A selection of soil samples from reclaimed mining areas were prepared to establish the quantitative inversion models of the soil heavy metal (As, Zn, Cu, Cr, and Pb) concentrations. The concentrations of the soil heavy metals and the visible and near-infrared spectra of the soil samples were obtained in a darkroom. Firstly, smoothing processing was used to smooth the noise in the original spectra, and the spectral transformation techniques of first derivative (FD), continuum removal (CR), and standard normal variate (SNV) were used to promote the model stability and the accuracy of the prediction. Through correlation analysis, the feature bands of the different transformed spectra were extracted. Finally, three different inversion models were adopted and compared, i.e., traditional multiple linear regression (MLR), partial least squares regression (PLSR), and least squares support vector machines (LS-SVM) modeling. The results indicated that: (1) the stability and accuracy of the inversion models established by the different transformed spectra was high, in which LS-SVM was better than PLSR, and PLSR was better than MLR (except for a few cases);and (2) the spectral features extracted from the different transformed spectra had a certain influence on the inversion model, in which the results based on CR transformation and SNV transformation were better than the FD transformation. Therefore, the quantitative estimation of heavy metal concentrations by the use of reflectance spectroscopy is feasible, and the pre-processing is essential to improve the accuracy of the model.