Sensitivity analysis of species distribution models to environmental variables from a GIS-perspective

Organised by Laboratory of Geo-information Science and Remote Sensing

Wed 29 August 2018 15:30 to 16:00

Venue Atlas, gebouwnummer 104
Room 1

By Lieven Slenders (the Netherlands)

Species Distribution Models are extensively used in a variety of studies ranging from disturbance and habitat loss to invasive and endangered species distribution mapping. These models are extremely suitable to predict the potential geographical distribution of a species by combining the species occurrence locations with geo-data. As more and more geo-data is becoming available, researchers are using more detailed model predictors for their species distribution models. However, despite an abundance of geo-data, it is often extremely difficult for ecologists to find environmental input data that is already tailored to the model they want to use. Since model predictors can largely affect the outcomes of the model it is highly important to consider the effects that these predictors have on the predictive performance of the model. Therefore, the main objective of this study was to evaluate the effect of the level of detail of predictors on SDM predictive performance (AUC). The predictors were altered into scenarios of six different measures of resolution. Then, used for model input, resulting in six model scenarios. Each scenario was trained on seven folds to account for abnormalities. Thus, a total of 42 distribution rasters was created. Comparing these model distribution rasters we did not find any relation between the resolution and the predictive performance. However, we only tested a limited range of resolution sizes, more repetitions and a wider range towards finer resolutions may give more conclusive results. Therefore, we suggest repeating this research with a targeted background sample and a broader range of resolutions. Here we conclude that, on basis of our simulations, a finer resolution of predictors cannot be statistically associated with a higher predictive performance, suggesting that a finer resolution (up to 350 meters) as input in species distribution models does not improve predictive performance. Therefore, our general advice to ecologists is to use the finest resolution possible in which predictors are available considering that the error on species location (uncertainty) should be smaller than the finest resolution for the environmental data.

Keywords: Species Distribution Models; Maxent; Predictor analysis; Resolution; Big data; Predictive performance; Sensitivity analysis; AUC; Geographical consistency; Veracity; SDM