The research object of the paper is, based on the Huainan Chuangda coal gangue filling reclamation experiment area, to analyze soil heavy metals (Cu, As, Cr) with the traditional sampling method. Reflectance spectra of soil samples measured by Analytical Spectral Devices FiSpec4, spectral features are extracted, and the spectra are averaged with the first order differential, the second order differential transformation, and the inverse logarithmic transformation, etc. Correlation analysis of spectral characteristic parameters and heavy metal content in soil is conducted, therefore, the selection of the relevant bands is related to the relevant factors. Multivariate stepwise regression analysis, partial least squares regression and artificial neural network are used to establish the prediction model of soil heavy metals by using soil spectral reflectance. The experimental results show that the spectral bands of the differential transformation are significantly correlated with the content of heavy metals. For heavy metal Cu and Cr, the artificial neural network model of the first order differential spectrum is the best prediction model and the partial least squares regression model of the two order differential spectra of heavy metal elements is obtained by the best prediction results.