Autoencoder Satellite Image Matching for UAV Geolocation in Long-Range High-Altitude Missions

Vision-based geolocation is a promising way to overcome the vulnerabilities of Global Navigation Satellite System (GNSS) methods, which are subject to signal degradation, intentional interference, and environmental obstacles. This paper presents a novel approach to Unmanned Aerial Vehicle (UAV) geolocation in long-range and high-altitude missions using satellite imagery. Our method is based on the matching of encoded vector representations in embedded space, demonstrating robust performance to changes in vegetation and landscape. The neural network is used to encode satellite images of a reference map into embedding representations. Image matching is performed in this embedded space using cross-correlation. We evaluated the accuracy and processing time of the proposed model by querying images along a 200 km northbound path at high altitude, covering an area larger than twenty thousand square kilometers. We also evaluated the network's generalization capability on an unknown map. Reference and query images are sourced from satellite images captured at different acquisition times to evaluate robustness due to appearance variations. The results demonstrate that the method can achieve up to 96.83% accuracy on a known map, while experiments on an unknown map averaged 90% accuracy. The processing time to match encoded images is 0.05 ms. These findings suggest the feasibility of integrating the method into more complex vision-based geolocation systems.