Incorporating geostatistical methods into the analysis of plant species distribution

Background/Question/Methods
Understanding the spatial distribution of species is an indispensable pre-requisite for their conservation. Although geostatistical models have been proven useful tools to modeling several environmental processes and their spatial characteristics, their use in biological sciences is rare. The main advantages of these models are their ability to describe and account for spatial dependence and using covariates as predictors. The main objective of this study is to investigate the usefulness of geostatistical methods to describe, quantify and predict the spatial dependence and distributional patterns of plant species. As a case study the species Combretum collinum Fresen. was chosen and modeled in West Africa. All species analyzed showed consistent results. Because our database consists of presence-only data, the Mahalanobis Distance algorithm was applied to identify poor suitable habitats and select pseudo-absences. To detect and describe the spatial structure of C. collinum, semivariograms together with prediction maps based on Indicator Kriging were calculated. A Generalized Linear Model (non-spatial) was used to estimate significant environmental variables and to predict the species distribution. Results were compared to predictions using Indicator Kriging with External Drift. The area under the curve (AUC) of the Receiver Operating Characteristic (ROC) curves was used to evaluate model performance.

Results/Conclusions
Considering only the spatial structure of collection localities, 93% of the variance in the distribution of C. collinum was explained by the fitted variogram. The absence of a nugget effect is evidence of the species’ non-random distributional pattern, which appears in a smooth and homogeneous fashion throughout the sudanian zone. The spatial dependence of collection records disappears at a distance of 600 km which is in concordance with the wide dispersion capabilities and range habitat of the species. The AUC value (0.94) for the Indicator Kriging prediction considering only spatial structure was highly significant (p < 0.0001) indicating better than random prediction. Predictions were calculated using Kriging with external drift by using the significant environmental variables and the model of residual spatial dependence. Model performance was again significantly better than a random prediction (AUC=0.79, p < 0.0001). Results from GLM predictions show a poor performance (AUC=0.67) although being significant (p < 0.01) and predicts species occurrence outside the species’ known range. Due to their higher performance values and better predictions within the known spatial range of C. collinum, geostatistical methods (i.e. variogram modeling and Kriging interpolation) appear to be useful techniques to describe, quantify and predict species distributional patterns.