A significant portion of publicly available municipal development plans is currently presented solely as raster graphics, derived from scanned paper development plans. This practice presents a challenge in terms of both accessibility and visualization of crucial urban planning information. For planners and citizens alike, it is imperative that both new and old development plans be made available within a unified XML-based information system. However, as of today, only new plans are machine-readable in that sense. This limitation hinders statistical analysis, impedes visualization, and restricts accessibility. In response to this issue, this study focuses on improving the representation of and accessibility of these plans, with the goal of enhancing their utility for professionals and the public. The presented approach involves the integration of deep learning-based methods, such as the pre-trained Segment Anything Model (SAM), which eliminates the need for extensive training and allows for direct application in digitization processes. For fine-grained information extraction, a specialized YOLOv8 image segmentation model is trained. This paper explores potential applications of this technology within urban planning contexts and offers insight into potential solutions for the segmentation of development plans. For example, the application of this deep learning-based approach not only enhances the visual representation of development plans but also facilitates their linkage with textual specifications, making them more informative and user-friendly. This research contributes the broader discourse on leveraging artificial intelligence to improve the utility of historical planning documents and engineering drawings, ultimately fostering more informed decision-making and enhanced public engagement.