Developing a toolbox to efficiently handle SDMs in collaborative projects. Package aims at providing varied and flexible model fitting and tools for SDMs. Meta information should be stored (author, date, taxon,...) efficiently. Model evaluation and prediction should flexibly handle various fitted model objects. Evaluation metrics for presence-only and presence-absence models should be mutually implemented. Filtering and testing tools should help the user to find adequate predictors to apply meaningful models. Poisson Point Process models (PPPM) implementation should remain user-friendly. Regularization and variable selection should be implemented in this framework. Efficient methods to correct sampling bias in model fitting should be implements (pseudo-absences sampling strategies, environmental bias correction, bias covariate correction...)
You can install the development version from GitHub with:
remotes::install_github("8Ginette8/wsl.biodiv")
library(wsl.biodiv)Load data:
# Take anguilla data set from dismo package
data("Anguilla_train")
vrs = c("SegSumT","USRainDays","USSlope")
env = Anguilla_train[,vrs]
# Alps data
data(AlpineConvention_lonlat)
data(exrst)
rst = rst[[1:6]]
data(xy_ppm)
mypoints = xy.ppm[,c("x","y")]Custom parameters for the ensemble model:
# Formulas
form.glm = as.formula(paste("Presence~",paste(paste0("poly(",vrs,",2)"),collapse="+")))
form.gam = as.formula(paste("Presence~",paste(paste0("s(",vrs,")"),collapse="+")))
form.gbm = as.formula(Presence ~ .)
feat = c("linear=true","quadratic=true","hinge=true","product=true","threshold=false")
# All options
modinp = list(multi("glm",list(formula=form.glm,family="binomial"),"glm-simple",step=TRUE,weight=TRUE),
multi("gbm",list(formula=form.gbm,distribution = "bernoulli",interaction.depth = 1,shrinkage=.01,n.trees = 3500),"gbm-simple"),
multi("gam",list(formula=form.gam,family="binomial"),"gam-simple",step=FALSE,weight=TRUE),
multi("maxent",list(args=feat),"mxe-simple"),
multi("randomForest",list(formula=form.gbm,ntree=500,maxnodes=NULL),"waud1"))Calibrate ensemble model:
modi5 = wsl.flex(pa=Anguilla_train$Angaus,
env_vars = env,
taxon = "Angaus",
replicatetype = "cv",
reps = 3,
strata = sample(1:3,nrow(env),replace=TRUE),
project = "multitest",
mod_args = modinp)
summary(modi5)Evaluate and display results:
# Evaluate the model
eval5 = wsl.evaluate.pa(modi5,crit = "maxTSS")
# Get outputs or evaluation summary
eval5
summary(eval5)
# Get thresholds
thr.5 = get_thres(eval5, mean=FALSE)Let's predict now (also works when 'predat' is a raster):
# Make some predictions (works also with Raster objects)
pred4 = wsl.predict.pa(modi5,predat=env)
pred5 = wsl.predict.pa(modi5,predat=env,thres=thr.5)Define a mask of your study area, to set a window and sample quadrature points:
# Define mask
maskR = mask(rst[[1]],shp.lonlat)
# Run 'wsl.ppm.window' function
wind = wsl.ppm.window(mask = maskR,
val = 1,
owin = TRUE)
# nDefine quadrature points for 'wsl.ppmGlasso'
quadG1 = wsl.quadrature(mask = maskR,
area.win = wind,
random = FALSE,
lasso = TRUE,
env_vars = rst)
# Define your environments
envG = raster::extract(rst,mypoints)# Spatial block cross-validation
to_b_xy = rbind(mypoints,quadG1@coords)
toSamp = c(rep(1,nrow(mypoints)),rep(0,nrow(quadG1@coords)))
block_cv_xy = make_blocks(nstrat = 5, df = to_b_xy, nclusters = 10, pres = toSamp)
# Environmental block cross-validation
to_b_env = rbind(envG,quadG1@Qenv[,-1])
block_cv_env = make_blocks(nstrat = 5, df = to_b_env, nclusters = 10, pres = toSamp)Here, we are basically applying environmental k-medoid clustering to the obervations, and k-nearest neighbors (relative to the observation clusters) to the quadrature points.
Fit a PPM with an Elastic Net regularization:
ppm.lasso = wsl.ppmGlasso(pres = mypoints,
quadPoints = quadG1,
asurface = raster::area(shp.lonlat)/1000,
env_vars = envG,
taxon = "species_eg1",
replicatetype = "cv",
reps = 5,
strata = NA,
save=FALSE,
project = "lasso_eg1",
path = NA,
poly = TRUE,
lasso = TRUE,
alpha = 0.5, # 0.5 = Elastic Net here
type.measure = "mse", # Other regularization parameters
standardize = TRUE, # ....
nfolds = 5, # ....
nlambda = 100) # ....see cv.glmnet() for more details
summary(ppm.lasso)Fit a simple PPM (without any regularization, nor polynomial terms) using environmental block cross-validation:
ppm.simple = wsl.ppmGlasso(pres = mypoints,
quadPoints = quadG1,
asurface = raster::area(shp.lonlat)/1000,
env_vars = envG,
taxon = "species_eg2",
replicatetype = "block-cv",
reps = 5,
strata = block_cv_env,
save = FALSE,
project = "lasso_eg2",
path = NA,
poly = FALSE,
lasso = FALSE)
summary(ppm.simple)Evaluation example using Boyce index:
eval.lasso = wsl.evaluate.pres(x = ppm.lasso,
env_vars = rst,
speedup = TRUE)
summary(eval.lasso)Evaluation example using other binary metrics:
eval.simple = wsl.evaluate.pa(x = ppm.simple,
crit = "maxTSS",
pres_only = TRUE)
summary(eval.simple)Evaluation example by resetting a potential fitted bias covariate to 0 values (e.g. to remove sampling bias):
eval.bias = wsl.evaluate.pa(x = ppm.simple,
crit = "maxTSS",
pres_only = TRUE,
bias_cov = c(1,1,1,1,1,0))
summary(eval.bias)Get calculated thresholds (mean may be chosen):
get_thres(eval.lasso, mean = FALSE)
get_thres(eval.simple, mean = TRUE)
get_thres(eval.bias, mean = FALSE)Now we can predict:
# e.g. without using the thresholds --> species 'abundances'
pred.lasso = wsl.predict.pres(x = ppm.lasso,
predat = rst,
raster = TRUE)
# e.g. using the thresholds --> species presences/absences
pred.simple = wsl.predict.pres(x = ppm.simple,
predat = rst,
thres = get_thres(eval.simple,mean=FALSE),
raster = TRUE)
# e.g. or resetting a potential fitted bias covariate to 0 values:
pred.bias = wsl.predict.pres(x = ppm.simple,
predat = rst,
thres = get_thres(eval.bias,mean=FALSE),
raster = TRUE,
bias_cov = c(1,1,1,1,1,0))Finally let's see what distributions we obtain across the European Alps by plotting:
par(mfrow=c(2,3))
sapply(1:5,function(x) plot(pred.lasso@predictions[[x]][[1]]))
par(mfrow=c(2,3))
sapply(1:5,function(x) plot(pred.simple@predictions[[x]][[1]]))
par(mfrow=c(2,3))
sapply(1:5,function(x) plot(pred.bias@predictions[[x]][[1]]))
Brun, P., Thuiller, W., Chauvier, Y., Pellissier, L., Wüest, R. O., Wang, Z., & Zimmermann, N. E. (2020). Model complexity affects species distribution projections under climate change. Journal of Biogeography, 47(1), 130-142. doi: 10.1111/jbi.13734
Chauvier, Y., Thuiller, W., Brun, P., Lavergne, S., Descombes, P., Karger, D. N., ... & Zimmermann, N. E. (2021). Influence of climate, soil, and land cover on plant species distribution in the European Alps. Ecological monographs, 91(2), e01433. doi: 10.1002/ecm.1433
Descombes, P., Chauvier, Y., Brun, P., Righetti, D., Wüest, R. O., Karger, D. N., ... & Zimmermann, N. E. (2022). Strategies for sampling pseudo-absences for species distribution models in complex mountainous terrain. bioRxiv, 2022-03. doi: 10.1101/2022.03.24.485693