Rotation and seasonal invariant UAV-satellite image matching via polar coordinate transformation and multi-scale fusion

Cross-platform image registration, particularly between Unmanned Aerial Vehicle (UAV) and satellite imagery, is a pivotal technology in intelligent Earth observation and geospatial analysis. However, such registration tasks frequently confront two major challenges: significant rotational discrepancies arising from different imaging platforms and flight trajectories, and dramatic cross-seasonal variations (e.g., snow cover, vegetation changes) in ground object appearance due to large temporal acquisition gaps. Existing methods often exhibit insufficient robustness and accuracy when dealing with these complex conditions. To address these issues, this paper proposes a novel feature extraction and matching framework that inherently possesses both rotation invariance and seasonal invariance, specifically designed for UAV-satellite image registration. To effectively achieve rotation invariance, we integrate a polar coordinate transformation-based module into the feature extraction network, ensuring the extracted features are highly stable against arbitrary rotations. For the substantial differences induced by cross-seasonal variations, we devise a multi-scale feature fusion strategy to capture multi-level semantic and structural information, thereby effectively coping with drastic changes in surface texture and brightness caused by seasonal shifts (e.g., snow cover). The framework further enhances registration accuracy by combining coarse-grained and fine-grained feature matching. Experimental results demonstrate that, compared to existing mainstream methods, the proposed approach exhibits significantly superior registration accuracy and robustness on UAV-satellite image pairs characterized by large rotational angles and cross-seasonal differences, providing strong technical support for multi-source image fusion and analysis in complex environments.