Pre-harvest and post-harvest plant-based food quality evaluation based on near-infrared spectroscopy coupled with multiscale deep ensemble learning.

Accurate assessment of internal quality attributes in pre- and post-harvest plant-based foods is essential for maturity evaluation and quality grading. Although near-infrared (NIR) spectroscopy offers rapid, non-destructive compositional analysis, real products exhibit marked chemical heterogeneity from variable growth and storage environments, while their spectra remain highly collinear due to overlapping overtone bands. This combination masks subtle compositional differences and limits the reliability of conventional chemometric models. This study develop a multiscale deep ensemble regression framework (MDER) that disentangles chemically informative variation from redundant spectral correlations through multi-receptive-field 1D-CNN branches and an attention-based meta-learner. Using a multi-season Vis-NIR mango dataset (dry matter) and a multi-origin NIR corn kernel dataset (protein), MDER achieved high predictive accuracy (mango: R2 = 0.9352; corn: R2 = 0.9046) on the test set. These results demonstrate that coupling MDER with NIR provides an effective and chemically meaningful approach for quality evaluation under heterogeneous agricultural conditions.