Istat is investigating whether Deep Learning (DL) could be used to derive automated Land Cover (LC) estimates of satisfactory quality from Sentinel-2 satellite images. This paper focuses on ‘automated land cover maps’, an important output artefact of the automatic LC estimation system we developed so far. The aim is to show how we succeeded in improving the quality of our automated LC maps by integrating a DL model for Semantic Segmentation (a U-Net) into our previous processing pipeline. The U-Net helped us solve the overestimation issue that affected our previous approach for linear narrow LC classes, like rivers and highways.