This paper proposes a novel end-to-end self-supervised learning-based method, dynamic memory-driven and neighborhood information learning (DMNIL), to overcome the high annotation costs and limited transferability of existing supervised learning-based methods for drone view geolocation (DVGL). DMNIL generates pseudo labels using a clustering algorithm and learns discriminative within-view representations via a dual-path contrastive learning framework. Furthermore, it includes a dynamic hierarchical memory learning (DHML) module that combines short- and long-term memories to enhance within-view feature consistency and discriminability, and an information consistency evolutionary learning (ICEL) module that captures implicit inter-view semantic correlations using a neighborhood-based dynamic constraint mechanism to improve feature alignment across views. By stabilizing and enhancing the self-supervised learning process through a pseudo-label refinement strategy, it outperforms existing self-supervised learning methods and several state-of-the-art supervised learning methods on three public benchmark datasets.
