Heterogeneous image matching is a key technology for unmanned aerial vehicle (UAV) self-localization in environments where global positioning System (GPS) signals are denied. This paper addresses the challenges traditional UAV captured ground images face when matched with satellite images, such as the extensive image retrieval and feature point matching, the complex design of dual-input networks, and the lack of real-time capabilities. It transforms cross-modal image localization into a multi-class image problem and uses an end-to-end classification network based on ResNet18 to achieve UAV self-localization. To tackle the issue of complex dual-input network designs, this paper proposes a method for preparing a dataset that integrates aerial images with satellite imagery by categorizing different areas of a satellite image into different classes and establishing a category-to-geographic information mapping library. By comparing with mainstream lightweight algorithm models, the feasibility of the proposed method is validated.