Objective: This study evaluates the effectiveness of Real-ESRGAN, a deep learning-based super-resolution method, in improving the radiometric and geometric quality of UAV imagery for urban applications. While UAV photogrammetry facilitates the generation of 3D models and orthophotos, its limited spatial resolution restricts accuracy in detailed urban analysis. Super-resolution techniques, particularly those based on deep learning, offer a promising solution by reconstructing finer spatial details from low-resolution inputs.
Method: High-resolution images were reconstructed from UAV-based aerial photographs using the Real-ESRGAN model. These outputs were compared against bicubic interpolation and original datasets to assess visual, radiometric, and geometric improvements. Photogrammetric products, including orthoimage mosaics and 3D mesh models, were generated from each image type. Standard quality metrics (e.g., RMSE, ERGAS, SAM, GRMSE) were used for evaluation.
Results: Real-ESRGAN substantially outperformed bicubic interpolation in radiometric quality, showing improvements of 57.76% in RMSE, 100% in ERGAS, and 56.59% in SAM. It also improved geometric accuracy in derived products, reducing average positional errors by 18.1% in 2D orthoimages and up to 93.2% in 3D mesh reconstructions based on Root Sum of Squares (RSS) analysis.
Conclusions: The findings demonstrate that Real-ESRGAN can effectively enhance both visual quality and spatial accuracy of UAV-derived imagery and photogrammetric products. However, slight geometric inconsistencies in raw SR images suggest a trade-off between perceptual enhancement and geometric fidelity. Future research should explore geometry-aware super-resolution models that integrate spatial constraints and training strategies suited for geospatial applications.