Geometric Accuracy Analysis of Tie Point-Based RFM Refinement Using Bundle Adjustment Framework Seunghwan Ban, Taejung Kim
Inha University
Abstract
Rational Function Models (RFMs) are commonly used for georeferencing satellite imagery due to their flexibility and sensor-agnostic nature. However, initial RFM parameters often contain inherent geolocation errors resulting from limited precision of attitude and orbit data used for sensor modeling. As a result, the geolocation accuracy of RFM-based products can be limited without ground control points (GCPs). This study analyzes the accuracy of our bundle adjustment framework that refines RFM parameters using only tie points extracted from overlapping satellite images. experiments were conducted by incrementally increasing the number of images from 2 to 9. We evaluated multi-view redundancy on adjustment stability during the tie point-based RFM refnment framework. In this framework, the bundle adjustment was performed in a relative model space using tie points rather than being fitted to absolute coordinate frame using GCPs. to ensure the adjustment results were quantitatively superior to the original RFM solution, GCPs were used only to evaluate the changes in absolute geolocation accuracy before and after the adjustment. Reprojection errors (RMSE) of the modeling tie points were consistently below 0.5 pixels across all cases, indicating stable adjustment performance. Independently acquired manual checkpoints from two images showed reprojection errors below 0.8 pixel. Subsequent evaluation using GCPs showed that image-space reprojection errors were reduced by approximately 1 to 5 pixels in most cases compared to the original RFM and the errors were dependent on the image configuration. Importantly, the spatial patterns of the residual errors exhibited directionally coherent trends, suggesting the possibility of systematic post-adjustment correction. These findings indicate that the proposed framework may be extended by incorporating minimal GCP input or reference image constraints. This would allow the adjusted model to be anchored to the absolute coordinate space, thereby enhancing global correction pipelines without full GCP dependence.
