Present only
#> Time used:
#>     Pre = 1.74, Running = 17.1, Post = 0.0217, Total = 18.8 
#> Fixed effects:
#>                   mean    sd 0.025quant 0.5quant 0.975quant mode   kld
#> Forest          -0.002 0.001     -0.005   -0.002      0.000   NA 0.099
#> eBird_intercept -0.244 0.047     -0.336   -0.244     -0.152   NA 0.505
#> Parks_intercept -0.536 0.178     -0.892   -0.535     -0.191   NA 0.000
#> Gbif_intercept  -0.553 0.048     -0.647   -0.553     -0.459   NA 0.285
#> 
#> Random effects:
#>   Name     Model
#>     shared_spatial SPDE2 model
#> 
#> Model hyperparameters:
#>                            mean    sd 0.025quant 0.5quant 0.975quant mode
#> Theta1 for shared_spatial -3.39 0.004      -3.39    -3.39      -3.38   NA
#> Theta2 for shared_spatial -8.00 0.004      -8.01    -8.00      -7.99   NA
#> 
#> Deviance Information Criterion (DIC) ...............: 4204.54
#> Deviance Information Criterion (DIC, saturated) ....: -23256.48
#> Effective number of parameters .....................: 229.82
#> 
#> Watanabe-Akaike information criterion (WAIC) ...: 3088.50
#> Effective number of parameters .................: 653.03
#> 
#> Marginal log-Likelihood:  -3891.83 
#>  is computed 
#> Posterior summaries for the linear predictor and the fitted values are computed
#> (Posterior marginals needs also 'control.compute=list(return.marginals.predictor=TRUE)')