The geographical origin of plants exerts a significant influence on their chemical compositions and quality. However, the spectral features of samples grown under similar habitats tend to converge remarkably, which compromises the accuracy of geographical origin tracing. Laser-induced breakdown spectroscopy (LIBS) offers distinct advantages of rapid and non-destructive elemental analysis, yet it is constrained by plasma fluctuations and matrix effects, making it challenging to accurately discriminate samples with analogous compositions. Deep learning, by contrast, provides an effective tool for extracting nonlinear features from high-dimensional spectral data. In this study, we propose a LIBS analysis system based on a Feedforward Residual Fusion Network (FRN), termed FRN-LIBS, for the geographical origin tracing of yellowhorn. Yellowhorn samples were collected from 5 regions around Tongliao, and 500 sets of LIBS spectra within the wavelength range of 220–880 nm were acquired. Principal Component Analysis (PCA) was performed to extract 353 principal components with a cumulative contribution rate of 90%, which were then used as the input of the models. The performance of the FRN model integrated with residual modules was compared with that of the conventional Feedforward Neural Network (FNN). The results demonstrate that the accuracy, precision, recall, and F1-score of the FRN model all reached approximately 94%, representing an improvement of around 6% compared with the FNN model. This study confirms that the residual fusion mechanism can enhance the characterization of nonlinear features in LIBS spectra and effectively suppress plasma fluctuations and matrix effects. The proposed FRN-LIBS system provides a versatile and accurate modeling framework for high-dimensional spectral recognition, as well as an efficient and reliable technical solution for the geographical origin tracing of plants
Bai et al. (Fri,) studied this question.