Optimizing gas sensor array to classify three edible ripening stages of date fruit using machine learning and feature selection (case study: cv. Shahani.)

Aim of study: This study aimed to analyze aroma changes in date fruits (Phoenix dactylifera L.) across three ripening stages (Khalal, Rutab, and Tamr) using an e-nose as a non-destructive monitoring method for fruit quality management. The study also evaluated feature selection methods to optimize the sensor array for a repeatable model based on external data. Area of study: This study was carried out at a date palm garden cultivated by cv. Shahani in Jahrom, west-south of Iran. Material and methods: An e-nose profiled the aroma of date fruits from a date palm garden across three ripening stages over two years. Classification models, developed using one year's data as internal data, were externally validated using data from the other year as external data. Feature selection optimized the sensor array, improving the prediction accuracy on the external dataset. Main results: The classification models had similar test accuracy, but their predictive performance varied on external data. Using F-test and principal component analysis (PCA) for feature selection to optimize the sensor array, with support vector machine (SVM) as the classification algorithm, resulted in a highly repeatable model. This study suggests that e-noses are a promising tool for monitoring aroma changes during date fruit ripening in artificial ripening or storage processes. Research highlights: Aroma profiles from edible ripening stages in data fruits were classified with high accuracy using e-nose; Feature selection methods can affect linear discriminant analysis (LDA) and SVM algorithms differently on external data; In SVM, some feature selection methods decreased prediction accuracy on external data compared to using the full sensor array; PCA effectively determined the optimal number of features for optimizing the sensor array.