Images provided by the European Copernicus Sentinel-2 satellites are valuable and easily accessible sources of remote sensing data for tasks across various fields. These data have a high spectral and temporal resolution, but a rather low spatial resolution, limiting their applicability for many tasks. In agricultural tasks, such as crop monitoring of small land parcels, the use of these data for fine-scale analysis is contingent upon the enhancement of spatial resolution while maintaining spectral fidelity. In this work, we propose a comprehensive single-image super-resolution reconstruction workflow that ensures both properties and is divided into two parts. First, a deep learning-based super-resolution reconstruction approach is applied to improve the spatial resolution of multi-spectral Sentinel-2 images to 2.5 m. For this purpose, a novel method is applied to achieve super-resolution of multiple spectral bands where associated real-word reference data is only partially available. It learns to increase the spatial resolution while preserving spectral accuracy of 10 m bands using high-resolution data from an auxiliary satellite with spectral correspondence, and 20 m bands without reference data using synthetic Sentinel-2 pairs. Second, the suitability of the method to subsequent agricultural tasks is evaluated by measuring the discrepancy between the super-resolved and reference data through a novel spectral knowledge-based validation method. This method leverages mappings of reflectances to spectral categories that enable assessing the spectral fidelity of super-resolved outputs, which is complementary to existing image quality assessment metrics, but with greater depth. The promising spectral validation results suggest that our super-resolution reconstruction pipeline has a great potential for agricultural applications.
Major et al. (2025) studied this question.