Performance of feature extraction method combination in arabica coffee roasting classification

Abstract

Feature extraction is vital in electronic nose technology, particularly for classification tasks. However, challenges like noise, temperature variations, humidity, drift, and unwanted aromas can introduce inconsistencies in feature extraction, diminishing the machine’s classification capabilities. This study aimed to assess the electronic nose’s performance in recognizing aroma patterns of Arabica coffee at four roasting levels. It involved comparing 63 feature extraction method combinations derived from six primary methods (mean, skewness, kurtosis, standard deviation, maximum, minimum). After extracting features, Linear Discriminant Analysis (LDA) was used for dimensionality reduction and analysis. Subsequently, a Support Vector Machine (SVM) model was trained and validated using Stratified K-Fold Cross Validation to comprehend feature patterns and labels and determine hyperplanes for distinct classes. The results revealed that the best feature combination to classify arabica coffee aromas at various roasting levels was achieved by using a combination of mean, kurtosis, and standard deviation, with an accuracy of 86.11%, precision of 86.73%, recall of 86.11%, and MCC of 0.8159 in a training time of 0.0574 seconds. Utilizing LDA improved accuracy by 9.81% and MCC by 15.01%.