LIBS combined with TrAdaBoost based transfer learning for quantitative analysis of heavy metals in soil particles

Abstract

Laser induced breakdown spectroscopy (LIBS) has been proven to be a feasible technique for rapid on-site analysis of soil heavy metals in recent years. However, despite advantages like no complex sample pretreatment, real-time analysis, and multi-element detection, LIBS still faces challenges in field applications, including instrument accuracy and soil matrix effects, which may cause inaccurate and inconsistent results. This study addresses the challenge of applying LIBS to the real-time on-site monitoring of particle soil heavy metals. Using various soil forms as research objects, a quantitative analysis model based on LIBS combined with the transfer adaBoost (TrAdaBoost) algorithm was developed. By investigating the spectral characteristics of both tablet and particle soil samples, a regression model was established using spectral data from both forms, enabling the transfer of spectral features from tablet to particle samples to improve quantitative accuracy. The model performance was first evaluated by examining the effects of parameters and preprocessing methods. Subsequently, the input variables were optimised by increasing the proportion of particle sample spectra in the training set, further enhancing prediction accuracy. The proposed method was then compared with the random forest (RF). Results show that the TrAdaBoost transfer model outperformed conventional approaches, achieving determination coefficients (R²p) of 0.9885 for Cu, 0.9473 for Cr, 0.8958 for Zn, and 0.9563 for Ni, with corresponding root mean square errors of prediction (RMSEp) of 8.7812, 5.8027, 33.9846, and 13.2258 mg·kg-1, respectively. These findings show that the proposed transfer learning approach greatly improves LIBS-based in-situ quantitative analysis of soil samples, providing a new technical solution and research direction for addressing matrix effect challenges, with strong engineering applicability and practical potential.