The specification of environmental extrema is a persisting problem, especially in soil with spatial heterogeneity owing to anthropogenic activities. Using a geographic detector, a Bayesian spatial model, and a copula-based spatial model, methods of identification of extremal regions in a mine area were compared. The results are as follows. (1) All of the heavy metals in anthropogenic soil, including As, Cd, Cr, Hg, and Ni, had a weak random spatial heterogeneity, but Cd and As exhibited strong stratification spatial heterogeneity (q = 0.21** a nd 0 .11*, respectively). (2) The Cr, Hg, and Ni predictions are very similar for both models (the improvements in the mean absolute percentage error (MAPE) and R² are 5.88% at most and 3.29%, respectively). The copula-based spatial model outperformed the Gaussian spatial model in the predictions of Cd (MAPE: 12.12%; R²: 16.67%) and As (MAPE: 4.16%; R²: 7.89%). (3) Based on the comparison with the Gaussian spatial model using a Bayesian process, the identification of the extremal regions using the copula-based spatial model had a higher accuracy for the extreme samples. In general, the prediction obtained using the copula-based model revealed the probability of exceeding a certain threshold at a location. Moreover, it uses the copulas fitting of the samples’ spatial heterogeneity obtained through maximum likelihood estimation, rather than variogram fitting, resulting in the random spatial heterogeneity summing to a nugget, which preserves more information about the samples. Thus, we conclude that the copula-based spatial model can be used to predict the heavy metal concentrations in soil with weak random spatial heterogeneity but strong stratification spatial heterogeneity.