Rapid identification and prediction of the botanical origins and main components of Chinese yam based on a multimodal data fusion of near-infrared spectra and visualization image information.

Yam was recognized as a nutritionally valuable crop, creating demand for rapid, and non-destructive quality control methods. A multimodal fusion strategy was developed by integrating near-infrared spectroscopy (NIRS) with image features to address this need. The machine learning models constructed from the fused data exhibited strong discriminatory performance. Specifically, a linear discriminant analysis model incorporating Savitzky-Golay smoothing and random forest feature selection of NIRS data achieved 94.59% accuracy in botanical origin identification. A support vector machine model utilizing image-texture features attained the identical accuracy. The combination of NIRS and image data significantly enhanced the prediction precision and model stability, yielding 97.30% discrimination accuracy. Additionally, the partial least squares regression models based on NIRS data accurately predicted key nutritional components including starch and polysaccharides. This integrated approach established an efficient, non-destructive framework for the quality evaluation of yam, demonstrating considerable potential for industrial applications in yam processing and quality assurance.