Abstract
This study presents a low-cost, integrated microwave cavity perturbation sensor operating in TM010 mode at 2.45 GHz for the non-destructive estimation of moisture content (MC) in wheat and chickpea grains. Unlike conventional methods that rely on expensive vector network analyzers (VNAs), this system uses a custom-designed circuit to derive a density-independent moisture content function, M(Ψ). Various machine learning models (ML) were trained to predict MC and classify grain types based on these dialectical metrics. The results demonstrate that bagging (BAG), k-nearest neighbors (k-NN), and reduced-error pruning trees (REPTree) models significantly outperform deep learning models. The BAG model achieved the highest predictive performance, yielding correlation coefficients (R) of 0.995 for wheat and 0.989 for chickpea, with root mean square error (RMSE) of 0.207% and 0.302%, respectively. Furthermore, k-NN variants achieved 100% accuracy in classifying the grain types. The proposed system offers a precise, rapid, and cost-effective alternative for real-time grain quality monitoring.