A three-wavelength spectral approach for automated detection of chilling damage and juice content in oranges

Early and accurate detection of chilling damage in citrus is essential for preserving fruit quality and reducing postharvest economic losses. This study presents a rapid, non-destructive approach for identifying early-stage chilling injury and predicting juice content in oranges by integrating visible–near infrared (Vis/NIR) spectroscopy with machine learning (ML). Spectral absorbance in the range of 400–2500 nm was collected from 200 field-grown oranges representing both sound (healthy) fruit and those exhibiting early chilling damage. The spectra were pre-processed using Diameter Normalization, Standard Normal Variate, Multiplicative Scatter Correction, and first- and second-order derivatives. Six ML classifiers such as Partial Least Squares Discriminant Analysis (PLS-DA), Support Vector Classifier, Random Forest, XGBoost, Gaussian Process classifier, and Extreme Learning Machine (ELM) were applied to differentiate sound oranges from chilling-damaged ones. Juice content was then estimated using six regression models, including Partial Least Squares Regression (PLS-R), Support Vector Regressor, Random Forest Regressor, XGBoost Regressor, Gaussian Process Regressor, and ELM Regressor. All models were trained and cross-validated on both full-spectra inputs and selected informative wavelengths sets. A multi-method wavelength importance analysis PLS VIP (Variable Importance in Projection) scores and Random Forest Gini/MSE metrics consistently identified three pivotal wavelengths 520, 675, and 1710 nm associated with chlorophyll degradation, pigment changes, and water/sugar dynamics. Remarkably, models using only these three wavelengths achieved performance comparable to full-spectra models: ELM reached 0.980 accuracy for chilling damage classification versus 0.990 for full-spectra PLS-DA and 0.871 R² for juice prediction versus 0.910 R² for the full-spectra Gaussian Process Regressor. Leveraging these wavelengths, we formulated a logistic regression index that provides transparent, computationally efficient chilling-damage probabilities highly aligned with ML predictions. This simplified spectral fingerprint enables real-time, on-site fruit quality screening while minimizing hardware complexity. Overall, the findings highlight that a concise set of informative wavelengths can effectively replace high-dimensional spectral models without compromising accuracy, offering a practical and scalable pathway for automated citrus quality assessment.