Day 4

Detailed paper information

Back to list

Paper title Deep Learning for Archaeological Object Detection on EO data: automating the process of buried sites' identification
  1. Arianna Traviglia Istituto Italiano di Tecnologia Speaker
  2. Marco Fiorucci IIT
  3. Paolo Soleni IIT
  4. Bertrand Le Saux ESA-ESRIN
Form of presentation Poster
  • D2. Sustainable Development
    • D2.12 Cultural and Natural Heritage
Abstract text Current straightforward access to remote sensing data for archaeological research provided by Open platforms, such as Copernicus, is putting the spotlight on the urgency of developing or advancing automated workflows able to streamline the examination of such data and unearth meaningful information from them. Automated detection of ancient human footprint on satellite imagery has seen so far limited (although promising) progress: algorithms developed to this end are usually specific for single-object categories or for a few categories, but show limited accuracy. This strongly limits their application and restricts their usability to other contexts and situations.
Advances in fine-tuning workflows for the automatic recognition of target archaeological features are being trailed within the framework of Cultural Landscapes Scanner (CLS) Project, a collaborative project involving the Italian Institute of Technology and ESA. This project tackles the shortcomings of site-specific algorithms by developing novel and more generic AI workflows based on a deep encoder/decoder neural network that exploits the availability of a large number of unlabelled EO multispectral data and addresses the lack of a priori knowledge. The methodology is based on the development of an encoder/decoder network that is pre-trained on a large set of unlabelled data. The pre-trained encoder is then connected to another decoder and the network is trained on a small, labelled dataset. Once trained, this network enables the identification of various classes of CH sites requiring only a small set of labelled data.
The experimental results on Sentinel multispectral datasets shows that this approach achieves performance close to the methods tailored for detecting only one object categories, while improves the identification accuracy in detecting different classes of CH sites. The novelty of this approach lies in the fact that it addresses the lack of both a priori knowledge and deficiencies of labelled training information, which are the prime bottlenecks that prevent the efficient use of Machine learning for the automatic identification of buried archaeological sites.