Billions of wireless devices are interconnected to provide services to many
aspects of life and form The Internet of Things. These devices which are
often battery-powered and energy efficient can benefit greatly from an ac-
curate localisation service that does not consume extra energy. Several loc-
alisation methods have been developed for Low-Power Wide-Area Networks
(LPWANs), with LoRa being of particular interest thanks to its long range
and cost effectiveness. Time Difference of Arrival (TDoA) is a common way
to find location in a LoRa network which works well in open areas but poorly
in the harsh radio environment of cities. In indoor settings where the radio
environment is more complicated than outdoor, RSSI fingerprinting tech-
niques have been sucessfully used for positioning using WiFi and Bluetooth,
with state-of-the-art solutions employing Artificial Neural Networks (ANN).
This work aims to provide accurate localisation in an urban LoRa network,
using an ANN-based fingerprinting approach. Two publicly available data
sets collected in the cities of Utrecht and Antwerp are used to evaluate our
method. We show that the ANN model can be trained on these data sets to
predict location with mean errors between 411m and 581m. We determine
that the presence of gateways in the fingerprint plays a major role in the
ANN’s estimation but RSSI information is crucial in improving the accuracy.
To realistically compare the ANN approach to TDoA, we train and test the
neural network with chronologically split data. Our ANN approach achieves
a mean error of 500m with 90% of cases having errors below 1070m. This
RSSI fingerprinting method is more effective than TDoA at limiting large
localisation errors in cities.