Hello all,

Thanks for your answers. In the meantime, I found the publication I was looking for :

Mao-Ning Tuanmu, Andrés Viña, Gary J. Roloff, Wei Liu, Zhiyun Ouyang, Hemin Zhang and Jianguo Liu 2011 Temporal transferability of wildlife habitat models: implications for habitat monitoring Journal of Biogeography, Volume 38, Issue 8, pages 1510–1523,

This paper introduced metrics to assess the spatial patterns of predicted habitat, which can be a useful complement to model accuracy metrics.

Cheers and thanks again,

Diederik

Hey Diederik,

this article by Mouton et al 2010 (Ecol. Modelling 221: 1995-2002) is also related to the topic (authors are from Belgium - pretty sure you know it already?).

http://izt.ciens.ucv.ve/ecologia/Archivos/ECO_POB%202010/ECOPO7_2010/Mouton%20et%20al%202010.pdf

Cheers

Darius

The two provided references are comparing different statistics; very useful, no doubt. But indeed with any set of spatial data you may generate numerous SDMs with the equivalent evaluation stats values (e.g. AUC) by different selections of your predictor data layers (see attached). However each SDM may have a more or less different output map in terms of spatial pattern and extent; and the differences in pattern is ecology relevant for conservation (attached). In other words, I join the question: how to compare patterns statistically.

Article Where the bears roam in Majella National Park, Italy

Hello all,

Thanks for your answers. In the meantime, I found the publication I was looking for :

Mao-Ning Tuanmu, Andrés Viña, Gary J. Roloff, Wei Liu, Zhiyun Ouyang, Hemin Zhang and Jianguo Liu 2011 Temporal transferability of wildlife habitat models: implications for habitat monitoring Journal of Biogeography, Volume 38, Issue 8, pages 1510–1523,

This paper introduced metrics to assess the spatial patterns of predicted habitat, which can be a useful complement to model accuracy metrics.

Cheers and thanks again,

Diederik

Hi Diederik,

A little late to the party, but you might also look into using Lin's coefficient of concordance. The CC statistic takes into account both the precision and accuracy of model predictions by measuring the covariation between expected and observed values and the degree that these relationships depart from a line with slope of 1 and y-intercept of zero. The nice thing about CC is it compares observed and predicted values, while accounting for habitat context by including the proportion of habitat in each predicted probability bin (ranging from 0 to 1). For some references:

Lin 1989. Biometrics 45: 255-268.

Lin 2000. Biometrics 56: 324-325.

Granadeiro et al. 2004. Journal of Sea Research 52: 227-240.

Johnson et al. 2006. Journal of Wildlife Management 70: 347-357.

Zeller et al. 2014. Landscape Ecology 29: 541-557.

Good luck,

Andrew

Andrew, many thanks for this answer and the references!

Cheers,

Diederik

Diederik, see another method for evaluating correspondence between model predictions in Martinez-Meyer et al. 2004 (GEB) and 2006 (JBI). This methods is based on binomial distribution.

Cheers

Matheus 

 In paper below, I used some technics based in confusion Matrix.

DE SOUZA, RODRIGO ANTÔNIO, DE MARCO, PAULO

The use of species distribution models to predict the spatial distribution of deforestation in the western Brazilian Amazon. Ecological Modelling. , v.291, p.250 - 259, 2014.

Hi Diederik.

Your concern might be more about the consistency (or efficacy) of model validation statistics. After all, having two models with similar cross-validated AUCs that make radically different extrapolations with highly different projection accuracies suggests that the original accuracy estimate (the CV AUC) is not useful. This might be more a problem of deficiencies in cross-validation approaches more than anything else (i.e. the CV AUC of the better-projecting model should be higher). Bahn & McGill looked at some stratified CV approaches which help generate a more accurate error estimate (AUC) from cross-validations, when models are intended for projection/extrapolation. Might be useful:

Bahn, V. and B. J. McGill (2013). "Testing the predictive performance of distribution models." Oikos 122(3): 321-331.