Soil heavy metal pollution caused by mining in mining areas seriously affects crop yield and causes human diseases. It is necessary to prevent soil heavy metal pollution from damaging health. Hyperspectral remote sensing can rapidly and dynamically acquire continuous spectra signals of ground objects, which provides a new idea for developing soil heavy metal content monitoring based on remote sensing. Aiming at the typical lead-zinc mining area in Laiyuan County, Hebei Province, soil samples from the mining area and surrounding areas are collected on-site, and the reflectance spectra of soil were obtained using SVC HR-1024i spectrometer (350~2 500 nm). Through the spectral data smoothing, first derivative (FD), multivariate scattering correction (MSC), standard normal variate transform (SNV), first derivative after multivariate scattering correction (MSC+FD), and first derivative after standard normal variatetransform (SNV+FD), six kinds of spectral transformations were performed. The difference index (DI), ratioindex (RI), and normalizeddifference index (NDI) methods were used to extract the spectral indices from the six pretreated data. The contents of heavy metals cadmium (Cd), lead (Pb) and zinc (Zn) in soil were obtained through laboratory chemical testing and analysis. Different spectral transformation methods pretreated different heavy metals. The optimal spectral transformation methods for heavy metal elements were obtained. The difference index, ratio index, and normalized vegetation index were used to extract the optimal band combination under different spectral indices to get the optimal independent variables for modeling different heavy metals. The inversion models of heavy metal elements were constructed based on random forest and partial least square method. The research indicated that the noise could be effectively reduced, and the spectral characteristics were enhanced by pretreatment of spectral data. The results showed that the correlation between the spectral data and the heavy metal content was improved after the pretreatment. The optimal independent variables for different heavy metal elements were selected to increase the practical features of inversion modeling. Random forest algorithm and partial least squares regression method were used to establish prediction models for three heavy metals: cadmium (Cd), lead (Pb), and zinc (Zn). The R2 of the optimal models reached 0.90, 0.91, and 0.84, respectively, which confirmed the validity of this research method. This study can provide a basis for the inversion modeling of soil heavy metal content in lead-zinc mining areas and a method reference for detecting soil heavy metal content in mining areas.