Prediction of bayberry SSC by ensemble model with random frog successively selected from the residual vis-NIR spectra

This study presents a robust for non-invasively predicting the soluble solids content (SSC) of Chinese bayberry using an affordable portable visible near-infrared (vis-NIR) analyzer, combined with an adaptive boosting (AdaBoost) ensemble learning framework. Interactance spectra of bayberry were pretreated with z-score normalization, and relevant variables to develop partial least squares (PLS) model were screened using random frog (RF) from the residual spectra in last RF selection round. AdaBoost ensemble models, were respectively constructed with a series of top RF-PLS models and compared against other ensemble strategies. Results revealed that the residual spectral variables retained valuable information after successive RF selection, with the fifth RF-PLS model obtaining a root mean squared error of prediction (RMSEP) of >0.77, only slightly less accurate than the full spectrum-based PLS model. The AdaBoost model, integrating the top three member models, obtained the best predictive capacity with an RMSEP of 0.713, represent a 7.2 % reduction over the full spectrum-based PLS model. This performance surpassed that of deviation, average and least absolute shrinkage and selection operator (Lasso) ensemble strategies. This ensemble framework, leveraging successively constructed RF-PLS member models from residual spectra, maximizes the utilization of spectral information and enhances the prediction accuracy for SSC in bayberry using an affordable NIR analyzer.