This paper investigates source localization in urban environments using machine learning methods. Both classification and regression schemes are examined using the random forest algorithm. In both approaches, the localization performance depends mostly on arrival time information of received signals. It is shown that the use of a relative arrival time difference rather than the direct time of arrival (TOA) also provide good performance, which is beneficial for practical application. It is found that with enough number of receivers, the signal power parameter can be omitted, which eliminates frequency dependency of developed models. It is also found that at least three receivers should be used to achieve acceptable prediction performance. Additionally, the number of training examples needed depends on the approach used as well as the desired level of accuracy. This factor is more important in the classification approach, as the amount of training data required closely relates to the number of sectors created by the localization problem. A regression scheme is more natural for predicting spatial coordinate values, and achieved higher localization accuracy with fewer training examples as compared to classification. Ultimately, the regression based localization scheme using the time difference of arrival (TDOA) parameter at three receiver locations achieved an average localization accuracy of 2.8m.