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

were derived from the WFDEI Near Surface Temperature data for the same period at 0.5 degrees resolution [36]. Top soil organic carbon was obtained from ISIRIC [37]. For future climatic conditions, the same climatic variables used in model fitting were derived from data on projected climatic conditions for Ghana. We used climate projections of the Inter-Sectoral Impact Model Inter-comparison Project (ISIMIP) for the period 2006 to 2016 (baseline) and for 2041 to 2050 (future). This climate data consisted of four different gen- eral circulation models (GCM) projections, namely GFDL-ESM2M [38, 39], HadGEM-ES2 [40], IPSL-CM5A-LR [41], and MIROC-ESM-CHEM [42]. These GCMs were chosen because they are available with bias-adjustment [36, 43]. For future projections, the RCP2.6 and RCP8.5 scenarios were selected to represent the 1.5–2˚C-target of the Paris Agreement and a scenario without climate policy, respectively, to capture the range of climatic possibilities (Table 2). The modelling for climate impact assessment were run on the assumption of no change in soil organic carbon as there are currently no spatial near-future projections for this variable. All these variables were clipped to Ghana and then scaled, ensuring that they have a matching spatial resolution and extent. The same set of responses, predictors and scenarios were used for each crop and each scenario. 2.4 Modelling approach Suitability models or their variants have been used in assessing the geography of crop suitabil- ity and in modelling impacts of climate change on agriculture for different crops. While the

common approach is to use a 2 class (suitable/unsuitable) approach for modelling crop suit- ability [44–47], we propose a method that models four suitability classes (optimal, moderate, marginal and limited) as a 2 class system may over-estimate climate impacts by not scaling the suitability. Scaled four-class (high, moderate, marginal and unsuitable) suitability models are an alternative for determining suitability classes of agricultural crops from machine learning algorithms [31, 48–51]. To model the four suitability classes of the four crops, we applied the eXtreme Gradient Boosting (XGBoost) machine learning approach to the variables. XGBoost is an improvement of the recursive tree-based partitioning method of gradient boosting machines (GBM) by Friedman [52]. Gradient boosting is a technique implemented in a com- plex prediction model by iterative combinations of ensembles of weak prediction models into Table 2. Projected rainfall and temperature changes as used in the suitability modelling under the RCP2.6 and RCP8.5 for Ghana. Variables are summarized across the country. Scenario Model Sum rainfall Mar -Sept(mm) Sum rain sowing month (mm) Rainfall coefficient of variation (%) Rainfall growing season (mm) Average Tmax —Tmin (T˚C) Mean temperature growing season (T˚C) Mean temperature Mar-Sept (T˚C) Current Current 1246 228 69 558 9.9 24.5 25.7 RCP2.6 GFDL +52 +2 -2 +30 -0.4 +1.5 +1.3 IPSL -43 +3 +3 -14 -0.2 +1.8 +1.5 HADGEM -4 +10 +3 +14 -0.3 +2.2 +1.4 MIROC +58 +21 +6 +41 -0.4 +1.1 +1.3 Model mean

+16 +9 +3 +18 -0.3 +1.7 +1.4 RCP8.5 GFDL +59 +27 +3 +40 -0.7 +2.1 +2.5 IPSL -117 -3 +2 -29 -0.1 +2.8 +2.6 HADGEM -13 +10 +3 +7 -0.5 +1.5 +2.5 MIROC +86 +21 +2 +51 -0.3 +2.3 +1.8 Model mean +3.8 +13.9 +2.8 +17.5 -0.4 +2.2 +2.3 https://doi.org/10.1371/journal.pone.0229881.t002 PLOS ONE Climate change and food crops suitability in Ghana PLOS ONE | https://doi.org/10.1371/journal.pone.0229881 June 29, 2020 5 / 21