Abstract: Due to the constraints of soil monitoring cost and project cycle, there are often monitoring data missing on contaminated soil sites. Accordingly, how to obtain more information from limited data is the focus of current study. This paper studied a site contaminated by metal sheds and used multivariate statistical methods to analyze the correlation between heavy metals (As, Zn, Cu, Pb, Ni, Cd, Cr) and polycyclic aromatic hydrocarbons (PAHs) (BaP, DBA, BkF, BbF, BaA, Nap, Chr) in soil samples. On this basis, a BP neural network model was established by using the known data to predict the content of heavy metals and PAHs in missing soil samples. The results suggested that: compared with the risk screening values in Soil Environment Quality Risk Control Standard for Soil Contamination of Development Land (GB 36600-2018), the rate of heavy metals exceeding the standard was w(Ni) > w(Cu) > w(As) > w(Pb) > w(Zn)=w(Cd) > w(Cr). Except that w(Chr) did not exceed the standard, the over-standard rate of the other six PAHs was w(BaP) > w(DBA) > w(BbF)=w(BaA) > w(Nap) > w(BkF). Zn and Pb, As and Cd correlated better, Cu and Ni correlated better, Cr correlated poorly with the other six heavy metals. Except for Nap, BaP, DBA, BkF, BbF, BaA and Chr had good correlation with each other in PAHs. The coefficient of determination (R₂) of the actual values of contaminant concentrations and the predicted values of the BP neural network model was in the range of 0.812-0.993. The Nash efficiency coefficient (NSE) was in the range of 0.779-0.959. The root means square error (RMSE) and mean absolute error (MAE) were small. The research showed that the contents of heavy metals and PAHs generally exceeded the standard to different degrees in the study area. The BP neural network prediction model was accurate and reliable to predict contaminant concentration. It was feasible to use BP neural network model for spatial analysis and evaluation of soil pollution. And weakly correlated factors as input parameters could equip the prediction model with higher precision.