Recreating ancient maps requires a rich understanding of history. Those cartographic documents are imbued with the intentions of their creators as well as their historical and sociopolitical contexts. The need for a high level of understanding of these documents has oftentimes proven to be a barrier for many social scientists conducting cross-disciplinary research.
A team of researchers is trying to break through this barrier with machine learning, using conditional generative adversarial networks (GANs) to recreate historical maps in the modern satellite style. The work may expedite research and provide a better understanding of the changes that our physical environment has gone through.
“Historical maps are an irreplaceable primary source of geographical and political information from the past,” explain researchers Henrique Andrade and Bruno Fernandes in their paper on the project.
The technique involves an image-to-image translation software called Pix2Pix, but it adds an intermediate step that incorporates an understanding of the supporting semantics involved in the translation.
Pix2Pix is a conditional GAN consisting of a generative network and a discriminator network. The GAN learns how to translate the maps into images, while the discriminator determines whether the images are real or computer generated. The framework is a supervised method, and the inputs are provided in pairs of the expected satellite-like image and its label map (see the second-to-last step in the diagram below).
The work involved in translating these historical cartographical documents requires the creation of sets of label maps from the inputs (such as map legends and other contextual indices), the translation of those label maps, and the combination of those results. The product is a synthesized satellite image that combines both sets of inputs. This allows those studying the documents to get a better understanding of what is depicted on the map, with all the contextual information already filled in.
While the satellite-style results are not instant, and the discriminator network must be trained over time in order to produce more accurate results, the method allows for faster image generation. The only bottleneck is the requirement for a human researcher to confirm the accuracy of the data. As the discriminator network for a particular task grows, the need for human interaction is lessened over time.
This use of computer science to help solve issues within other domains of inquiry (history, epidemiology, sociology, civil engineering) is a great example of engineering at work. The reconstruction of these documents should prove helpful in providing a historical understanding of the way the world we live in has changed over time.