Near-infrared spectroscopy-based parallel assessment of soybean fungal contamination and protein content using an adaptive-threshold Monte Carlo UVE strategy.

Soybean protein content and fungal contamination level are critical determinants of economic value and consumption safety. However, conventional analytical methods rely on destructive chemical analyses or single-parameter measurements, severely limiting assessment efficiency and multi-attribute evaluation. Although near-infrared spectroscopy (NIRS) is widely used, it frequently suffers from spectral redundancy, model instability, and limited generalization in complex agricultural matrices. To address these challenges, this study proposes a parallel NIR analytical framework driven by Ensemble Monte-Carlo Stability-based Uninformative Variable Elimination (EMS-UVE). EMS-UVE integrates repeated stability estimation with a noise-referenced adaptive threshold for high-efficiency informative wavelength selection. For fungal contamination classification, EMS-UVE-XGBoost reduced the feature subset to 12 wavelengths while improving the validation accuracy and macro-F1 score to 0.9529 and 0.9530, respectively, outperforming the baseline ensemble Monte Carlo uninformative variable elimination-based extreme gradient boosting (EMCUVE-XGBoost) (21 wavelengths; accuracy = 0.9412, macro-F1 = 0.9418). For protein content regression, EMCUVE-SVR retained 18 wavelengths, with Rcv=0.8834, RMSECV = 1.6027%, Rp=0.8875, and RMSEP = 1.3338%, whereas EMS-UVE-SVR retained 13 wavelengths, achieving superior performance (Rcv=0.9012, RMSECV = 1.4720%, Rp=0.9070, and RMSEP = 1.2042%). Across 50 independent runs, EMS-UVE-XGBoost yielded a mean validation accuracy of 0.9518 ± 0.0062 and a macro-F1 score of 0.9516 ± 0.0068, while EMS-UVE-SVR achieved af (Rp) of 0.9055 ± 0.0108 with an RMSEP of 1.219 ± 0.058%, demonstrating exceptionally low run-to-run variability. Furthermore, the finalized dual-task models were successfully deployed on a Raspberry Pi 5-based embedded platform. External validation confirmed that the system completed shared spectral acquisition and simultaneous multi-attribute inference within an end-to-end measurement cycle of approximately 5.6 s, achieving a classification accuracy of 0.9200 and a protein prediction correlation coefficient of 0.9067. These findings demonstrate an effective, robust, and edge-deployable strategy for rapid, non-destructive, and multi-attribute evaluation of complex agricultural products.