HeleneBlt
commented
6 months ago Hi,
It seems you have some overfitting for RF and GAM.
You can try
- to increase nodesize :
myBiomodOptions <- BIOMOD_ModelingOptions(RF = list(nodesize = 10)) - to try different maxnodes. For example :
myBiomodOptions <- BIOMOD_ModelingOptions(RF = list(nodesize = 10, maxnodes = 5)) - to balance RF with a downsampling. You can have a look to the issue #393.
For GAM, you can play with the smoothness of the curve. You can switch to gam with the package mgcv if it is easier for you with BIOMOD_ModelingOptions(algo = 'GAM_mgcv' )
Please let me know if you think this is not a case of overfitting or if you have any other questions.
Hélène
Hello, I'm trying to do a species distribution modeling with BIOMOD2 package. I am encountering a problem regarding TSS value, more specifically the TSS value of RF and GAM are 1 or close to 1 and the TSS of ensemble model is lesser than individual model. I dont know where the problem is. My study area is 650 sq km. I have 110 presence points and am generating 300 PA (though have tried with multiple PA) myRespName<- 'Presence' myResp <- as.numeric(data[,myRespName]) myRespXY <- data[,c("X","Y")] myBiomodData <- BIOMOD_FormatingData(
resp.var = data["Presence"], resp.xy = data[, c('Y', 'X')], expl.var = myExpl, resp.name = "Presence", PA.nb.rep = 3, PA.nb.absences = 500, PA.strategy = 'random' ) myBiomodOption <- BIOMOD_ModelingOptions() myBiomodModelOut <- BIOMOD_Modeling(bm.format = myBiomodData, models = c('GLM','GBM','GAM','CTA','ANN','FDA','MARS','SRE','RF', 'MAXENT' ), bm.options = myBiomodOption, var.import = 1, CV.strategy = 'random', CV.nb.rep = 3, CV.perc = 0.7, prevalence=NULL, metric.eval = c('TSS')) myBiomodModelEval <- get_evaluations (myBiomodModelOut) myBiomodModelEval myBiomodEM <- BIOMOD_EnsembleModeling(bm.mod = myBiomodModelOut, models.chosen = 'all', em.by = "all", em.algo = c('EMmean'), metric.select = c('TSS'), metric.select.thresh = c(0.6), metric.eval = c('TSS'), var.import = 2, EMci.alpha = 0.05, EMwmean.decay = 'proportional' ) I dont know if the problem is with code, modeling or data itself.