High-Accuracy Classification of Peanut Origin via Feature-Level Fusion of Multimodal Spectroscopy

Abstract

The geographical origin of peanuts directly determines their quality and market value. To overcome the destructive and time-consuming nature of conventional detection methods, this study proposes a novel nondestructive traceability strategy by integrating near-infrared (NIR) spectroscopy and low-field nuclear magnetic resonance (LF-NMR). Whole kernels of the peanut cultivar “LuhuaNo.9” were examined, from which NIR spectra reflecting chemical composition and LF-NMR signals characterizing moisture–lipid dynamics were simultaneously acquired. Two fusion strategies combined with several classifiers were evaluated: (1) low-level fusion, in which the original NIR and LF-NMR spectra were concatenated and modeled using support vector machine (SVM), extreme gradient boosting (XGBoost), random forest (RF), and a self-designed squeeze-and-excitation convolutional neural network (SE-CNN); (2) feature-level fusion, where informative spectral variables were selected using competitive adaptive reweighted sampling (CARS) to eliminate redundancy and noise. The results demonstrate that multisource data fusion substantially outperforms single-modality analysis, and the SE-CNN exhibits superior discriminatory ability compared with traditional machine learning models. In particular, the combination of CARS-based feature-level fusion and the SE-CNN achieved 100% classification accuracy on the test set, highlighting its outstanding performance. Overall, this work establishes a high-precision traceability strategy that integrates multidimensional spectral information with deep learning, markedly improving identification efficiency while preserving sample integrity. The proposed approach holds significant promise for the protection of geographical indications and quality monitoring of agricultural products.