Rapid and explainable classification of ZEN contamination in stored multi-grain based on microwave frequency sweeping coupled with PKO feature selection, PKO-optimized RF, and SHAP analysis

Zearalenone (ZEN) contamination is widespread in stored wheat and maize, posing a serious threat to food safety. This study proposed a novel approach for identifying ZEN contamination levels in wheat and maize. Microwave transmission frequency-sweeping technology was employed to collect response signals from cereal samples with different ZEN concentrations, and the contamination levels were categorized into three classes: low, medium, and high. To address the redundancy of high-dimensional microwave features and the insufficient stability of conventional models, the Pied Kingfisher Optimizer (PKO) was introduced to simultaneously perform key frequency feature selection and optimize the hyperparameters of the Random Forest (RF) model. Moreover, the Shapley Additive exPlanations (SHAP) interpretation framework was integrated to reveal the contribution of key microwave frequencies to classification decisions. The experimental results demonstrated that the optimized PKO-RF model exhibited excellent classification performance on independent test sets, achieving a classification accuracy of 97.44% for both wheat and maize samples. SHAP analysis further revealed the characteristic frequency-response patterns associated with ZEN contamination levels in different cereal matrices and identified important microwave frequencies related to ZEN contamination-level discrimination, thereby improving the transparency and reliability of model predictions. Therefore, the PKO-RF model combined with the SHAP interpretation framework provides a novel strategy for rapid, non-destructive, and interpretable detection of ZEN contamination levels in cereals, demonstrating the potential of microwave transmission frequency-sweeping technology integrated with intelligent algorithms for applications in stored-product research.