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

Machine learning models implementation Using 5-fold cross-validation and Grid Search, the results of the stacking ensemble learning model, which includes the optimal parameters of the base models (Gradient Boosting, Linear Regression, Decision Tree, Support Vector Machine, and Random Forest), are shown in Table 5. Table 5 provides a thorough summary of these ideal parameter values, showing the precise setups that produced the best results when used with the stacking ensemble building. Utilizing the unique capabilities of each base model, the stacking ensemble learning employs the prediction of the optimal outcomes of each base model to build the final prediction. Each region’s unique ideal parameters were achieved (see Table 5), and the optimal parameter for each region was identified. This will be useful in predicting maize yield based on climate variables. Results of the performance metrics for the models The statistics in Table 6 demonstrate that, in terms of MSE and the MAE, the LR and the SVM models perform much worse than other single models. As a result, the nonlinear relationship between the climate factors and maize production makes it difficult to predict using either the linear regression model or the support vector machine. The performance of various machine learning models in predicting maize yield across different regions demonstrates notable varia- tions. By evaluating the MSE and the MAE, it is evident that the linear regression model dem- onstrates excellent performance in the Ashanti region. Additionally, the decision tree

outperforms other models in the Eastern region. In the Central region, the random forest model emerges as the most effective. Meanwhile, the support vector machine model shows effectiveness in the Northern and Brong Ahafo regions. The gradient boosting model also stands out as the optimal choice in the Volta and Central regions. The Stacking ensemble model also emerged as a top performer in regions many regions like Western, Upper East, Upper West, and Greater Accra, achieving the lowest MSE and MAE values. It is essential to highlight that the regions, particularly Greater Accra, Upper East, and Upper West, exhibit relatively high coefficients of variation, as shown in Table 3. This Table 5. The results of the optimal parameters for each base model after performing Grid Search. Regions Optimal parameters values of each base model after Grid Search with 5-fold CV DT RF SVM GB max depth min sample_split min_sample_leaf n_estimators max depth min sample_split min_sample_leaf C Kernel Degree n_estimators Learning rate Ashanti 5 10 4 100 15 5 4 1 poly 3 100 0.01 Eastern 5 10 4 200 5 10 1 1 rbf 2 100 0.01 Volta 5 2 4 100 15 10 2 1 sigmoid 2 100 0.01 Central 5 10 4 500 5 10 1 1 linear 2 100 0.01 Brong Ahafo 15 2 1 100 5 2 1 0.1 poly 3 100 0.01 Greater Accra 5 5 2 100 10 10 1 0.1 poly 2 100 0.01 Western 5 10 4 200 5 10 4 1 rbf 2 100 0.01 Northern 5 2 4 100 10 10 1 10 poly 2 100 0.01 Upper east 5 2 1 200 10 10 4 1 sigmoid 2 100 0.01 Upper west

15 2 2 100 15 10 4 10 poly 4 100 0.01 https://doi.org/10.1371/journal.pone.0305762.t005 PLOS ONE Impact of climate variability on maize yields in Ghana PLOS ONE | https://doi.org/10.1371/journal.pone.0305762 June 25, 2024 16 / 24