Wireless underground sensor networks (WUSNs) based on magnetic induction (MI) have been recently proposed as a promising candidate for underground networking. The benefit of the MI-WUSNs compared to other solutions (e.g. so-called Through-The-Earth communication) is related to the substantially lower path loss and lower vulnerability to the changes of the soil properties. In the past, some efforts have been made to characterize the signal transmission in MI-WUSNs. Those investigations, however, refer mostly to the information transmission. One of the target applications of the WUSNs is the object localization in the underground medium, which remains an open issue due to the complicated characteristics of the MI channels corrupted by the influence of soil. In this work, we propose a machine learning based solution for localization. In addition, a novel passive localization technique is introduced, which requires no signal from the target node and thus proves useful for rescue operations, where the battery of the node to be localized is either empty or damaged.