Storage is crucial because paddy is a seasonal crop where it is only harvested twice a year. Moisture content is a factor that significantly affects the storage of paddy. Numerous methods have been developed to measure the moisture content in paddy, however, most of the existing techniques are unable to measure the moisture content (MC) of the paddy grain in real-time during storage, destroy the samples during the measurements, unable to localize the grain wetspot in bulk density storage, etc. Recently, radio frequency (RF) based methodshave been introduced as a potential method to determine the moisture content. However, the existing work requires complex and expensive equipment such as Vector Signal Generator (VSG) and Vector Network Analyser (VNA) that utilised only one frequency band without localization of grain spoilage wetspots in the storage. Hence, this study aims to characterize the RF frequency band for moisture content determination in real-time in bulk density rice grain storage with a machine learning-based classification algorithm. Multiple radio frequency bands were investigated and identified in order to increase the accuracy of the determination of moisture content and the localization of grain spoilage wetspots. From the characterization test in terms of RSSI measurement, two wireless devices operating based on the established protocol IEEE802.15.4/WSN and Radio Frequency Identification (RFID) was found to have a decreasing RSSI value when the moisture content of the rice was increased. Hence, the two wireless technologies operating in 868 MHz and 2.4 GHz bands, respectively were selected as the dual-frequency band in the developed measurement system.There are six test conditions investigated in this research which is 0% (empty container), 12% (normal moisture condition), and 5 kg of sample conditioned to 14%, 20%, 25%, and 30% moisture content. The samples are set at different locations in the storage container (a large container that contains 20 kg of grain) for wetspot localization purposes. Four machine learning models which are support vector machine, random forest, gradient descent tree, and multilayer perceptron are developed for the classification of moisture content. Among the algorithms, the gradient boosting tree provides the highest accuracy of 94.8% for one output feature. Meanwhile, with 95.4% accuracy, random forest is the best classification algorithm for two output features. As the result, the gradient boosting tree is only suitable for predicting the moisture content of rice. On the other hand, random forest is a suitable algorithm for both predicting the moisture content as well as the localization of the location of the rice spoilage wetspot in the storage. An ensemble method was introduced in this research to further improved the classification accuracy of the weak learners. The ensemble of random forest and gradient boosting trees has provided the highest accuracy for the determination of rice moisture content and localisation of the wetspot in the storage with an accuracy of 99.5%.
