University of Twente Student Theses


Factors affecting variable importance estimations from species distribution models: A virtual ecologist approach

Harisena, Nivedita Varma (2019) Factors affecting variable importance estimations from species distribution models: A virtual ecologist approach.

[img] PDF
Abstract:Aim: The aim of this study was toinvestigate three factors of ‘spatial dependency’ in input parameters namely: 1) spatial autocorrelation in predictor variables; 2) types of species response curve geometries; 3) varying sampling densities and analysetheir effect on model-independentvariable importance estimations derived from species distribution models. Method:This study uses simulated data of both the environmental predictors and the species response curves. Twenty-fivelevels of spatial autocorrelation(SAC)in combination with four types of species based on response types (linear, unimodal and combinations of these) and three levels of sampling density were analysed. The simulations were also run for two scenarios of relative SAC (0% background and 12.5% background).The choice of models includes eight models: Generalised Linear model (GLM); Generalised Linear Mixed Model (GLMM) with spatial random effect; Generalised additive model (GAM); Maximum Entropy (MaxEnt); Random Forest (RF); Boosted Regression Trees (BRT); Support Vector Machines (SVM) and Artificial neural networks (ANN). From each of the models,the variable importance (based on a model-independentrandomisationtechnique) is calculatedon an independent simulated test dataset, along with the reporting of the area under the Receivers Operating Characteristic curve, kappa and the autocorrelation in the residuals.Results:The results showed thatfor all fourspeciesall the models estimated biased importance towards the highly autocorrelated predictors, but the magnitude of bias was higher for the linear response species.The threshold relative SAC within which there was no bias was also narrower for the said species.Another significant result is the importancebias towards the unimodal responseswhen compared to a linear responsefrom all the models. Changing sampling densities did not have an observable effect. The RF and SVM were the most robust amongst all the models.Main conclusions:The type of response curve geometry, which are mainly dictated by species characteristics (i.e. narrow or wide-ranging species),along with the relative SACof the covariates werethe most determining factorsfor bias in relative variable importance estimations due to spatial autocorrelation in predictors. Therefore, species response characteristics along with the relative spatial structure of predictor variables must be givendue consideration for making proper inferences about variable importance from species distribution models.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page