Determining Philippine coconut maturity level using machine learning algorithms based on acoustic signal

Advanced intelligent systems are becoming significant to many sectors, including farming. In agriculture, the intelligent classification of post-harvested fruits seems to have a direct impact on farmers, mainly for export products. Unlike other popular fruits, coconuts tend to have limited studies due to its tropical nature grown in developing countries as well as its unique physical structure. In this study, a classification of real coconut datasets is performed based on acoustic signals acquired through a developed tapping system and learned by three widely-used machine learning techniques - artificial neural network (ANN), random forest (RF) and support vector machine (SVM). There are 129 coconuts samples, each classified into three maturity levels - pre-mature, mature, and over-mature. A three-second tapping system gathered from each sample a total of 132,300 data points, which underwent noise reduction and signal processing. Each machine learning model predicts the class of the fruit by learning the patterns of the transformed frequency spectrums of each sample signal. Based on ten times cross-validated results, the three machine learning algorithms satisfactorily predicted the maturity level of coconuts with at least 80% classification accuracy. All models correctly predicted over-mature coconuts but confused in classifying pre-mature with mature and mature with over-mature coconuts. RF model outperformed the other models with efficiencies of 90.98% and 83.48% accuracies for training and testing, respectively. The imbalance data for each coconut class can be addressed to give better results. Additionally, the prepared coconut dataset may use more advanced deep learning techniques.