Assessing the impact of climate variability on maize yields in the different regions of Ghana—A machine learning perspective

observation emphasizes the stacking ensemble model’s capability to effectively manage data instances characterized by significant variability in these regions. However, due to the nature of the stacking ensemble learning model, where the deficiencies of underperforming models are compensated by the strengths of other models, the MSE of some base models tended to be slightly lower compared to the MSE of the stacking ensemble learning model in some regions. This emphasizes the importance of careful consideration when choosing the base models for the stacking ensemble approach. As indicated by the findings in Table 6, it becomes apparent that the stacking ensemble learning model excels in predicting maize yield based on diverse cli- matic variables across various regions in Ghana. Our findings corroborate the impactful per- formance of the Stacking ensemble model as documented in the study by LI et al. [35]. Moreover, our observations underscore its proficiency in effectively handling small datasets [34], reaffirming its practical utility in diverse scenarios. The stacking ensemble learning model has also proved its effectiveness in handling nonlinear data as suggested by Wang et al. [51]. Feature importance of the climate variables in predicting maize yield Maize yield in Ghana have been experiencing an upward trend as shown in Fig 2. This increas- ing trend can be attributed to the favorable climatic conditions and optimal policies in maize cultivation. Fig 4 illustrates the feature importance of climate variables in forecasting maize

yield. Since random forest performs better across all regions, we base our feature importance calculations for the climatic variables on the random forest Regressor. The statistics below show that CO2 is a key factor in predicting maize output throughout all of Ghana’s regions, Table 6. Performance metric of machine learning models. REGIONS METRIC MACHINE LEARNING MODELS LR DT RF SVM GB STACKING ASHANTI MSE 0.208 0.258 0.254 0.219 0.241 0.300 MAE 0.328 0.376 0.377 0.336 0.357 0.497 EASTERN MSE 0.051 0.041 0.049 0.053 0.044 0.052 MAE 0.148 0.133 0.145 0.155 0.135 0.143 VOLTA MSE 0.148 0.170 0.152 0.145 0.135 0.047 MAE 0.258 0.280 0.262 0.272 0.238 0.281 WESTERN MSE 0.028 0.028 0.027 0.026 0.027 0.025 MAE 0.116 0.118 0.112 0.111 0.115 0.109 CENTRAL MSE 0.244 0.214 0.153 0.178 0.148 0.200 MAE 0.381 0.374 0.307 0.350 0.311 0.401 NORTHERN MSE 0.085 0.088 0.083 0.082 0.085 0.083 MAE 0.220 0.226 0.222 0.220 0.225 0.224 UPPER WEST MSE 0.160 0.278 0.143 0.136 0.157 0.133 MAE 0.332 0.434 0.299 0.279 0.328 0.270 UPPER EAST MSE 2.259 2.915 1.268 0.953 1.178 0.866 MAE 1.080 1.328 0.818 0.725 0.865 0.725 BRONG-AHAFO MSE 0.209 0.282 0.211 0.165 0.232 0.215 MAE 0.441 0.515 0.443 0.359 0.449 0.410 GREATER ACCRA MSE 0.397 0.079 0.090 0.077 0.085 0.065 MAE 0.500 0.225 0.246 0.245 0.246 0.244 Note: The numbers in bold indicate the best performing model in the particular region. https://doi.org/10.1371/journal.pone.0305762.t006 PLOS ONE

Impact of climate variability on maize yields in Ghana PLOS ONE | https://doi.org/10.1371/journal.pone.0305762 June 25, 2024 17 / 24