Without Paired Labeled Data: End-to-End Self-Supervised Learning for Drone-View Geo-Localization

Area of Science:

• Computer Vision
• Geospatial Intelligence
• Machine Learning

Background:

• Drone-view geo-localization (DVGL) traditionally requires extensive labeled drone-satellite image pairs for supervised learning.
• Existing methods struggle with distribution shifts and high annotation costs, limiting practical open-world deployment.
• Limited transferability of current DVGL techniques hinders adaptability to new geographical regions.

Purpose of the Study:

• To propose a novel end-to-end self-supervised learning method for DVGL that overcomes limitations of supervised approaches.
• To enhance drone localization accuracy and robustness in diverse, open-world scenarios.
• To reduce reliance on costly annotated data and improve model transferability.

Main Methods:

• Developed the dynamic memory-driven and neighborhood information learning (DMNIL) method, featuring a shallow backbone network.
• Employed a clustering algorithm for pseudolabel generation and a dual-path contrastive learning framework for intra-view representation learning.
• Integrated dynamic hierarchical memory learning (DHML) for feature consistency and information consistency evolution learning (ICEL) for cross-view alignment, enhanced by a pseudolabel enhancement (PLE) strategy.

Main Results:

• The DMNIL method demonstrated superior performance compared to existing self-supervised DVGL techniques.
• The proposed approach outperformed several state-of-the-art supervised methods on three benchmark datasets.
• Achieved significant improvements in localization accuracy and robustness, validating the effectiveness of the self-supervised approach.

Conclusions:

• The DMNIL method offers a highly effective self-supervised solution for drone-view geo-localization, addressing key limitations of supervised learning.
• The integration of dynamic memory and neighborhood information learning significantly enhances feature representation and alignment.
• This research paves the way for more practical and scalable DVGL applications in real-world environments.