An efficient machine learning framework for optimizing hyperspectral data analysis in detecting adulterated honey

Abstrak

Honey adulteration detection involves employing spectral data, often utilizing machine learning (ML) techniques, to identify the presence of impurities or additives in honey. This study aims to explore ML models through the collection of a hyperspectral honey dataset with limited samples and 128 features. Three distinct feature selection (FS) methods i.e., Boruta, repeated incremental pruning to produce error reduction (RIPPER), and gain ratio attribute evaluator (GRAE) are applied to extract important features for decision-making. Then, the feature-selected dataset is classified through four effective ML algorithms, such as support vector machine (SVM), random forest (RF), logistic regression (LR), and decision tree (DT). Accuracy, F1-score, Kappa Statistics, and Matthews correlation coefficient (MCC) are the performance metrics used to assess the results of ML algorithms. RIPPER FS technique gave the best results by improving its accuracy values from 79.05% (primary data) to 91.89% (augmented data) for the RF classifier model and 74.93% (primary data) to 91.89% (augmented data) for the DT classifier model. These detailed examinations of the experiments demonstrate that proper finetuning of the ML methods can play a vital role in optimizing hyperspectral data analysis for detecting adulteration levels in honey samples.