As drones become increasingly utilized across various fields, related multimedia applications are also emerging. One significant application is cross-view geo-localization, which leverages aerial drone and satellite imagery data to facilitate drone navigation and geo-localization. In this paper, we focus on the robustness and generalization of retrieval under various extreme weather conditions. Considering the significant gap between training and testing data, our research emphasizes exploring and employing a powerful self-supervised backbone and an unsupervised aggregator to achieve domain adaptation. Additionally, from a data perspective, we simulate various weather conditions to bridge the gap between training and testing drone data through data augmentation. Futhermore, a cross-weather triplet loss is utilized to minimize the domain differences between drone and satellite images under extreme weather conditions. Our method achieves 94.07% Recall@1 accuracy on University-160k-WX, and ranks 4th in the UAVM2024 Challenge. Code will be released at https://github.com/SunJ1025/WAGL.