WorldView-3 Super-resolution Training Dataset Construction Using the MTF-GLP Method

The demand for high-resolution satellite imagery is rapidly increasing in various fields such as national defense, disaster response, and environmental monitoring, drawing significant attention to super-resolution (SR) techniques in remote sensing. In this study, a high-quality high resolution (HR)-low resolution (LR) training dataset was generated using a single sub-meter satellite image acquired from WorldView-3. The enhanced correlation coefficient image registration method was employed to correct geometric distortions between the panchromatic and multispectral images, and the modulation transfer function-generalized Laplacian pyramid algorithm was applied to produce HR images with improved spatial resolution while preserving spectral fidelity. LR images were synthesized through sensor-specific degradation modeling, resulting in a total of 14,776 HR-LR patch pairs. The constructed dataset covers diverse land cover types, including urban areas, agricultural fields, and water bodies, enabling comprehensive evaluation of super-resolution models in both spatial detail reconstruction and spectral information preservation. Unlike conventional datasets that rely on cross-sensor fusion with complex preprocessing, this study simplifies the data generation pipeline using a single-sensor approach and ensures spatial-spectral consistency. The proposed dataset is expected to serve as a reliable benchmark for training and evaluating deep learning-based satellite image SR models.