Potential and Limitations of Designing a Deep Learning Model for Discovering New Archaeological Sites: A Case with the Mesopotamian Floodplain

We have tried to provide an answer to the question whether a collection of satellite images, with notable archaeological sites, is informative enough to instruct a deep learning model that discovers new archaeological sites, well before archaeologists venture out in the field. Convolutional neural networks and satellite images in the visible light range were employed to detect sites in the Iraqi region of Qadisyah. The preliminary results we achieved are interesting yet not still fully convincing. The AUC value we got is near 70%, while more interesting findings have come from the idea to map the numerical predictions into heat-maps, revealing the regions where a site can lie. Several motivations can explain this controversial output. Not least is the fact that our model was instructed to learn archaeological sites of a very different form and size