Application of meta-heuristic feature selection method in low-cost portable device for watermelon classification using signal processing techniques

•To identify the optimal ripening time the watermelon audio features were analyzed through signal process techniques.•A low-cost portable device for watermelon classification was developed.•Sample categorization was accomplished using four classical techniques, namely Discriminant Analysis, K-Nearest neighbor classifiers, Support Vector Machine and Decision Tree.•The watermelon samples were classified into three categories: unripe, ripe and overripe. Recognizing the stage of fruit maturity and thus determining the optimum harvesting time is critical since the competitive market requires high-quality products at a competitive price. Furthermore, in the context of the global water crisis, harvesting watermelons at an inappropriate ripening stage causes water wasting of around 280 kg per kilo of watermelon as a virtual water footprint, highlighting the necessity of criteria for harvesting ripe watermelons by the farmers. This research thus aims to classify the Charleston Gray watermelon type into three categories i.e. unripe, ripe, and overripe using portable acoustic signal processing, data mining methods, and artificial intelligence approaches. Signal processing in the time and frequency domains and Wavelet Transformation were used to extract essential features from acoustic signals of the watermelons. This was followed by the selection of the significant features in classification using a t-test mean comparison. Sample categorization was accomplished using Support Vector Machines and the K-Nearest Neighbor techniques. Considering the effect of three impact positions (the stem side, middle, and flower side), two materials (Polyoxymethylene and metal balls) and two diameters (15 and 10 mm diameter) for striker ball, using Principal Component Analysis and application of various signal processing methods, 3360 models were established. Based on the frequency equipped with the Mahalanobis distance, the K-Nearest Neighbor classifier presented the best result with a 77.3 % classification rate for signals obtained from the small metal ball and stem side impact position. In comparison, experts accurately had categorized 52 % of the samples in total.