The assessment of soil variables relative importance for cereal yield prediction under rainfed cropping system in Morocco

In this study, Explainable Artificial Intelligence (XAI) techniques were applied to identify the most important factors influencing crop yield prediction, with a focus on strategies for sustainable agriculture. Using permutation importance and residual plot analysis, the results showed that nitrogen (N) content, Bandera variety and potassium oxide ( K 2 O ) are the most important traits influencing yield prediction. Extreme Gradient Boosting (XGB) was used to predict yield using a large Moroccan national cereal dataset spanning 3 seasons. Residual Plots Analysis, Partial Dependent Plots (PDP), Permutation Importance (PI) and SHapley Additive ExPlanations (SHap) were used to select the features that influence yield prediction. The results indicate that optimizing soil nitrogen and potassium oxide levels together with strategic selection of crop varieties can significantly increase productivity. Residual analysis of the eXtreme Gradient Boosting (XGB) model confirmed its high predictive accuracy. This study underlines the value of XAI methods in improving the interpretability of predictive models. The insights gained can contribute to better soil management and informed crop selection, ultimately reducing yield losses under environmental stress. By increasing the resilience of agricultural systems, we aim to contribute to sustainable and data-driven farming practices and, in particular, address some of Morocco's unique agricultural challenges. • Evaluation of key soil variables affecting cereal yield predictions in Morocco's rainfed cropping systems. • Application of SHAP-based Explainable AI to interpret soil variable impacts on cereal yield predictions. • Providing insights for targeted soil management to boost cereal production and food security in arid regions.