# A Spatial Point Feature-Based Registration Method for Remote Sensing Images with Large Regional Variations

## TL;DR

This paper introduces a new method for accurately aligning remote sensing images, even when there are large differences in resolution and regional features.

## Contribution

The novel method uses spatial point features and a new edge keypoint extraction technique to improve registration accuracy in challenging image pairs.

## Key findings

- The proposed method achieves nearly 100% registration precision with pixel-level accuracy.
- It demonstrates strong rotation and scale invariance through extensive testing.
- The method outperforms existing algorithms in handling images with large regional variations.

## Abstract

The accurate registration of image pairs is an indispensable key step in the process of disaster assessment, environmental monitoring, and change detection. However, obtaining correct matches from input images is difficult, especially from images with significant resolution and regional variations. The current image-registration algorithms perform poorly in this application scenario. In this article, a spatial point feature-based registration method is proposed for remote sensing images with large regional variations. First, a new edge keypoint extraction method is designed that selects points with gradient magnitude maxima around the neighborhood of the edge line segments as keypoint features. Then, the feature descriptors for each keypoint are constructed based on the geometrical distribution (distance and orientation) of each keypoint. Considering the stability of the distribution of the edge contours, our constructed descriptor vectors can be well used for image pairs with large resolution and regional variations. In addition, all feature descriptors in this method are constructed and matched in the rotated image pyramid. Finally, the fast sampling consensus algorithm is applied to eliminate mismatches. In test images with various scales, rotation angles, and regional variations, the proposed method achieved pixel-level root mean square error, and the average registration precision is nearly 100%. Meanwhile, our proposed method’s rotation and scale invariance are verified by rotating and downsampling the image pairs extensively. In addition, compared with the comparison algorithms, the method proposed in this paper has better registration performance for images with resolution and regional variations.