Classification for Orange Varieties Using Near Infrared Spectroscopy

A reduced classification model is developed in order to discriminate the three orange varieties (i.e., Kaew Wan, Number One, and Sai Nam Pung). A diversity of classification methods, including kNN, Linear Discriminant Analysis (LDA), Logistic Regression (LGR), Multi-Criteria Quadratic Programming (MCQP), and Support Vector Machine (SVM), were first evaluated on the complete data. The 10-fold cross-validation results demonstrate that the best performing model - LGR reaches 100% classification accuracy with all the 255 Near Infrared NIR spectrum features. From these 255 features, a subsequent feature selection identified four spectra of good discriminative ability. Based on these four NIR spectrum features, a reduced LGR is developed and its classification accuracy is as high as 95%. This finding suggests that the oranges can be classified with satisfying accuracy by measuring only four NIR spectra instead of all the 255 ones.