DeepSatLoc: A Multimodal Fusion Strategy for Enhanced Localization in Urban Canyons Exploiting GPS and LEO Satellite Communication Signals

The positioning accuracy provided by the Global Positioning System (GPS) heavily deteriorates in urban canyons. Recent research has leveraged deep learning models trained on large GPS and environmental information datasets to enhance pseudorange and position estimates, mitigating the urban canyon effect. In addition, signals of opportunity (SoOP) from Low Earth Orbit (LEO) communication satellites offer a promising approach for localization, particularly in GPS-denied environments. This paper presents an algorithm that combines both strategies, employing a multi-modal fusion of localization information from GPS and LEO-based systems to achieve enhanced robustness and accuracy in position tracking. The proposed system exploits first a deep neural network (DNN) to refine the GPS position estimate by leveraging only GPS data, eliminating the need for detailed environmental information required in previous methods. Then, it integrates the refined GPS estimate with measurements from LEO satellites and an inertial measurement unit (IMU) using a Kalman filter, further improving localization precision. Simulation results, based on ray tracing-generated satellite channels, demonstrate that the method achieves an accuracy of approximately 10 meters in challenging deep urban canyon scenarios.