2.4. Ecological predictive models’ development and evaluation
To calibrate our models, we employed the maximum entropy method (Elith et al. 2010; Phillips et al. 2005) implemented in Maxent ver. 3.4.4. The algorithm has been extensively tested and benchmarked (Phillips & Dudík 2008; Richards et al. 2007). Many studies have reported Maxent as one of the highest performing presence-background algorithms (Elithet al. 2006; Merow et al.2013). As the selection of sample points can influence model performance in Maxent (Phillips et al. 2009), we restricted the selection of background points using a regularization of 10,000 background points (Elith et al. 2006). For simplification of the modeling algorithms, we used the default settings (feature class and regularization) in Maxent for each of the three models. Models were trained with data from the present and projected in the future. Three Maxent models were generated with different occurrence datasets: first, all accessions location points grouped together without genetic information were used in projecting the entire distribution of KG (model 1); model 2 and model 3 were developed using separately occurrence data of Pop 1 and Pop 2, the genetically defined populations (Supplemental Table S1).
We used a tenfold cross-validation method, which uses 90% of the data for model training and 10% for model testing for 10 iterations (Elith et al. 2010).
Each model performance was evaluated using traditional Receiver Operator Characteristics (ROC) - area under the curve (AUC) scores (Merow et al. 2013) by specifying 500 iterations with the omission threshold set at ten percent (Peterson et al. 2008). A model is considered as having a good fit when its AUC is close to one (AUC≥0.75) (Elith et al. 2006). Minimum training presence (MTP) values were also used as thresholds for testing the performance of each model (Phillipset al. 2005).
The outputs from Maxent were processed in ArcGIS ver. 10.7.1 to construct maps of the distribution of Kersting’s groundnut areas cultivability. The continuous probabilities generated by Maxent (Ten Percentile Training Presence) were converted into binary presence–absence maps to identify the levels of areas suitability. Two different levels were therefore, defined: unsuitable and suitable. Finally, we quantified the dynamic of the cultivated zones of the crop in the scenarios RCP 4.5 and RCP 8.5 of the horizon 2055 using the following equation:
\begin{equation} \bigtriangleup(\%)\ =\frac{\left(\text{FA}_{\text{ij}}-\text{CA}_{j}\right)*100}{\text{TA}}\nonumber \\ \end{equation}
Where, FAij corresponds to the extent value (in number of pixels) under the scenario i of future horizon in the environment j (area suitability); CA is the extent value of current condition; TA corresponds to the total extend of all cultivated zones of the present day. Negative, null and positive values represent range lost, stable and gained, respectively.
Furthermore, to visualize the potential changes of suitable areas for Kersting’s groundnut production, we compared current and future distribution ranges of the crop and of genetic populations using package “tmap” version 3.3-1 (Tennekes 2018) in R.