Journal of Food Science Food Science

From Memorization to Generalization: L1‐BO‐SVR for Robust Vis‐NIR Prediction of Mango Dry Matter Across Diverse Cultivars, Origins, and Harvest Years

Qing‐xiao Ma, Hao‐xuan Kan, Rui Chen, Wen‐fei Zhang, Guo‐ao Xie, Hao Liu
Wiley · 2026-11-01 · DOI 10.1111/1750-3841.71616

Abstract

ABSTRACT Mango ( Mangifera indica L.) dry matter (DM) content is a critical quality indicator determining consumer acceptance and commercial value, yet conventional oven‐drying methods are destructive and time‐consuming. Visible and near‐infrared (Vis‐NIR) spectroscopy offers a rapid and non‐consumptive alternative, but existing models suffer from poor generalizability due to spectrally homogeneous training sets. To address this limitation, we designed a diverse sampling strategy encompassing three cultivars, two production regions, and three harvest years, yielding 2,787 mango samples with DM content ranging from 9.46% to 22.95%. We propose a hybrid modeling framework named L1‐BO‐SVR integrating L1‐regularized feature selection with Bayesian optimization (BO) enhanced support vector regression (SVR). The L1‐BO‐SVR model selected 94 informative wavelengths from 234 variables and achieved excellent predictive performance, with R 2 P of 0.947 and RMSE P of 0.443%, substantially outperforming full‐spectrum models with R 2 P of 0.895 and conventional preprocessing‐model combinations. Critically, the small performance gaps among training, cross‐validation, and test sets confirm robust generalization with minimal overfitting. These results demonstrate that diverse sampling combined with L1‐BO‐SVR effectively overcomes the “machine memorization” pitfall, enabling reliable and generalizable Vis‐NIR prediction of mango DM content across diverse production conditions, a critical advantage for practical deployment in real‐world supply chains.

Read at publisher →