Heavy metal pollution in the soil affects the yield and quality of crops. The traditional detection method has complicated procedures, high detection costs, and slow detection speed. The X-ray fluorescence (XRF) analysis technology to detect heavy metal content in soil has the advantages of being simple in processing, on-site, rapid and non-destructive. Due to the complex soil background including much noise and irrelevant information, before the establishment of the XRF correction model, the pre-processing of the spectrum can effectively remove irrelevant information and maintain useful information, which has an important influence on the accuracy of the XRF prediction model.This article focuses on the effects of spectral pre-processing method on the accuracy of heavy metal content prediction model. Firstly, forward interval partial least squares (FiPLS) was taken as a correction model to compare the detection accuracy of the heavy metal model in eight different conditions, namely non-pre-processing, detrending processing (DT), standard normal variable transformation (SNV), multiple scatter correction (MSC), wavelet denoising (WT), SNV+DT, convolution smoothing (SG) + first derivative and convolution smoothing (SG) + second derivative. The preliminary results showed that the multiple scatter correction pre-treatment method had desirable effects. Compared with the original spectrum, the determination coefficient R rised from the original 0.988 to 0.990, and the prediction of root mean square error (RMSEP) and the relative error respectively declined from the original 20.809 and 0.166 to 19.051 and 0.121. Secondly, on the basis of the multi-dimensional scattering correction pre-processing method, the localized weighted linear regression multiple scatter correction (LWLRMSC) and partial least squares multivariate scatter correction (PLSMSC) were proposed in terms of the restriction of describing non-linear relationships with linear representations, and the modeling effects of LWLRMSC and PLSMSC were compared. LWLRMSC was based on the weighted idea. In the prediction of the value of a point, the proper kernel function and weight distribution strategy were selected to perform linear regression of the prediction point, and the under-fitting condition of the simple linear regression was resolved. PLSMSC, based on the PLS modeling idea and taking into account the maximum correlation between the independent variable and the dependent variable, reduced the fitting error and distortion. The results showed that PLSMSC has the best pre-treatment effects. The R values of the predicted and actual values of the five heavy metals (Cu, Zn, As, Pb and Cr) were 0.989, 0.973, 0.991, 0.989 and 0.986, with the RMSEP respectively being 8.805, 58.360, 7.671, 12.549 and 20.851. Compared with the traditional MSC method, PLSMSC not only has a significant improvement in accuracy but also has better generalization performance. It can eliminate spectral noise and improve the contribution of effective information, thus providing theoretical support for the soil heavy metal content model to select the suitable pre-treatment method.