Self-orthorectification of satellite images using multi-view geometric constraints without external geospatial data

초록

Traditional orthorectification methods rely on ground control points (GCPs) and external digital elevation models (DEMs), which may be unavailable or costly to obtain. This study presents a self-orthorectification framework for high-resolution satellite imagery that eliminates the need for such external references by using only overlapping images and their rational polynomial coefficients (RPCs). The proposed method extracts tie points under RPC-based geometric constraints, performs multi-view bundle adjustment to jointly refine relative image orientation and reconstruct 3D ground coordinates, and generates a virtual terrain model for orthorectification. Experiments on two KOMPSAT-3A datasets covering urban and mountainous regions showed that the proposed method reduced the 2D check-point MAE from 8.391 to 0.355 pixels in Seoul and from 2.335 to 0.454 pixels in Jeju, while the corresponding RMSE decreased from 10.151 to 0.512 pixels and from 3.117 to 0.686 pixels, respectively. Compared with existing GCP-free image registration methods that mainly improve relative alignment and DEM-assisted orthorectification approaches that still depend on external elevation data, the proposed framework enables internally consistent orthorectification using only the input image block. The reported accuracy reflects internal geometric consistency among overlapping images rather than absolute planimetric geolocation accuracy.

키워드