Many researchers have been interested in historical and geographical analysis over the years. Particularly in the field of historical research, the digitization of historical documents has become increasingly important, which facilitates the ability to compare and process a large number of historical documents. One of the most challenging historical data encountered in digitization are cadastral maps.
A cadastral map is a map that describes the boundaries and ownership of land parcels. It has the benefit of presenting the spatial relationships between features illustrated on the map such as location and buildings shape in addition to ownership and parcel values. Usually the digitization can be partially solved by archiving maps as images; however, it adds some new concerns.
Raster images are composed of a grid of pixels, which cannot be scaled without loss of resolution. Moreover, the grid representation of the raster images can produce inaccuracies in spatial and spectral information. Since raster images consist of millions of pixels, processing, altering, and transferring the images require high computational power. Therefore, vectorizing cadastral maps is the key for solving these issues. Vectorization means that the maps are redrawn in a geometrical form, i.e. in polygons and lines with points. The main advantage of vector graphic data is that it is lightweight, scalable for arbitrary size, and can be analyzed interactively. This also known as a raster-to-vector conversion.
In this presentation, a pipeline is introduced to automate the raster-to-vector process of historical cadastral map images, which consists of transforming raster historical cadastral map into vector graphic format. The pipeline initiates a solution to solve this problem using a combination of modern and traditional Machine Learning techniques. It has been tested on different cadastral maps and it achieves a modular, extensible application with satisfactory results. The results indicate that the pipeline geometrically approximates 79% of the parcels as polygons with an 𝐼𝑜𝑈 score higher than 0.8 and classifies different map objects with an accuracy of 90% using Feature Engineering. Finally, a web application with a user-friendly interface is implemented, which provides interactive visualization and a simple control panel for the parameters used in the pipeline.
digitalhistory: Benjamin Risse/Sufian Zaabalawi: Automatische Vektorisierung historischer Katasterkarten, in: Digital History Berlin (Blog), erschienen am: 4. Januar 2021, Internet: https://dhistory.hypotheses.org/346.