Center point matching network based on multi-coverage templates for high-resolution satellite orthorectification

High-resolution (HR) satellite images are provided with rational polynomial coefficients (RPCs) that map ground coordinates to image space. However, initial RPCs are derived from approximate models and contain geometric errors. To correct these errors, accurate matching of ground control point (GCP) locations in the image is required. Unlike conventional template matching methods that focus on global similarity, orthorectification requires precise localization of the template center. This study presents a deep learning-based template center point matching method that aims to reduce the limitations of conventional matching when used for orthorectification. The network extracts feature maps from both wide and narrow observation areas based on Siamese U-Net, and computes similarity maps through channel-wise fast Fourier transform-based cross correlation. A dual threshold strategy is applied to filter unreliable results by evaluating both the response strength and its spatial distribution. To train the model, patch pairs were automatically generated from temporally different aerial orthophotos. A filtering strategy based on multi-scale matching consistency was applied to select regions that exhibit robust alignment across observation ranges. The model shows stable performance, achieving an average root mean square deviation (RMSE) of 0.95 pixels for ground truth. This method is tested on satellite and aerial image pairs from multiple regions with varying land covers. The selected points were used for orthorectification, leading to improved geometric accuracy.