Development of Multifrequency-Swept Microwave Sensing System for Moisture Measurement of Sweet Corn With Deep Neural Network

Moisture measurement has long been a challenge for agricultural products with high moisture content (MC). In this article, a novel microwave sensing system embedded with multifrequency-swept technique was built with off-the-shelf components and applied to moisture measurement of sweet corn [MC is approximately 80% wet basis (w.b.)]. In order to collect sufficient moisture information, a frequency-swept signal (contains 41 frequencies from 2.60 to 3.00 GHz) was taken as the original measurement signal. A total of 20 redundant frequencies were removed from the original measurement signal according to the frequency selection for further measurements. Four different algorithms, including deep neural network (DNN), random forest (RF), adaptive boosting (AdaBoost), and extreme gradient boosting (XGBoost), were employed to establish moisture prediction models. The proposed six-layer DNN showed the best performance (R
²
= 0.980, RMSE =2.023%, and MAE =1.556%) in predicting the MC of sweet corn (ranging from 15.45% to 81.19% w.b.). The results showed that the developed microwave sensing system was capable of measuring the MC of sweet corn and could potentially be applied to moisture determination of other agricultural products with high MC in food processing industry.