A methodology for the self‐training and self‐assessing of new GPR practitioners

Measuring diagnostic proficiency illustrated by a case study of a historic African‐American cemetery for unmarked graves

Bibliographic Data

ID12429591
AuthorsJ M Martin (0000-0001-5585-2356, Department of Geology and Geophysics Texas A&M University College Station Texas USA, corresponding author), Mark E Everett (0000-0001-6453-0316, Department of Geology and Geophysics Texas A&M University College Station Texas USA)
Year2023
Volume30
Issue3
Pages311-325
Publication date2023-01-13
Peer ReviewedYes
Open AccessYes
TypeARTICLE
VenueArchaeological Prospection (JOURNAL)
Journal identifiersISSN: 1075-2196 • E-ISSN: 1099-0763
PublisherWiley (PUBLISHER • GB)
DOI10.1002/arp.1893
OpenAlexW4316011122
LanguageEN
Citations received7
References cited36

In ground penetrating radar (GPR) surveys of African‐American cemeteries suspected to contain unmarked graves, determining the proficiency of a new GPR practitioner is vital and perhaps even more fundamental than that of the GPR hardware, deployment configuration and software. Proficiency may be defined as the practitioner's true‐positive, true‐negative, false‐positive (i.e., false alarms) and false‐negative (i.e., misses) percentages. We embarked on this research as a means to improve our own proficiency in unmarked grave detection and to develop an algorithm by which to improve the proficiency of any new GPR practitioner. After surveying the Salem Cemetery in Brazos County, TX, and the Old Danville‐Shepherd Hill Cemetery in Montgomery County, TX, we first classified subsurface targets based on in‐field, visual inspection of the real‐time, onscreen, unprocessed (except for an automatic gain control) GPR B‐scans. We then developed a proxy for ground‐truthing that allowed us to calculate the proficiency of the in‐field classifications. From this proxy, we established a quantitative prevalence threshold for identifying or rejecting a subsurface object as a target of interest. Its quantitative nature allowed us to quantitatively control and adjust that threshold, a threshold we set at 70% likely to be to a specific target of interest. We show that our classification accuracy increased from 66.2% at the Salem Cemetery to 75.0% at the Old Danville‐Shepherd Hill Cemetery and, through use of diagnostic evaluations originally developed for medical imaging and herein applied to geophysics, showed that the accuracy improved due to increases in true‐negative classifications, that is, in examining real‐time, onscreen, mostly unprocessed GPR B‐scans, discerning a potential target and correctly concluding it was not an unmarked grave. This research outlines the procedure we developed to measure the proficiency of a new GPR practitioner

Archaeology · Ground-Penetrating Radar · Machine learning · Proxy (statistics · Radar · Computer Science · Geophysical and Geoelectrical Methods · Geophysical Methods and Applications · History · Seismic Imaging and Inversion Techniques

Tools

Open DOIOpen Access