 |
Dura Europos is an ancient city located in Syria. Notable artifacts have been unearthed, including numerous temples, wall decorations, inscriptions, military equipment, tombs, numerous papyri. Especially after the Syrian civil war in 2011, the site has been heavily looted. |
 |
Ebla is an ancient city in Syria of the Early Bronze Age, rebuilt twice, and ultimately destroyed in the middle of the 2nd millennium BC. Its remains constitute a tell, known as Tell Mardikh, located near the modern-day Mardikh. After the Syrian civil war, it was looted. |
 |
Aswan is one of the richer archaeological areas of Southern Egypt, and it acts as an interface between the Nile Valley and both the Western and Eastern deserts. It is the setting of a large number of necropolises, temples and sites. During the last years, the looting has increased and shows no sign of decreasing. |
 |
Arpinova is a rural village in south Italy, surrounded by highly significant archaeological sites. Among these, the most extensive and one of the oldest in Europe is the Neolithic village of Passo di Corvo. Additionally, the area contains the remains of the ancient Daunian city of Arpi, including structures and various necropolises. The area is well known to Law Enforcement Agencies for being a site of illegal excavations, especially in the past years. |
 |
Cerveteri is a town in central Italy. Caere was one of the most important and populous Etruscan city-states. Little is known of the ancient city. Portions of the city's walls remain visible to this day. Three necropolises have been discovered, and are between the most significant due to the magnificence of their tombs, as well as the significance of the excavated artifacts. The area is well known to Law Enforcement Agencies for being a site of clandestine excavations. |
|
|
|
|
Selected multitemporal satellite images have been chosen for each site.
A geodatabase of the looted traces already identified and mapped has been created to train the Deep Neural Network developed.
The metadata used to label pit-polygons in the images have been carefully defined. Each feature class contains information about the singles looting pits, their shape/type and their reliability.
The developed Machine Learning Model is a Supervised Deep Neural Network for Change Detection on Satellite Imagery. This model analyses images of the same scene at two different time steps and automatically segments the newly appeared looting pits.
Most of the training dataset is used for optimising the Deep Neural Network performance by learning from its prediction errors. While the rest is used to estimate how the model performance generalizes on “new data”.
A ground truthing campaign has been carried out in Aswan (Egypt), where 129 features of the curated geodatabase were verified. The change detection model has been re-trained removing the misidentified looting pits.
Risk Maps have been elaborated and created for each case study, relying on the density and distance of the looting pits, and the proximity to known archaeological areas.