In the perspective of landscape archaeology, remote sensing is a very important tool that allows to recognize and locate potential sites, which will then be "groundtruthed"through a surface survey. Remote sensing is, unfortunately, a very time-consuming process that scales terribly with the size of the area under investigation. In this paper we explore the possibility of using semantic segmentation models to detect and highlight the presence of archaeological sites present in the Mesopotamian floodplain. Whereas archaeologists usually combine information from a variety of basemaps, including aerial and satellite photos taken from the 1950s onwards, we investigated the possibility of using an easily accessible online maps (in our case, Bing Maps). Trying to build an accessible and lightweight system also dictated the choice of trying pretrained segmentation models and use transfer learning. The preliminary results obtained (from different models and parameters choices), as well as the dataset, its idiosyncrasies and how we can deal with them are discussed in this paper.
Casini L., Orrù V., Roccetti M., Marchetti N. (2022). When Machines Find Sites for the Archaeologists: A Preliminary Study with Semantic Segmentation applied on Satellite Imagery of the Mesopotamian Floodplain. Association for Computing Machinery [10.1145/3524458.3547121].
When Machines Find Sites for the Archaeologists: A Preliminary Study with Semantic Segmentation applied on Satellite Imagery of the Mesopotamian Floodplain
Casini L.;Orrù V.;Roccetti M.
;Marchetti N.
2022
Abstract
In the perspective of landscape archaeology, remote sensing is a very important tool that allows to recognize and locate potential sites, which will then be "groundtruthed"through a surface survey. Remote sensing is, unfortunately, a very time-consuming process that scales terribly with the size of the area under investigation. In this paper we explore the possibility of using semantic segmentation models to detect and highlight the presence of archaeological sites present in the Mesopotamian floodplain. Whereas archaeologists usually combine information from a variety of basemaps, including aerial and satellite photos taken from the 1950s onwards, we investigated the possibility of using an easily accessible online maps (in our case, Bing Maps). Trying to build an accessible and lightweight system also dictated the choice of trying pretrained segmentation models and use transfer learning. The preliminary results obtained (from different models and parameters choices), as well as the dataset, its idiosyncrasies and how we can deal with them are discussed in this paper.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.