A method for geolocating a terminal of a wireless communication system, based on a learning method making it possible to estimate the geographical position of a terminal using both a radio signature corresponding to a set of values representative of the quality of radio links existing between the terminal located at the sought position and a plurality of base stations of the wireless communication system, as well as a reference data set associating radio signatures with known geographical positions. To limit the complexity of the learning algorithm and to make it resistant to topology changes of the access network, each radio signature contains a selection of N values among the set of measured values, as well as the geographical positions of the corresponding base stations.