이종 시기 항공영상 기반 템플릿 정합 학습 데이터셋 구축

This study proposes a patch generation and automatic refinement based train dataset construction method using bi-temporal aerial orthophotos to improve GCP (Ground Control Point) positional accuracy for RPC (Rational Polynomial Coefficient) correction and to overcome the limitations of existing deep learning training datasets. First of all, patch pairs centered on the same feature point are generated from bi-temporal orthophotos. Afterwards, we crop the patch pairs to different sizes based on the same center point to generate various matching scenarios involving structural changes and noise conditions. We applied the ZNCC (Zero-mean Normalized Cross-Correlation) to the generated patch pairs, compared the peak points derived from each patch, and removed inappropriate patches to build a dataset robust to matching. We performed matching on each aerial image pair and aerial image-satellite image pair using the Siamese U-Net model trained by the proposed dataset. As a result, the aerial image pair showed a lower RMSE (Root-Mean-Square Error) of 2.436 pixels compared to the ZNCC technique which showed a result of 2.659 pixels, confirming the effectiveness of using bi-temporal aerial orthophotos. In the aerial image-satellite image pair, the model based on the refined dataset achieved the lowest RMSE at 1.966 pixels, while the unrefined dataset actually showed lower performance. These results suggest that leveraging bi-temporal aerial orthophotos to construct a training dataset incorporating diverse temporal and structural variations can contribute to improving the accuracy of deep learning-based image matching.