LoRa technology allows long-range transmissions with low power consumption and it can also be used indoor. For these reasons, the introduction of a precise timestamping of LoRa frames provides the possibility to use this technology for accurate localization in many scenarios. However, this is still very challenging to achieve in non-line-of-sight environments such as urban landscapes. In this paper, we present a “fingerprinting” method to perform outdoor geolocation based on machine learning (Random Forest and Neural Networks) applied to a reference map. The map combines Time Difference Of Arrival (TDOA) measurements generated by a LoRa network and GPS location as ground truth. We tested our approach on simulated data achieving promising results with a Root Mean Squared Error below 9 meters by using a Long Short-Term Memory (LSTM) network.