An integrated CBLA-Net with fractional discrete wavelet transform and frequency-based CARS to predict heavy metal elements by XRF.

Background: X-ray fluorescence (XRF) emerges as a promising technique for estimating heavy metal elements. However, XRF spectra typically contain a significant amount of environmental information and signal noise, and the relationship between spectral intensity and element concentration is difficult to quantify using a single model, thereby reducing the predictive performance for low concentration elements.

Results: This paper proposed a comprehensive framework for predicting elemental concentrations, encompassing preprocessing, variable selection, decision-making, to enable fast, non-destructive, and accurate estimation of element concentrations in soil. Firstly, an optimal denoising method based on fractional discrete wavelet transform (FDWT) was introduced to enhance signal quality. Furthermore, the frequency-based competitive adaptive reweighted sampling (FCARS) algorithm was employed for feature selection of XRF spectral variables, allowing extraction of the most informative features from the complex spectral data. Finally, a novel deep learning network, called ConvBiLSTM-Attention (CBLA-Net), was designed to achieve precise estimation of heavy metal elements concentration. Compared with other advanced algorithms, The CBLA-Net demonstrated the highest accuracy for V, Cr, Mn, Zn, Cd, and Pb, achieving the coefficient of determination (R) of 0.9730, 0.9874, 0.9952, 0.9921, 0.9518, and 0.9741, respectively. The CBLA-Net not only effectively extract local features and capture global information, but also combines attention mechanism to focus on key information.

Significance: The proposed novel deep learning quantitative framework, including preprocessing, feature selection, and CBLA-Net decision-making, significantly enhances the accuracy of elemental content prediction. It provides a new approach for accurately assessing the concentration of heavy metal elements in soil.