UAVs increasingly require GNSS-independent positioning for operation in contested or infrastructure-denied environments. This paper presents a vision-based celestial navigation system with automatic adaptation to light pollution through dynamic star catalog selection. The algorithm employs DDR pattern matching with novel polar-star rejection and consensus-driven magnitude refinement to robustly identify observable stars under varying environmental conditions. Evaluation on 200 synthetic night-sky images demonstrates substantially improved star identification robustness compared to fixed-catalog baselines, achieving 71.5% recall at visual magnitude 7 (Bortle 3) and maintaining non-zero performance under severe light pollution (27.6% recall at magnitude 5.0 and 4.5% at magnitude 4.5), where the baseline fails entirely. Across higher limiting magnitudes (6.5–8.0), the adaptive method consistently attains 71.5–82.5% recall. Including misidentifications, the end-to-end system achieves a median geolocation error of 6.80~km, supporting coarse global localization, GNSS integrity monitoring, and long-duration drift bounding in GNSS-denied environments. These results indicate that adaptive catalog selection significantly extends the operational envelope of celestial navigation into light-polluted conditions previously considered infeasible.