Agricultural scientists investigating geographical indication authenticity have demonstrated that tracking fruit sugar and organic acid dynamics through machine learning models provides a highly accurate method for identifying both the geographic origin and the developmental stage of the Korla fragrant pear. The Korla fragrant pear, scientifically classified as Pyrus sinkiangensis Yu., is a prized fruit variety indigenous to the Xinjiang region of northwestern China that is celebrated for its tender skin, delicate flesh, high juice content, pronounced sweetness, and aromatic profile. In recent years, expanding consumer demand has driven extensive orchard cultivation beyond the historical Korla territory into newly developed agricultural regions such as Alar. Because local variations in regional climate, soil composition, and orchard management influence how internal fruit quality develops, establishing an analytical system that can trace geographical provenance and verify harvest timing has become critical for protecting regional brand authenticity and optimizing commercial picking schedules.
To establish baseline quality metrics, the research team focused on two key production regions in the Xinjiang Uygur Autonomous Region: Korla City and Alar City. The Korla production district occupies the northern margin of the Tarim Basin between 85 degrees 14 minutes and 86 degrees 34 minutes east longitude and 41 degrees 10 minutes and 42 degrees 21 minutes north latitude. This territory experiences a warm temperate arid climate characterized by an average annual temperature of 11.4 degrees Celsius, mean annual precipitation of 58.6 millimeters, and an annual sunshine duration of 2,890 hours. By contrast, the Alar growing district is located on the northern perimeter of the Taklimakan Desert between 80 degrees 30 minutes and 81 degrees 58 minutes east longitude and 40 degrees 20 minutes and 40 degrees 57 minutes north latitude. Alar exhibits a typical continental arid desert climate with an average annual temperature of 10.7 degrees Celsius, annual precipitation of 48.5 millimeters, and an annual sunshine duration of 2,956 hours, providing an environmental contrast in thermal accumulation and light exposure.
Sampling was carried out during the 2023 growing season across orchards containing mature Korla fragrant pear trees aged ten to twelve years that were managed under uniform cultivation protocols. The investigators harvested fruit at three distinct physiological periods: the young fruit stage at approximately 60 days following flowering in late May, the fruit expansion stage at approximately 110 days after flowering in mid-July, and the final fruit ripening stage at approximately 150 days after flowering in late August. To account for natural field growth conditions, the researchers collected 20 fruits per region during the young stage yielding 40 samples, 120 fruits per region during the expansion stage yielding 240 samples, and 40 fruits per region at maturity yielding 80 samples, generating a comprehensive dataset of 360 individual fruits. Following collection, intact disease-free specimens of uniform physical size were immediately transported to the laboratory, peeled, cored, cut into uniform tissue sections, flash-frozen in liquid nitrogen, and stored in an ultra-low-temperature freezer at minus 80 degrees Celsius until chemical extraction.
Chemical profiling of the pear tissues relied on standardized high-performance liquid chromatography procedures to measure three soluble sugars and four organic acids across all specimens. Frozen two-gram tissue aliquots were homogenized in an ice bath with 80 percent ethanol, extracted in an 80-degree water bath for 30 minutes, and centrifuged at 10,000 times gravity at four degrees Celsius, repeating the extraction on the remaining residue to collect 25 milliliters of combined supernatant. Soluble sugars—specifically fructose, glucose, and sorbitol—were separated on a carbohydrate analysis column using an acetonitrile-water mobile phase at a flow rate of 1.0 milliliter per minute and quantified with an evaporative light-scattering detector. Organic acids including malic acid, tartaric acid, citric acid, and fumaric acid were resolved on a reversed-phase C18 column using a potassium dihydrogen phosphate buffer at pH 2.5 and methanol at 0.8 milliliters per minute, detected with a diode array detector at 210 nanometers. Total sugar content was calculated as the sum of fructose, glucose, and sucrose; total acid content was calculated as the sum of malic, citric, tartaric, and fumaric acids; and the sugar-to-acid ratio was computed by dividing total sugar by total acid.
The quantitative chromatographic analyses revealed distinct developmental trajectories for individual sugar constituents between the two geographic origins. Throughout fruit maturation, the concentrations of fructose, glucose, and total sugars exhibited continuous upward trajectories in both regional populations, whereas sorbitol showed a sharp and continuous downward decline. During the young fruit phase, sorbitol represented the dominant carbohydrate fraction, accounting for 64.50 percent of total soluble sugars in Korla pears and 75.10 percent in Alar pears. As development advanced into the expansion and ripening stages, sorbitol was steadily metabolized and converted into hexoses, dropping to 16.12 percent of total sugars in mature Korla fruit and 25.18 percent in mature Alar fruit. Concurrently, fructose proportions expanded from 21.91 percent to 45.85 percent in Korla pears, while glucose rose from 13.59 percent to 38.03 percent. At full commercial maturity, Alar pears contained a significantly higher fructose concentration of 4.12 grams per 100 grams compared to 3.64 grams per 100 grams in Korla pears, while total sugar values converged to 9.23 grams per 100 grams in Alar and 9.01 grams per 100 grams in Korla.
The organic acid profiles underwent equally pronounced transformations as fruit advanced toward harvest maturity, driving major shifts in overall sensory balance. Concentrations of malic acid, tartaric acid, citric acid, and cumulative total acids decreased steadily across development, whereas fumaric acid maintained an extremely low baseline concentration that showed a slight upward trend. Malic acid persisted as the predominant organic acid throughout all developmental phases, comprising 56 to 77 percent of total acid content across both regions, while citric acid remained minimal at under one percent and fumaric acid contributed less than four percent of the total acid pool. While young Korla fruits exhibited significantly higher tartaric and citric acid concentrations than young Alar fruits, organic acid differences diminished by harvest. However, because mature Alar pears accumulated higher fructose alongside reduced acid levels, their final sugar-to-acid ratio reached 101.2, which was significantly higher than the ratio of 78.6 recorded for mature Korla pears. This divergence highlighted a distinctly elevated perceived sweetness in fruit from Alar, contrasted with a more balanced sweet-tart flavor equilibrium in Korla fruit.
Multivariate statistical evaluations and Pearson correlation matrices confirmed strong biochemical linkages among the measured metabolites during specific developmental windows. During the young fruit period, fructose and glucose showed no statistical correlation with one another or with sorbitol, whereas sorbitol correlated almost perfectly with total sugar content at a correlation coefficient of 0.97. As pears expanded, the correlation between fructose and glucose rose to 0.71, malic acid correlated with total acid at 0.99, and the sugar-to-acid ratio exhibited an inverse correlation of minus 0.86 with malic acid and minus 0.84 with total acid. Principal component analysis demonstrated clear separation between Korla and Alar samples along the first principal component during the young fruit stage, explaining 48.9 percent of total variance. However, score plots showed substantial spatial overlap between the two growing regions during the expansion and mature stages, indicating that unsupervised linear dimensionality reduction could not fully resolve regional origin as metabolic profiles converged later in the growing season, highlighting the necessity for advanced non-linear supervised classifiers.
To establish predictive origin classification across all developmental periods, the researchers constructed three distinct supervised machine learning architectures: a backpropagation neural network, a random forest classifier, and a support vector machine utilizing a radial basis function kernel. Modeling protocols partitioned the 360-sample dataset into a 70 percent training subset and a 30 percent independent validation subset using stratified sampling across growth periods and regions, supported by five-fold cross-validation. In full-stage provenance discrimination, the support vector machine attained 100 percent accuracy on training data but declined to 92.9 percent accuracy on the validation set, reflecting overfitting caused by complex kernel boundaries on sample subsets. The backpropagation neural network achieved 98.4 percent training accuracy but dropped to 86.0 percent on validation samples. In contrast, the random forest ensemble demonstrated superior generalization and structural stability, securing 94.8 percent accuracy on training data alongside a validation accuracy of 92.6 percent, with macro-precision, macro-recall, and macro-F1 scores all reaching 0.926. These metrics proved that the ensemble of decision trees handled nonlinear feature interactions with high resilience against regional noise.
To overcome the inherent interpretability challenges of machine learning, the researchers applied game-theoretic SHapley Additive exPlanations, or SHAP, to determine the exact marginal contribution of each chemical feature to model predictions. Across the global random forest origin model, glucose emerged as the single most discriminative feature with a mean absolute SHAP value of 0.30, followed by tartaric acid and fumaric acid. SHAP dependence analysis revealed a distinct critical threshold at 2.0 grams of glucose per 100 grams of fruit tissue: samples falling below this concentration produced negative SHAP values directing classifications toward Alar, while concentrations exceeding 2.0 grams per 100 grams shifted SHAP values positive toward Korla. For sorbitol, values under 2.0 grams per 100 grams favored Alar, whereas levels between 2.0 and 3.0 grams per 100 grams rose sharply and stabilized above 3.0 grams per 100 grams toward Korla. When evaluating stage-specific models, key biomarkers shifted dynamically: sorbitol was the primary origin indicator during the young fruit phase with an importance score of 0.35 in neural network models, citric acid dominated origin discrimination during the fruit expansion stage with an importance of 0.29 in random forest models, and the sugar-to-acid ratio served as the primary diagnostic metric at commercial harvest with an importance score of 0.077.
Beyond geographical origin, the team trained three-class models using full-stage chemical profiles to automatically classify pears into young fruit, expansion, or maturity stages, finding that the backpropagation neural network delivered the highest validation accuracy at 0.972, outperforming random forest at 0.940 and support vector machines at 0.750. Feature attribution identified fumaric acid as the decisive stage biomarker across fruit development, exhibiting a monotonic relationship where concentrations below 0.03 milligrams per gram corresponded to the young fruit phase, concentrations from 0.03 to 0.05 milligrams per gram signaled the expansion phase, and levels exceeding 0.05 milligrams per gram denoted complete harvest maturity. The authors noted several operational limitations, emphasizing that their sample collection was conducted across a single harvest year and restricted to representative orchards within two geographic regions, meaning multi-year seasonal climate fluctuations and broader geographical variations require further validation. Furthermore, non-destructive optical tools such as near-infrared spectroscopy and broader metabolomic profiles should eventually be coupled with these chemical algorithms to build portable digital devices capable of real-time field grading and origin verification.
Subject of Research: Machine learning origin traceability and maturity discrimination in Korla fragrant pears
Article Title: Pear origin traceability and maturity discrimination based on sugar-acid dynamics and machine learning: multi-model comparison and SHAP interpretability analysis
Article References: Yu, M., Wang, H., Zeng, J., Fan, W., Wang, L., & Bao, J. (2026). Pear origin traceability and maturity discrimination based on sugar-acid dynamics and machine learning: multi-model comparison and SHAP interpretability analysis. Food Chemistry: X, 40, Article 104772. https://doi.org/10.1016/j.fochx.2026.104772
Image Credits: AI Generated
DOI: 10.1016/j.fochx.2026.104772
Keywords: Korla fragrant pear, Pyrus sinkiangensis, geographical indication, origin traceability, maturity discrimination, sugar-acid dynamics, random forest, backpropagation neural network, support vector machine, SHapley Additive exPlanations, fumaric acid, sorbitol accumulation, high-performance liquid chromatography, fruit quality evaluation, Alar production region, Korla production region, food authenticity
News Source: Blake Davidson. (November 18, 2026). Machine Learning and Sugar Acid Dynamics Track Pear Origin and Ripening. Scienmag.



