X-ray fluorescence (XRF) spectroscopy is a crucial technique for the rapid on-site detection of heavy metals. However, when applied to soil heavy metal analysis, the presence of high-intensity and complex background spectra due to soil matrix effects significantly hinders the accurate acquisition of characteristic spectral peaks and the precision of quantitative analyses. To address this issue, this paper proposes a background subtraction method for soil XRF spectra, combining peak-valley recognition using an extremum method with penalized correction for cubic smoothing spline fitting, termed PIEspline. Initially, the extremum method identifies peak and valley points in the complete soil XRF spectra to extract data points representative of the background. These points are then used to fit a cubic smoothing spline curve with penalized corrections, forming the background baseline and thus enabling the subtraction of complex backgrounds from soil XRF spectra. The performance of the PIEspline method is further validated by comparing it with three traditional spectral background subtraction methods: adaptive iteratively reweighted penalized least squares (airPLS), iterative wavelet transform (IWT), and statistical sensitive nonlinear iterative peak-clipping (SNIP). The results indicate that, for simulated soil XRF spectra, the root mean square errors (RMSE) between the background spectra obtained by the PIEspline method and the true background spectra are 0.425 8 and 0.644 1, respectively, which are lower than those of the other three methods. Additionally, the PIEspline demonstrates the fastest background subtraction efficiency. For three different soil types cinnamon soil, saline-alkali soil, and loess and three different soil uses agricultural, industrial, and construction the average relative error of fluorescence intensity at 10 characteristic valley points in the XRF spectra fitted by PIEspline is 10.87%. Compared to the traditional methods, this error is reduced by 84.88%, 76.30%, and 16.51%, respectively. Furthermore, for quantitative analysis of Cr, Pb, and Cd in the aforementioned six soil types, the average relative errors using PIEspline are 4.01%, 2.50%, and 5.20%, respectively, representing reductions of 22.39% to 84.07%, 60.15% to 71.92%, and 79.18% to 84.07% compared to airPLS, IWT, and SNIP. Notably, the PIEspline method exhibits minimal fluctuation in relative error when soil type and use change, showcasing superior stability. This suggests that the PIEspline method offers the best applicability for simultaneous XRF quantitatively analyzing multiple heavy metals across various soil types and uses. Therefore, the PIEspline method proposed in this study enables precise background subtraction of XRF spectra from different soil types and uses, improving heavy metal XRF quantitative analysis accuracy. This research provides a vital methodological foundation for the rapid and accurate on-site detection of soil heavy metals using XRF.