Localization without Global Navigation Satellite Systems (GNSS) is an area of interest forautonomous operations of aerial vehicles. A promising navigation method involves using onboardimages and comparing them to geo-tagged reference images for global localization. This studyinvestigates algorithms for global localization of flying vehicles at altitudes around 1-3 km usingimages. The focus is on matching onboard camera images with georeferenced images undersignificant appearance variations due to seasonal changes or different image sources. Fourmethods are evaluated: two traditional correlation techniques, the cross mutual informationfunction (CMIF) method, and the Deep Phase Correlation Network (DPCN) which uses neuralnetworks. Synthetic data from the X-Plane 12 flight simulator, featuring built-in graphics andsatellite imagery, is used to generate datasets over varying locations and under diverseconditions. The results indicate that the DPCN method, combining deep learning and correlation,achieves the highest accuracy across most test scenarios. This study underscores the potentialof DPCN for robust aerial vehicles image matching in GNSS-denied environments, while alsonoting its limitations in challenging weather conditions. Future improvements could involve largertraining datasets, the use of real-world images, and integration with additional navigation methodslike inertial and visual odometry.

No description available.