A new approach in generating feature-rich graphical maps using machine learning and its application in peach juice processing

This study presents a novel approach for converting compound concentration profiles derived from GC-IMS analysis into detailed graphical feature maps, designed to more effectively leverage machine learning (ML) techniques. The proposed approach utilizes a mathematical framework based on a row weighted loss function (RWLF) combined with sample reweighting gradient descent (SRGD), which is incorporated into weighted-artificial neural networks (w-ANNs) along with a convolutional neural network (CNN). This methodology enables the extraction of richer features through advanced scalar matrix and vector manipulation. These enhanced features are then transformed into graphical feature maps to streamline ML tasks, particularly in fruit juice classification and processing. The method was tested on classifying two peach varieties from different regions, as well as the same peach type being processed at various juicing stages. The results showed promising accuracy, with loss metrics suggesting that the proposed RWLF-SRGD approach effectively captures key features-not just for peach juices from different locations, but also for the same peach juice across different processing steps. The average accuracy reaches 0.93 in classifying peaches from different locations, while it reaches 0.96 in classifying different processing steps for the same peach juice. This demonstrates the model's reliability, robustness, and potential for broad applications in fruit processing and production. This research establishes a new benchmark for converting GC-IMS compound-concentration profiles into feature-rich graphical maps.