Reading the Past from the Surface of the Earth

Human interventions into the environment leave traces: Depending on the construction materials that were used, these traces disappear sooner or later. Whereas remains of Roman roads are still to be found in many places, we know relatively little about some important trading arteries of the 18th century. During the past 250 years, significant data have accumulated documenting human interventions into the surface of the past thus construction activities such as roads and houses, yet also agriculture, like forestry and field cultivation for some regions of the world. For the territory of the Habsburg Empire for example, so called cadastral maps are available starting with the very early 19th century. These provide us with detailed information on the location of houses and roads back then, yet also on the crops and vegetables farmed back then on which fields. Many of the buildings that show up in these old maps still exist, and frequently, the course of roads has not changed significantly over decades and sometimes even centuries! Owing to high-resolution satellite imagery, everybody can study the surface of the earth today, and have a closer look at urban sites, important highways, or untouched nature. If one lays the old cadastral maps on top of modern satellite imagery, one can easily identify those spots that have remained almost untouched since the date of the cadastral map survey, and those, where landscapes have transformed significantly. We use satellite imagery and an artificial intelligence (AI)-based computer system to analyse, where in Austria buildings, roads and agricultural areas have disappeared. To that end, we put cadastral maps and satellite images on top of each other and train our AI-system to mark those spots, where human activities were once visible and are not anymore. Then, our AI-system must compare those marked spots, where there used to be streets, buildings or gardens years, decades or centuries ago, with the surrounding areas that have never seen human activity. We aim at the identification of the slightest traces human activities can possibly leave in the surface of the earth, to use these for the further detection of archeological sites. We are searching for an extremely wide range of different signals that we expect to be so subtle and at the same time complicated that only extremely powerful computers will be able to recognize them. The aim of our project is to train an artificial intelligence, which can identify certain signals and combinations of signals in satellite photography (and further aerial photography), that are most likely a hint for historical human activities in a certain place. These signals can consist in slightly changed colors in the soil, tiny differences in nivels, certain structures, etc. Archeologists and researchers will use our system as a new tool to uncover the remains of human settlements and activities in satellite imagery.

In the course of major construction projects - from tramway extensions in the city centre of Graz to the construction of the Koralm railway - one repeatedly comes across structural remains from the past, from old city gates, Roman roads to residential buildings. Such discoveries cause delays in construction plans, because they require the intervention of archaeologists who uncover and document the discovered structures in emergency excavations. Our knowledge of our material heritage is surprisingly limited and only very good in a few places - such as medieval city centres or former Roman settlements. In the RePaSE project - Reading the Past from the Surface of the Earth - a multidisciplinary research group of historians and data scientists dedicated themselves to our past, which has sunk into the earth. The aim was to find out whether scientists can teach artificial intelligence to detect traces of buildings from the past in aerial and satellite images. The basic idea behind this project was that often in satellite images, digital terrain models, high-resolution aerial photographs and other image data sets that are freely accessible today, building structures can be recognised that no longer exist (sometimes for centuries) but have left traces in the earth. The problem with this is that it can depend on when a satellite or aerial photograph was taken whether anything is still visible, in agriculture it can make a difference what crop was grown in a field. It may be that a Roman road is visible in a certain satellite image but not in another satellite or aerial image or in the digital terrain model, or vice versa. For Styria alone, there are so many and such comprehensive image data that archaeologists could never evaluate them. Artificial intelligence can process a lot of data quickly, but it has to be very carefully prepared and trained for its task. In order to teach it to "read" the traces of our material past in this image material, the RePaSE team, consisting of two historians from the University of Graz and three data scientists from the Know Center at TU Graz, trained several different AI models for different tasks: First, a Deep Learning model was trained to recognise houses in historical cadastral maps from the 1820s. Then another model was used to detect these 1820s houses or their remains in aerial and satellite images. Through RePaSE, it was found that the idea behind it works very well. Soon, a demonstrator of RePaSE will go online, allowing Styrians to try out the detector. The next step will be to start a follow-up project that will extend the technology behind it from buildings to roads and from Styria to the whole of Austria.