Climate suitability predictions for the cultivation of macadamia (Macadamia integrifolia) in Malawi using climate change scenarios

We evaluated the predictive accuracy of 18 algorithms of species distribution models (SDM) using a cross-validation technique in the first stage. The SDM algorithms used in our analysis were those that can distinguish between suitable and non-suitable areas without need- ing absence locations [35]. Following work by Brotons et al. [46] and Thuiller et al. [47], we divided the occurrence data into two distinct sets by randomly assigning 70% of the data as a training dataset to fit the model, and the remaining 30% were used as test data to evaluate the model’s predictive accuracy. A five-fold (partition) cross-validation replicate was performed in each of the model algorithms to evaluate the stability of the prediction accuracy as described by Rabara et al. [48] and Mudereri et al. [39]. Each SDM algorithm’s performance was evalu- ated from each partition separately after individual algorithms were assessed with data from the other four partitions. Cross-validation validates the performance of models and prevents overfitting, particularly in cases where the amount of data may be limited [39,49]. The area under the curve (AUC) criterion computed by the R package "PresenceAbsence" [50] was used to evaluate the performance of each algorithm. The AUC value is a specific mea- sure of model performance that demonstrates the model’s ability to locate a randomly chosen present observation in a cell with a higher probability than a randomly selected absence obser- vation [45,48]. Based on the recommendation by Kindt and Cole [42], we used an AUC value

of 0.77 as a threshold to select the best-performing algorithms for this analysis. SDM algo- rithms that did not meet this criterion were not used to calculate the final ensemble model’s suitability [51]. AUC values of 0.75 are considered reliable, 0.80 as good, and 0.9 to 1 as having excellent discriminating ability [52]. We utilized the presence-only approach for our study, and this is because, for agricultural applications of niche models, it is inappropriate to treat areas without current production as entirely unsuitable. Further, determining whether a species is absent in a specific location is difficult and rare, so absence data may not be a true representation of naturally occurring phe- nomena [53]. As an alternative, we randomly generated 500 background pseudo-absence points for our analysis. A caveat to this approach is the recommendations of Barbe-Massin et al. [33] regarding the use of lower pseudoabsences in some algorithms. Then, we combined these background pseudo-absence points with the 84 occurrence points "presence only" for the niche modelling of macadamia. The second step consisted of retaining only the algorithms that contributed at least 5% to the ensemble suitability (Se) [43]. This procedure generated AUC values for each and the parameters of the response functions (model training) to estimate the probability values of spe- cies occurrence based on the climate of each grid cell of the study area. The AUC values for the selected SDM algorithms are shown in Table 2. The results of all the models were then com-

bined by calculating for each the weighted average (weighted by AUC for each model) of the probability values from each model to generate the ensemble suitability map. The AUC values obtained by each algorithm were weighted using the following equation: Ensemble ðSeÞ ¼ P iWiSi P iWi ð2Þ Where the ensemble suitability (Se) is obtained as a weighted (w) average of suitabilities pre- dicted by the contributing algorithm (Si). The predicted suitable area for the probability of macadamia was calculated using threshold values, i.e., � 0.34 for the suitable area, while < 0.34 was regarded as unsuitable [39]. To gener- ate the probability maps, we used the maximum sensitivity (true positive+) and maximum specificity (true negative-) approach [54], where we reclassified the probability maps to a binary raster image (suitable/unsuitable areas). Then, using the Malawi shapefile in R, the PLOS ONE Macadamia integrifolia: Climate suitability of Malawi PLOS ONE | https://doi.org/10.1371/journal.pone.0257007 September 9, 2021 8 / 20