Quantitative analysis of potentially toxic elements in soil by XRF based on efficient reinforcement learning and sparse partial least squares

Measuring the accurate concentration levels of potentially toxic elements in agricultural soils is necessary to prevent hazards to human health. X-ray fluorescence spectroscopy (XRF) allows rapid, non-destructive detection of potentially toxic elements in agricultural soils. However, the high-dimensional spectral information associated with low concentrations of potentially toxic elements is difficult to be mined effectively, affecting the quantitative performance. Reinforcement learning is a potent paradigm for variable spatial search that allows effective modeling of high-dimensional data. This paper proposes a novel XRF spectral variable selection algorithm based on efficient Q-learning and sparse partial least squares (SPLS) to establish the relationship between XRF spectra and a single potentially toxic element of low concentrations (i.e., QSP). Firstly, an energy-dispersive X-ray fluorescence (EDXRF) spectrometer is used to obtain the XRF spectral information of soil samples. Then, a novel efficient Q-learning XRF spectral variable optimization algorithm is proposed to select spectral variables related to a single potentially toxic element of low concentrations. Finally, the efficient Q-learning algorithm is combined with the proposed SPLS to predict the concentrations of four potentially toxic elements, respectively. Compared with other advanced algorithms, the QSP achieves the lowest root mean squared error of prediction (RMSEP) of 8.540, 2.217, and 40.667 for Cr, Cd, and Ba, respectively, with the highest coefficients of determination (R2) of 0.988, 0.854, and 0.958. The concentration prediction results of QSP for Pb are similarly competitive. These results show that combining the QSP with XRF provides a new solution to improve quantitative performance for potentially toxic elements of low concentrations. The relationship between XRF spectra and potentially toxic elements in soils is accurately established using a combination of efficient Q-learning and sparse partial least squares.