Abstract:Rapid identification of the geographical origin of Orah mandarins is important for ensuring product quality and facilitating automated sorting. To develop a rapid, nondestructive method for origin screening, hyperspectral images of Orah mandarins from Guangxi, Yunnan, and Sichuan were acquired, and mean reflectance spectra were extracted. Raw spectra (RAW) were used as the baseline for comparison with two preprocessing approaches, Savitzky–Golay smoothing combined with standard normal variate transformation (SG+SNV) and multiplicative scatter correction (MSC). Three feature sets were evaluated: principal component analysis (PCA) scores, histogram of oriented gradients features extracted from Markov transition field images (MTF-HOG), and a combination of the two. Support vector machine (SVM), random forest (RF), and one-vs-rest logistic regression (OVR-Logistic) classifiers were used to identify fruit origin. Among the 27 model combinations, RAW+PCA-SVM achieved the highest mean accuracy (0.9593) and macro-averaged F1 score (0.9569) in cross-validation, with a macro-averaged area under the receiver operating characteristic curve (AUC) of 0.9816. Its mean accuracy exceeded those of SG+SNV+PCA-SVM (0.9178), MSC+PCA-SVM (0.9222), RAW+MTF-HOG-SVM (0.8993), and RAW+PCA+MTF-HOG-SVM (0.8676).Under the experimental conditions, PCA and SVM applied to raw hyperspectral reflectance spectra distinguished Orah mandarins from the three geographical origins, providing a basis for rapid origin screening and product traceability.