Quick and accurate detection of mine water source is of great importance for the safety of mine production. In recent years, the impact of human activities on different aquifers has increased. In order to improve the accuracy of water source in mine water, this paper propose to select seven common ion concentrations in groundwater, and the nitrate ion concentration as well as chemical oxygen demand which can fully reflect the traces of human activity as the water chemical discriminant index. And then the RF model would be used to identify the mine water source. Taking the Dagushan iron ore as an example to verify the validity of the selected indicators and methods. The data was input into the RF model for 100 cross-validation, and the verification results were compared with the SVM model and the ELM model. The results of the RF model are stable and the prediction accuracy is not susceptible to fluctuation. RF have a wide range of adaptation parameters in the modeling process. When the number of trees is 50, the training error tends to be stable, and changing the number of trees has no practical impact on the prediction results, while the other two models are more sensitive to parameter selection. The parameters of the RF can be obtained simply by the OOB error rate, but the other two model parameters need to be obtained through cross-validation. The RF model verifies the training samples, and the correct rate can reach 100%. The RF model also verifies the test samples, and the correct rate can reach 97.38%. The accuracy of the two correct rates is better than the SVM and the ELM; The random forest model has higher prediction accuracy and robustness, and has a better application prospect in the determination of mine water source.