Quantitative prediction of active components in turnip using near-infrared spectroscopy (NIRS) and a multimodal...

Traditional feature selection algorithms dominantly rely on isolated, localized data points, affecting the extraction of comprehensive information from near-infrared spectra, thus limiting the predictive performance of the employed models. To address this limitation, we proposed a method that integrates partial least squares regression (PLSR) with a BGCA (a multifeature fusion framework combining Bi-Mamba, GAF-CNN, CARS, and KPCA) method for the rapid quantitative analysis of saponins and flavonoids in turnip. The multimodal model exhibited exceptional prediction accuracy for both saponins and flavonoids, achieving an R²p of 0.908 and an RPD of 3.360 for saponins, as well as an R²p of 0.734 and an RPD of 1.973 for flavonoids. The modeling performance significantly surpassed traditional feature selection algorithms, meeting the required standards for quantitative prediction, suggesting that the BGCA fusion method is effective in extracting relevant information from near-infrared spectral data, offering a novel strategy for the rapid and non-destructive detection and quality control of active pertinent ingredients for the traditional Chinese medicine.