Deep spectral feature fusion network combined with attention mechanism to predict heavy metal elements in soil by XRF

Abstract

With the rapid advancement of industrialization and urbanization, the accumulation of heavy metal elements in soil has become a significant environmental and public health concern. Accurate determination of heavy metal concentrations is therefore of great practical significance for preventing pollution, guiding ecological restoration, and formulating rational agricultural policies. X-ray fluorescence (XRF) spectroscopy technology has been widely used in soil heavy metal analysis due to its advantages of rapidity, non-destructiveness, and in situ measurement. However, due to the matrix effect of soil XRF and spectral line interference, obtaining high-precision detection results using traditional analytical methods is challenging, leading to significant bias in estimating heavy metal concentrations. To address this challenge, this paper proposes a deep feature fusion model (Spec-SAFFNet) that integrates a feature fusion module (FFM) with a self-attention-enhanced one-dimensional convolutional neural network (SA-1DCNN). Spectral data collected by a handheld energy-dispersive XRF device are first input to FFM to extract crucial fused features. Then, an SA-1DCNN is proposed to capture complex relationships between spectral fused features and the concentration of a single heavy metal element, enabling accurate prediction. Compared with other advanced soil analysis algorithms, the experimental results show that Spec-SAFFNet achieved coefficients of determination (R2) values of 0.9324, 0.9774, 0.9622, and 0.9577 for V, Cu, Zn, and Pb, respectively. The above results, along with a frequency-domain interpretable analysis, demonstrate that Spec-SAFFNet effectively provides a practical and promising approach for quantitative soil heavy metal analysis.