RPC Bias Compensation of KOMPSAT-3 Optical Imagery Based on KOMPSAT-5 SAR

Vendor-provided rational polynomial coefficients (RPCs) for high-resolution optical satellite imagery often contain systematic biases that are commonly corrected using field-surveyed ground control points (GCPs). Acquiring such points can be costly or infeasible in inaccessible regions. This study proposes an RPC bias compensation framework for KOMPSAT-3 (K3) optical imagery based on the geometric accuracy and stability of KOMPSAT-5 (K5) X-band SAR enhanced high-resolution (EH) imagery, without measuring image coordinates of control points. Three-dimensional ground control data (GCD) are projected into both sensor image spaces via the K5 Range–Doppler (RD) model and the K3 RPC model. A GCD-driven projective transformation provides coarse alignment, followed by fine registration using FFT-based phase correlation (FPC). Registration-derived correspondence coordinates are then used to estimate a strongly regularized image-space affine bias model, yielding a predominantly translational correction. Three GCD sources—Korean public control points (PCPs), ICESat-2 altimetry, and the 30 m SRTM DEM—were evaluated using one K3–K5 image pair over Suncheon, Korea. A sensitivity analysis showed that control points on flat terrain (elevation ≤60 m, slope ≤6°) yielded stable results, whereas points on steep or forested terrain degraded the registration, and increasing the number of GCD points did not systematically improve accuracy. Under these conditions, all three sources reduced the RMSE at independent GNSS checkpoints from 7.1 and 10.6 pixels in the column and row directions to 1.6–2.0 and 1.6–2.1 pixels, respectively. This accuracy is comparable to that of conventional GCP-based compensation (1.7/1.3 pixels), whereas two structural-feature-based matching methods (HOPC and CFOG) did not improve on the original RPCs under the tested implementations. These results demonstrate the potential of SRTM-derived control data for meter-level directional refinement of optical imagery when reliable SAR–optical registration is achieved. Generalization beyond the tested image pair requires further validation.