Recently, deep learning-based image super-resolution (SR) has emerged as an effective and cost-efficient solution for obtaining images that surpass the physical limitation of sensors. However, the training datasets commonly used for SR tasks—particularly in remote sensing—are constrained by limited scale and spatial resolution, as they are often developed for different purposes. To address these issues, we propose LACAS2K, a large-scale dataset of aerial and satellite images specifically tailored for image super-resolution of very high-resolution (VHR) satellite imagery. The dataset is constructed with a clearly defined target input spatial resolution of approximately 50 cm, aligning with the typical resolution of panchromatic (PAN) bands in VHR satellite imagery. Moreover, it provides high-quality high-resolution (HR) images organized into 2K-resolution patches for scale factors of 2 and 4, offering richer spatial features. The experimental results demonstrate that LACAS2K effectively supports the training of data-intensive models, such as transformer-based SR networks. The LACAS2K dataset will be made publicly available through IEEE DataPort, providing a valuable resource for advancing SR research in the remote sensing domain.



Keywords : Spatial resolution, Satellite images, Superresolution, Remote sensing, Training, Satellites, Earth, Internet, Transformers, Optical sensors