Recent advances in Artificial Intelligence and Computer Vision have been showed to be promising for automated land use classification of remotely sensed data. However, current state-of-the-art per-pixel segmentation networks fail to accurately capture geometrical and topological properties on land use segmentation, as these methods have inherently a lot of freedom. These geometrical and topological properties of land use structures are crucial for describing land usage on topographical maps, as their purpose is to present insight into topology and borders of land use structures. In order to preserve the geometrical and topological properties of land use structures, a novel segmentation method is introduced and tested on road structures in this thesis. Unlike current state-of-the-art segmentation networks, this new method performs the segmentation task by utilizing shape regression techniques as currently applied by state-of-the-art object detection networks. As modern object detection methods are only able to perform regression on simplistic shapes, and road structures generally describe complex shapes, a new topology preserving annotation generation method is introduced that subdivides a complex road structure into a set of oriented rectangular shapes. Since not many publicly available land use datasets contain both aerial images and per-pixel annotations, a new dataset based on aerial images and land use annotations, covering large areas of the Netherlands, is introduced as well. The results show that the novel segmentation method is capable of learning the newly introduced road structure representation, which preserves geometrical and topological properties. The connectedness property, however, is lost. The novel method does currently not outperform current state-of-the-art per-pixel segmentation networks, although several directions for future work are proposed to improve the segmentation performance of the shape regression based technique and preserve the connectedness property.