Detecting graves in GPR data: assessing the viability of machine learning for the interpretation of graves in B-scan data using medieval Irish case studies.

Authors: Green, A.

Conference: Bournemouth University, Faculty of Science and Technology

Abstract:

As commercial archaeogeophysical survey progressively shifts towards large landscape-scale surveys, small features like graves become more difficult to identify and interpret. In order to increase the rate and confidence of grave identification before excavation using geophysical methods, the accuracy and speed of survey outputs and reporting must be improved. The approach taken in this research was first to consider the survey parameters that govern the effectiveness of the four conventional techniques used in commercial archaeogeophysical evaluations (magnetometry, earth resistance, electromagnetic induction and ground-penetrating radar). Subsequently, in respect of ground-penetrating radar (GPR), this research developed machine learning applications to improve the speed and confidence of detecting inhumation graves. The survey parameters research combined established survey guidelines for the UK, Ireland, and Europe to account for local geology, soils and land cover to provide survey guidance for individual sites via a decision-based application linked to GIS database. To develop two machine learning tools for localising and probability scoring grave-like responses in GPR data, convolutional neural networks and transfer learning were used to analyse radargrams of medieval graves and timeslices of modern proxy clandestine graves. Models were c. 93% accurate at labelling images as containing a grave or no grave and c. 96% accurate in labelling and locating potential graves in radargram images. For timeslices, machine learning models achieved 94% classification accuracy. The >90% accuracy of the machine learning models demonstrates the viability of machine-assisted detection of inhumation graves within GPR data. While the expansion of the training dataset would further improve the accuracy of the proposed methods, the current machine-led interpretation methods provide valuable assistance for human-led interpretation until more data becomes available. The survey guidance tool and the two machine learning applications have been packaged into the Reilig web application toolset, which is freely available.

https://eprints.bournemouth.ac.uk/34284/

Source: Manual

Detecting graves in GPR data: assessing the viability of machine learning for the interpretation of graves in B-scan data using medieval Irish case studies.

Authors: Green, A.

Conference: Bournemouth University

Abstract:

As commercial archaeogeophysical survey progressively shifts towards large landscape-scale surveys, small features like graves become more difficult to identify and interpret. In order to increase the rate and confidence of grave identification before excavation using geophysical methods, the accuracy and speed of survey outputs and reporting must be improved. The approach taken in this research was first to consider the survey parameters that govern the effectiveness of the four conventional techniques used in commercial archaeogeophysical evaluations (magnetometry, earth resistance, electromagnetic induction and ground-penetrating radar). Subsequently, in respect of ground-penetrating radar (GPR), this research developed machine learning applications to improve the speed and confidence of detecting inhumation graves. The survey parameters research combined established survey guidelines for the UK, Ireland, and Europe to account for local geology, soils and land cover to provide survey guidance for individual sites via a decision-based application linked to GIS database. To develop two machine learning tools for localising and probability scoring grave-like responses in GPR data, convolutional neural networks and transfer learning were used to analyse radargrams of medieval graves and timeslices of modern proxy clandestine graves. Models were c. 93% accurate at labelling images as containing a grave or no grave and c. 96% accurate in labelling and locating potential graves in radargram images. For timeslices, machine learning models achieved 94% classification accuracy. The >90% accuracy of the machine learning models demonstrates the viability of machine-assisted detection of inhumation graves within GPR data. While the expansion of the training dataset would further improve the accuracy of the proposed methods, the current machine-led interpretation methods provide valuable assistance for human-led interpretation until more data becomes available. The survey guidance tool and the two machine learning applications have been packaged into the Reilig web application toolset, which is freely available.

https://eprints.bournemouth.ac.uk/34284/

Source: BURO EPrints