UNSUPERVISED MACHINE LEARNING APPROACH TO QUALITY MONITORING OF STRAWBERRIES DURING DRYING

GAMBOA, JULIANA

Congreso; II Argentine Congress of Bioinformatics and Computational Biology (XII CAB2C); 2022

ABSTRACTBACKGROUNDThis work aims to analyse the potential of unsupervised machine learning (ML) models to be applied during food quality monitoring of drying operations. In this sense, a microwave (MW) assisted drying (1.2 W/g, 100 min) of strawberry was carried out to validate a machine vision system composed of a digital camera and a 6-modules system using image embedding, programmed in Python. The system collected information about several morphological and colour features (area, enclosing rectangular area, height, width, radius, brightness and saturation retentions). Feature extraction tasks were performed from 490 individual images of fresh (FR) strawberry samples during MW drying processing. Final dataset accounted the information of 19 features (9310 samples), among them, drying kinetics parameters (moisture losses at each drying time condition) were also computed. With the aim to identify drying time thresholds of quality changes a k-Means clustering model was performed combining the selected features.RESULTSAfter applying a dimensionality reduction to 3 principal components, FR samples subjected to MW drying were naturally clustered among two (k=2) groups, with high clustering metrics performance. By analysing 16 different clustering configurations (with k values of 2 or 3) the better results were obtained for two drying time categories (k=2) at configuration 4 (C4, ARI: 0.957, AMI: 0.915), suggesting a quality threshold at drying times above 60 min.CONCLUSIONSThe results presented here showed the potential of unsupervised classification methods coupled with image embedding as an attractive and economical alternative system to define quality thresholds in order to in-line monitoring of quality changes during drying.