Impacts of climate change on agro-climatic suitability of major food crops in Ghana

model predictions by comparing it with the accuracy expected to occur by chance with k values ranging from -1 (poor) to 1 (good) [22]. The multiclass area under receiver operating charac- teristic curve (AUC) was used to validate model fit by comparing and averaging all pairwise class AUC. Sensitivity for suitability class is the percentage of a category on the reference data that is correctly modelled as belonging to that category, and measures proportion of pixels omitted from a reference suitability class (omission error). Specificity expresses the proportion of a category on the reference data that is included erroneously in another suitability class (commission error) [67]. Other metrics used for class accuracy are described in full in litera- ture [68, 69]. This analysis was done in R v.3.3 (R Core Team, 2013). 2.7 Assessment of climatic suitability for cultivation of multiple crops In order to determine the climatic suitability for cultivation of the four key food crops for Ghana, we combined the suitability of the crops to understand which areas are suitable for which crops and to what degree. At first the maps were stacked to determine the number of crops that were suitable for each cell. To determine suitability for multiple crops, we summed the modelled crop suitability with each class ranked from 1 (limited) to 4 (optimal). This pro- duced climatic suitability scale crops for the four crops on a scale from 4 (very low) to 16 (very high). After that, realizing that suitability of two-crops was most frequently observed, further

analysis of determining which pairs of crops were suitable at each of the pixels. Pairs of crops were summed to produce a potential between 2 (very low) to 8 (very high). Changes in suit- ability proportion and distribution between the current and the projected climatic conditions were assessed by counting and comparing areas and proportions of cells between times and scenarios. 3. Results 3.1 Model performance evaluation To reliably assess crop suitability, we first evaluated the fit of the model on an independent test data set. There were differences in the model fit between crops, but all crops showed a good fit. The best accuracy was for modelling sorghum (OA = 0.82, k = 0.75 and AUC = 0.87) (see Table 3). Modelling evaluation metrics for each of the four suitability classes such as sensitivity, specificity, positive and negative prediction values, precision, recall, F1-score, prevalence, detection rate, detection prevalence and balanced accuracy are shown in S1 Table. These accu- racy metrics indicated that the model was able to match observed classes for all the crops and thus could be used with confidence in assessing suitability of the four crops in Ghana under current and future climate. 3.2 Contribution of variables to crop suitability The relative contribution of each variable to modelling crop suitability was determined by ana- lyzing the importance of each variable to the model. The percent contributions of each variable

Table 3. Overall accuracy, kappa and multi-class AUC values as indicators model performance for maize, sorghum, groundnut and cassava in determining crop suitability classes in Ghana. Crop Accuracy Kappa Multiclass-AUC Maize 0.75 0.66 0.81 Sorghum 0.82 0.75 0.87 Groundnuts 0.75 0.66 0.85 Cassava 0.71 0.59 0.84 https://doi.org/10.1371/journal.pone.0229881.t003 PLOS ONE Climate change and food crops suitability in Ghana PLOS ONE | https://doi.org/10.1371/journal.pone.0229881 June 29, 2020 7 / 21