SUPPLEMENTARY MATERIALS
Performance and uncertainty of crop suitability prediction
For model validation and uncertainty measures we adopted a similar approach to that of Ramirez-Villegas (2010). Using 294 evidence points, 30 MAXENT runs were performed, each using 80% of the points (235) for model training and the remaining 20% for model testing. Default settings were used in MAXENT so that the complexity of the model varied depending upon the number of data points used for model fitting. Two measures of model skill were used: the area under the receiver operating characteristic curve (AUC; Peterson et al., 2008) and the maximum possible Cohen’s kappa (kmax; Manel et al., 2001). Both measures were calculated using a fixed background area (i.e. the whole area of Ghana and Côte d'Ivoire) from which 542 random pseudo-absences were drawn [i.e. fixed-area AUC; VanDerWal et al. (2009)]. These two measures were used because of the potential caveats that can arise from the use of AUC as the only model evaluation measure (Lobo et al., 2008). Using these 30 model runs, a baseline was projected onto the 30 arc-sec grids of WorldClim. This produced a total of 30 equally-plausible suitability predictions for the baseline.
The performance of the MAXENT model was generally high, with AUC (kappa) values ranging between 0.990 (0.907) and 0.999 (0.979) for the test data (20%), and almost no variation for the train data (80%) (Fig. S1). The high performance of the model also produced relatively low baseline uncertainties. These uncertainties are mostly caused by model parameters (i.e. a slightly different MAXENT regression model is generally obtained for each of the duplicates) and by the locations of input evidence data.
=> Figure S1: near here
In order to calculate the uncertainty among GCMs to predict changes in crop suitability, we performed a suitability prediction using current conditions (WorldClim) and future conditions (19 GCM), then for each predicted conditions, the change in suitability was calculated on a pixel basis and the ensemble mean, 25th and 75th percentiles were calculated as measures of suitability impact. The standard deviation among GCMs was also calculated (Ramirez-Villegas et al., 2011). The average predicted changes in suitability have been discussed extensively in the main paper. The average suitability prediction using the first quartile of GCMs, these are the GCM’s predicting a more drastic change in climate, show an overall much more negative impact of climate change. All the cocoa areas will become significantly less suitable by 2050. The third quartile of GCM’s show a less significant decrease in suitability in the main growing areas and even some increase in suitability for southern Ghana and western Côte d’Ivoire.
Figure S2: near here
References
Lobo, J.M., Jiménez-Valverde, A. and Real, R., 2008. AUC: a misleading measure of the performance of predictive distribution models. Global Ecology and Biogeography, 17(2): 145-151.
Manel, S., Williams, H.C. and Ormerod, S.J., 2001. Evaluating presence–absence models in ecology: the need to account for prevalence. Journal of Applied Ecology, 38(5): 921-931.
Peterson, A.T., Papes, M. and Soberón, J., 2008. Rethinking receiver operating characteristic analysis applications in ecological niche modeling. Ecological Modelling, 213(1): 63-72.
Phillips, S.J., Anderson, R.P. and Schapire, R.E. (2006), Maximum entropy modeling of species geographic distributions. Ecological Modelling 190:231-259.
Ramírez-Villegas J, Khoury C, Jarvis A, Debouck DG, Guarino L (2010) A Gap Analysis Methodology for Collecting Crop Genepools: A Case Study with Phaseolus Beans. PLoS ONE 5(10): e13497. doi:10.1371/journal.pone.0013497
Ramirez-Villegas J, Jarvis A, Läderach P., 2011. Empirical approachesfor assessing impacts of climate change on agriculture: the EcoCrop model and a case study with grain sorghum. Agricultural and Forest Meteorology,170: 67-78.
VanDerWal, J., Shoo, L.P., Graham, C. and Williams, S.E., 2009. Selecting pseudo-absence data for presence-only distribution modeling: How far should you stray from what you know? Ecological Modelling, 220(4): 589-594.
Figure S1: Predicted mean annual temperature, evapotranspiration, mean annual precipitation and cumulative dry month changes by 2050 according to nineteen GCM models (SRES A2) for Ghana and Côte d'Ivoire.
Figure S2: Performance of the MAXENT model across the 30 replicates: (A) AUC, and (B) maximum Cohen’s kappa. Thick black horizontal line shows the median, box extends to 25-75% and whiskers show 5-95% of the distributions. Both AUC and Cohen’s kappa show very high agreement between replicates and the difference between the test and training set is only minimal.
Figure S3: Predicted changes in cocoa suitability and uncertainties. The red tones represent a decrease and the green tones an increase of suitability. a) average of the first quartile, b) average of the 19 GCMs, c) average of the third quartile of GCMs. d) Standard deviation of 19 GCM’s values represented in gray.
Figure S4: Relationship between cocoa suitability and altitude today (blue line) and in 2050 (red line) in Ghana and Côte d'Ivoire.
Red and yellow lines show projections before and after inclusion of potential evapotranspiration in the model, respectively. The green line shows land availability at different altitudes.