A Robust Multi-Modal Wide-Field Satellite Image Registration Algorithm Based on Weighted Random Partition Optimization

Continuous advancements in sensor technology have expanded the dimensions and diversity of satellite imagery, yet intensified geometric processing challenges for multi-modal remote sensing images (MRSI). Enlarged image scales amplify susceptibility to suboptimal alignments under heterogeneous correspondence point distributions, manifesting as localized misregistration and structural artifacts. To address these challenges, this paper presents a weighted random partition optimization (WRPO) method for the robust registration of widefield MRSI. The method contributes primarily in two ways: (1) The construction of a strong-weak information inter-feeding randomized optimization block model, which divides the widefield image into sub-regions and ranks local strong and weak feature blocks. Random sampling and overlapping of strong and weak blocks are employed to ensure a uniform distribution of corresponding points. (2) The development of multi-channel directional gradient feature descriptors, which enable efficient multi-directional feature matching for fast registration. The experimental framework employed four MRSI datasets to benchmark WRPO against six state-of-the-art algorithms. Quantitative results establish key advantages of WRPO: (1) The proposed method achieves a 400% higher matching efficiency compared to five benchmark methods; (2) Its sub-pixel matching precision effectively eliminates local misregistration and artifacts; (3) WRPO uniquely succeeds in processing wide-field images. The algorithm exhibits exceptional generalization— integration with baseline methods enhances their performance. WRPO achieves efficient high-precision registration with dense correspondence points, while maintaining strong transferability across heterogeneous satellite data. The dataset and code are available at https://github.com/crazy0930/WRPO.