Quantitative analysis of cesium in kelp using laser-induced breakdown spectroscopy combined with multivariate models

Kelp is a marine product with a strong bioaccumulation capacity, and its Cs concentration can effectively reflect both marine environmental pollution and seafood safety. In this study, LIBS spectra of kelp samples with varying Cs concentrations were acquired using a nanosecond laser under ambient air conditions. A total of five quantitative models were developed, including a univariate calibration curve (CC) model and four multivariate PLSR models based on different spectral preprocessing methods (Raw, SNV, Total, and Max), to achieve accurate prediction of Cs concentrations in kelp. For raw spectra, the CC model yielded a calibration curve with an <i>R</i>² of only 0.7866, and predictions for 400 and 800 ppm samples exhibited RE of 18.88% and 22.42%, respectively. By contrast, the Raw-PLSR model substantially improved predictive performance, reducing RE to 2.59% and 7.40% for the test-set samples. Further enhancement was achieved with spectral preprocessing, which improved both accuracy and precision. Among the four PLSR models, SNV-PLSR demonstrated the best overall predictive performance. Although its RMSEC was slightly higher than that of Total-PLSR, it produced the lowest RMSEP-reduced by 13.57% relative to the other three models-and the highest <i>R</i>² (0.9802). Consistently, SNV-PLSR yielded the most accurate predictions at both low and high concentrations, with RE values of 0.02% and 2.34%, respectively. In terms of stability, all four PLSR models showed relatively high RSDs for the 400 ppm sample, reflecting the challenges of low-concentration prediction; nevertheless, SNV-PLSR achieved the lowest RSD (18.58%). For the 800 ppm sample, all models performed well, with RSDs below 13.25%. Additionally, three other multivariate methods (RF, PCR, and SVR) were also employed to construct new models by combining different spectral preprocessing methods. The model combinations with the best performance in terms of calibration and prediction indicators were selected and compared with the SNV-PLSR model. The results of the RMSE, <i>R</i>², and test-set MRE further demonstrated the superiority of the PLSR model in the field of multivariate analysis methods. These results demonstrate that LIBS combined with PLSR and appropriate spectral preprocessing can significantly enhance the accuracy and stability of Cs quantification in kelp, providing a promising approach for rapid detection of heavy metals in marine products and monitoring of marine environmental pollution.