Rapid classification of heavy metal soils from different mining areas by using a GSCV quadratic merit seeking network combined with MF-LIBS

Abstract

Focusing on the problems of complicated sample processing and long detection time of traditional soil composition testing in mining areas, this paper proposes a rapid soil classification method of grid search and cross validation (GSCV) quadratic optimization-seeking network combined with magnetic field enhanced laser-induced breakdown spectroscopy (MF-LIBS), which can quickly and accurately classify standard soil samples from 10 different mining areas. First, the effects of different magnetic field strengths on the enhancement factor, signal-to-noise ratio (SNR) and plasma electron temperature of the characteristic spectral lines were investigated. Next, 5400 12 248-dimensional spectral data under the 0.98 T magnetic field constraint were preprocessed using principal component analysis (PCA). Then the back propagation (BP) network was optimized by combining the sparrow search algorithm (SSA) external optimization network with the genetic algorithm (GA) internal optimization network. Finally, the hyperparameters of the optimal network architecture are refined by GSCV for secondary optimization. By comparing with the unoptimized network, the results show that the classification accuracy of GSCV-SSA-GA-BP is 0.99997, the precision is 0.99977, the recall is 0.99981, and the F1-score is 0.99979. The proposed technique of GSCV quadratic optimization network combined with MF-LIBS soil classification can not only play an important role in soil research, but also provide a new scheme for rapid detection and high-precision qualitative analysis of soil components.