Robust detection of soluble solids content in winter jujube using contrastive learning and spectroscopy: Mitigating the impact of measurement position

Visible and near infrared spectroscopy (Vis-NIRS) has been widely used for the detection of soluble solids content (SSC) in fruits. However, uncertain measurement positions, which significantly alter the spectral data, limit the effective application of traditional models in practical applications. This study established a global multi-position SSC prediction model for winter jujube across multiple maturity stages using Vis-NIRS. First, maturity probabilities obtained from a classification model were used to guide expert regression. Then, contrastive learning was applied to suppress prediction uncertainty caused by sampling position differences. This led to our proposed Contrastive Learning Expert Regression Convolutional Neural Network (CLERCNN). This study developed prediction models for SSC using single-position spectra, global multi-position spectra, and averaged spectra. Results showed that while the averaged spectra model performed best, it was inferior to the global multi-position detection model when applied to the single-position spectra prediction set. The CLERCNN further enhanced the predictive performance of the global multi-position model, with R p 2 and RMSE p values of 0.9131 and 1.5000. It also outperformed both the averaged spectra model and other comparative models. This research can effectively eliminate the impact of sample measurement position variability, providing a reliable solution for online quality monitoring of winter jujube. • SSC in winter jujube was predicted using a global multi-position spectral model. • A novel Contrastive Learning Expert Regression CNN (CLERCNN) framework is proposed. • The model mitigates prediction uncertainty arising from measurement position. • It seamlessly integrates maturity information to guide expert regression.