Robust prediction of mandarin fruit quality from repeated portable visible-near infrared measurements using error covariance penalized regression

Abstract With the development of spectrometer miniaturization, portable visible-near infrared (Vis-NIR) devices provide new opportunities for on-site and real-time fruit quality detection. However, under practical complex measurement scenarios, the measurement stability and predictive robustness of portable devices remain limited. Soluble solids content (SSC) is widely used as a representative indicator in fruit quality evaluation and Vis-NIR quantitative analysis. To improve the predictive performance of portable Vis-NIR models, error covariance penalized regression (ECPR) was employed to explicitly incorporate the error covariance matrix estimated from repeated Vis-NIR measurements into SSC prediction modeling. Two mandarin cultivars, Panzhihua Orah mandarin and Australian mandarin, were investigated using repeated spectra acquired by a portable Vis-NIR spectrometer. Experimental results demonstrated that ECPR consistently outperformed conventional regression models. With optimal spectral pretreatment, ECPR achieved coefficients of determination of 0.949 and 0.841, with corresponding root mean square errors of prediction of 0.192 and 0.441, respectively. These results indicate that ECPR modeling based on the covariance structure of repeated measurement errors provides an effective strategy for improving the predictive accuracy and robustness of portable Vis-NIR models.