Distribution Modeling, evaluation and niche metrics
I have used BIOMOD2 package 3.4.11 (Thuiller et al. , 2009) in RStudio Version 1.2.5033 (R Core Team, 2020) to perform ensemble modeling using a combination of five algorithms. These are Generalized Additive Models (GAMs), Surface Range Envelope (SRE), Random Forest (RF), Generalized Boosting Model and Artificial Neural Network (ANN). GAM is a regression-based algorithm that uses smoothers which are a set of polynomials for generating curves by local fitting to the subsection of data. SRE is an envelope-style method that creates arrays of values for each environmental variable at the given presence points. RF is a bagging approach classifies data into different classes based on homogeneity and generates result from classification tree. Under the Generalized Boosting Model which is a boosting method, each of the individual models possesses regression trees. ANN is a mathematical technique which makes an attempt to simulate biological neural networks and are trained by back propagation algorithm. Unlike modeling through single algorithm, ensemble modeling are considered to produce better accuracy (Araujo and New, 2007).
The models were selected with True Kill Statistic (TSS) evaluation metric with a threshold of 0.7. Data partitioning of 80% training dataset and 20% testing dataset was used. The entire procedure was repeated three times. Each model was evaluated by ROC (Relative Operating Characteristic) (=AUC) and TSS metrics. The AUC has a range from 0 to 1 and TSS range from -1 to 1. As per the refined AUC scale, AUC > 0.9 are considered excellent, 0.80 < AUC < 0.90 are considered good and so on (Araújo et al. , 2005) whereas the TSS score +1 indicates perfect agreement and values of zero or less indicate a performance no better than random.
Niche overlap values for the taxa were calculated using Schoener’s D (Schoener, 1968) and Hellinger’s I niche similarity metrics (Warrenet al. , 2008b) using the overlap function in ’dismo’ package (Hijmans et al. , 2017). D and I qualitatively similar results although I metric tends to give higher values than D for any given comparison (Culumber and Tobler, 2016). The niche equivalency test was performed using ’phyloclim’ package (Heibl et al. , 2018) with 99 replicates. This test asks whether the niches of the taxa under study are effectively indistinguishable (Warren et al. , 2008a). The equivalency of two niches are rejected if the niche overlap falls outside the 95% of the null hypotheses. The niche equivalency test using this package was chosen over some of the other popular methods such as use of ’ecospat’ package as it uses occurrence density grids (Cola et al. , 2017). This is problematic in this case because the South Asian Dolphins do not share same geographical space and the models built would need to be projected into the species’ geographical distribution elsewhere.
Results