...ibution to ecological niche and back.pdf

Evaluation exploring the SDMs’ performance behaviour more precisely !!. Spatial distribution of errors local evaluation of Comparison of different modelling techniques…!... 15 models

Babak Naimi, Thomas Groen, Nicholas Hamm,

Andrew Skidmore and Miguel B Araújo

A Summary Global Statistic

Would it be informative enough?

Relying on a single summary statistic to assess model performance may not be enough to fully understand whether a model really performed well!

AUC

AUC = 0.9

0%

100%

False Positive Rate

0

%

100%

TSS

1

0

Evaluation (exploring) the SDMs’ performance (behaviour) more precisely !!

Spatial distribution of errors ( local evaluation of

Comparison of different modelling techniques…!

10 species were simulated and mapped into Spain!

SI2=0.5 * (SI.L1.E1+SI.Ga.E2)

SI6=SI.L1.E1*SI.Ga.E2

15 models ( including ensemble approaches ) for presence/absence & presence-only data were used!

Global statistics

Local statistics

AUC TSS Sensitivity Specificity

Fuzzy map comparison Niche similarity

Model Uncertainty

Known truth! Predicted!

Known truth! Predicted!

Known truth BRT

Presence-only

GLM BRT

Presence-Absence

Ensemble_PO Ensemble

I

Model Uncertainty (between model inconsistency)

Presence Absence Presence Only

Model Uncertainty

Putting all together!!!!

the ensemble-based model strongly outperforms the other models either when accuracy or uncertainty in the predictions is a matter!

 The set of evaluation methods we

introduced can be useful for further investigations of model behaviours and provide deeper insights into the causes of

18