Species Distribution Modelling of Family Sapindaceae in West Africa

MATERIALS AND METHODS

Species data: A total of 153 Sapindaceae growing localities (without duplicates) were assembled from diverse sources. About 32% of the localities came from recent fieldwork while 68% were herbarium records. In order to reduce the effect of temporal bias and to match the species data with the environmental climatic data (produced over the period 1950-2000), herbarium records collected before 1950 were eliminated.

Environmental data: Growing potential was assessed by characterising the climate at the presence locations. Twenty spatial datasets from the WorldClim database (at 5' resolution) (Hijmans et al., 2005) were used. The Worldclim dataset included altitude and 19 bioclimatic variables derived from temperature and rainfall. Although 21 spatial datasets were used in the beginning, some were eliminated during modelling due to their low contribution to the final model. In agricultural systems, abiotic factors are more limiting than biotic factors (such as competition or dispersal); therefore no biotic factors were included in this study. Also, it was assumed that at a continental scale, climatic variables were much more limiting than human population.

Species distribution modelling: For each plant sample, GPS coordinates of the collection sites were recorded and subjected to modelling using the Maximum Entropy Programme (MAXENT) (Phillips and Dudik, 2008). Maxent predicts the potential species distribution by estimating the probability distribution of maximum entropy across a specified region, subject to a set of constraints that represent the missing information (lack of absence data) about the target distribution (Phillips et al., 2006). The Maxent method is currently considered to be the most accurate approach to modelling presence-only data (Elith et al., 2006; Pearson et al., 2007).

Model validation: Model performance was evaluated using several methods. First, model performance was determined by assigning a subset of the presence records for training and using the remaining records to test the resulting model. A good model should predict correctly the presence of Sapindaceae in the test locations. As performance can vary depending upon the particular set of data selected for building the model and for testing it, 10 random partitions of the presence records were made to assess the average behaviour of Maxent, following Phillips et al. (2006). Each partition was created by randomly selecting 75% of the total presence records to build the model and the remaining 25% of presence records were used for testing. However, the full set of presence records were used to build the final model to obtain the best estimate of the species distribution (for all records model, West Africa model). Second, Receiver Operating Characteristic (ROC) analysis was used to evaluate how well the Maxent model compared to a random prediction. The area under the ROC curve (AUC) serves as a measure of model performance in terms of sensitivity versus specificity. The sensitivity for a particular threshold is the fraction of all positive instances that are classified as present and specificity is the fraction of all negative instances that are classified as not present. The value of the AUC is typically between 0.5 (random) and 1.0. The closer the AUC value to 1, the better the model performance. Moreover, the success of the model was also evaluated by visually examining how well the mapped probability values matched the presence records. A good model should produce regions of high probability that cover the majority of presence records and areas of low probability should contain few or no presence points.

RESULTS

Potential distribution of Sapindaceae in West Africa: All records model, the resulting Maxent cultivation potential map for Africa predicted members of Sapindaceae to occur in most of West Africa. For the West Africa model, the cultivation potential map predicted Sapindaceae to grow in most of the Sahel from Senegal to south-west Cameroon and in southern Ghana-Nigeria (Fig. 1).

Model performance: All the ten generated training/testing models showed a high level of performance when compared to random (where the AUC would be 0.5). Training gained in this model ranged from 2.139 - 2.668 (Fig. 2) while the AUC value was 0.974 (Fig. 3). Test AUC values was lower but close to training AUC i.e., 0.960 (Fig. 3) while the test gained was 2.117 (Fig. 4). The training/testing models correctly predicted most of the test locations in all models and standard deviation is 0.007.

Variable contribution: The model indicated that the presence of members of Sapindaceae was mainly associated with minimum temperature of the coldest month (35.4%), precipitation of the coldest quarter (16.9%) and precipitation of wettest month (14.9%) (Table 1).

Fig. 1:	Species distribution: Training vs. test locations, Green: The typical condition of where species are found, Blue: Low probability of suitable condition, Red: High probability of suitable condition, White dots: The presence locations used for training, while violet dots show test locations

Fig. 2:	Jack-knife of regularized training gain for Sapindaceae

Fig. 3:	Sensitivity vs. specificity test on Sapindaceae

Maxent’s jackknife test of variable importance also suggested that the variables which contributed the most to the model were good predictors as they had the most information not contained in other variables and they could describe the distribution of Sapindaceae on their own (without the other variables).

Table 1:	Variable contributions to the MAXENT model

Fig. 4:	Jack-knife of regularized test gain for Sapindaceae

DISCUSSION

Although members of the family Sapindaceae have been recorded to be widely distributed in Africa, their occurrence is being threatened by high rate of deforestation and agricultural practices leading to loss of forest and threatened status of the family as recorded in the IUCN R.L. (2008). MAXENT, a machine-learning method based on the principle of maximum entropy, are used to predict distribution for each species of butterfly under current and future climates (Sofy et al., 2012). Data from this model is used to test the efficiency of the modeling program through the evaluation of the area under the Receiving Operator Curve (AUC) (Elith et al., 2006; Phillips et al., 2006).

Species distribution modelling with climatological data analysis using nineteen (19) standard environmental variables from the WorldClim database revealed that the species distribution is most affected by the minimum temperature of the coldest month (bio 6) with a variable contribution of 35.4%. This is followed by precipitation of coldest quarter (bio 19) and precipitation of the wettest month (bio 13) with 16.9 and 14.9% contribution, respectively. These results support the location of Sapindaceae mostly in the tropics rather than the temperate region.

Also, from the AUC curve generated, it can be concluded that this present model is more reliable as it predicts the species distribution better than a typical random model as shown in Fig. 3. This makes it useful for future projections. Recommendations for different conservation strategies include in situ conservation in Protected Areas; ex situ conservation in seed banks; and conservation through ‘sustainable utilization’.

ACKNOWLEDGMENTS

Special thanks to the Head of Molecular Systematics Laboratory, Department of Botany, University of Lagos for providing enabling environment for this study. Also, we are grateful to the Explorer’s Club USA for providing the exploration funds.