This dissertation proposes a multi-temporal reference-guided framework that integrates colorization and super-resolution for restoring historical panchromatic aerial photographs under ground-truth-absent conditions. Unlike conventional image enhanceme...
This dissertation proposes a multi-temporal reference-guided framework that integrates colorization and super-resolution for restoring historical panchromatic aerial photographs under ground-truth-absent conditions. Unlike conventional image enhancement, it treats aerial photographs as geospatial data and therefore preserves their coordinate reference system, ground sampling distance, and GeoTIFF metadata throughout restoration. The main contribution is a geospatially consistent restoration pipeline that couples multi-temporal color references, change-aware color transfer, comparative super-resolution, metadata-preserving GeoTIFF/BigTIFF output, and a multi-layer evaluation strategy that addresses the absence of contemporaneous color ground truth.
The study targets Yeonje-gu, Busan, using three panchromatic epochs(1995, 1999, 2002) as restoration targets and three color epochs(2004, 2008, 2010) as references aligned to a common 2004 grid. Change detection separates stable and changed areas. In stable areas, chromatic information from the multi-temporal references is transferred with relatively higher confidence, whereas changed areas are estimated more cautiously using a K-means-based multi-temporal color palette; in both cases the original luminance is preserved and only the chromatic components are estimated, confining the risk of plausible but misleading hallucination to the chromatic dimension. On the colorized outputs, three ×2 models—SwinIR, BSRGAN, and RealESRGAN—are compared under identical conditions, and all products are stored as BigTIFF with updated affine transforms for direct use in GIS.
Because no contemporaneous ground truth exists for the target epochs, results are interpreted through an uncertainty-aware, multi-layer evaluation—synthetic-pair and leave-one-out cross-validation, self-consistency, no-reference metrics, and change-based inspection-priority maps—rather than as claims of absolute color accuracy. Detected change ratios decrease monotonically with temporal distance (17.41%→15.10%→13.95%), the changed-area HSV KL divergence likewise decreases (1.751→1.331→1.067), and luminance consistency remains near 0.998. Cross-validation shows that the change-detection-and-palette mechanism yields a consistent gain (+0.594dB PSNR across all epochs), whereas the multi-temporal ensemble provides a larger but conditional benefit (+0.822dB on average, with an exception at the 2004 target); the super-resolution models exhibit clear trade-offs, implying use-dependent selection. The restored images are therefore algorithmic estimates that support archives and historical GIS as auxiliary geospatial layers, not substitutes for contemporaneous color photographs or legal survey records.