This paper presents training and validation strategies for the application of a Single Image Super-Resolution (SISR) architecture to Sentinel-2 imagery using the SEN2VENµS open dataset, in order to generate 5m resolution images from the initial 10m and 20m bands. Our approach focuses on preserving radiometry and geometry of the input images and avoiding the introduction of artifacts, and leverages a smaller ESRGAN generator network, as well as high and low spatial frequencies separation in loss calculation. Bands B11 and B12, which are missing in the SEN2VENµS dataset, are also process by means of the Wald protocol. We demonstrate that those adaptations allow to enhance the resolution of all Sentinel-2 bands to 5m without substantial loss in reliability. Benefits of the 5 meter super-resolved images are demonstrated through a water bodies detection use-case, allowing to significantly improve detection performances on smaller water bodies.