Academia.eduAcademia.edu

Figure 7 – uploaded by Juergen Landauer

See full PDFdownloadDownload figure
Figure 8 shows the distribution of resulting confidence scores as seen in the histogram function of QGIS (from left: low to right: high). A large number of unconfirmed hill- forts have scores close to zero, followed by relatively few in the 0.4-0.7 range. Surpris- ingly, a distinct number of hillforts show scores of 0.8 or above (marked by the black box). With some justification, one can claim that the latter ‘have features very similar’ to

Figure 8 shows the distribution of resulting confidence scores as seen in the histogram function of QGIS (from left: low to right: high). A large number of unconfirmed hill- forts have scores close to zero, followed by relatively few in the 0.4-0.7 range. Surpris- ingly, a distinct number of hillforts show scores of 0.8 or above (marked by the black box). With some justification, one can claim that the latter ‘have features very similar’ to

Related Figures (6)

For Britain and Ireland, Lock and Ralston (2017) published an excellent online data- base with the location of more than 17,000 Bronze and Iron Age hillforts (fig. 1). More- over, this database contains detailed object information in an easily retrievable format, including condition (e.g., ‘destroyed’), type (e.g., promontory fort), reliability of archae- ological interpretation (‘confirmed’ or ‘un- confirmed’) and topography class (such as hilltop or lowland). This information proved valuable in the construction of a suitable da- taset for the CNN training (see Section 3).
For Britain and Ireland, Lock and Ralston (2017) published an excellent online data- base with the location of more than 17,000 Bronze and Iron Age hillforts (fig. 1). More- over, this database contains detailed object information in an easily retrievable format, including condition (e.g., ‘destroyed’), type (e.g., promontory fort), reliability of archae- ological interpretation (‘confirmed’ or ‘un- confirmed’) and topography class (such as hilltop or lowland). This information proved valuable in the construction of a suitable da- taset for the CNN training (see Section 3).
LiDAR data is widely available across Eng- land. In fact, in 2020, more than 93% of the land surface was covered (see fig. 2). The Department for Environment, Food & Rural Affairs (DEFRA) offers a LiDAR Composite Digital Terrain Model (DTM) web service. This contains LiDAR data for England in various data formats and resolutions, return- ing a hillshade-like representation. For the case study, LiDAR data with a 2m resolution was downloaded from this portal. Figure 2: Overview of LiDAR coverage for England (source: DEFRA, https://environment.data.gov.uk)
LiDAR data is widely available across Eng- land. In fact, in 2020, more than 93% of the land surface was covered (see fig. 2). The Department for Environment, Food & Rural Affairs (DEFRA) offers a LiDAR Composite Digital Terrain Model (DTM) web service. This contains LiDAR data for England in various data formats and resolutions, return- ing a hillshade-like representation. For the case study, LiDAR data with a 2m resolution was downloaded from this portal. Figure 2: Overview of LiDAR coverage for England (source: DEFRA, https://environment.data.gov.uk)
Figure 3: Excerpts of LiDAR data, down- loaded from the DEFRA web service, show- Se: map nanwew lan At ke llfhAwéa
Figure 3: Excerpts of LiDAR data, down- loaded from the DEFRA web service, show- Se: map nanwew lan At ke llfhAwéa
Figure 5: Three known hillforts (black dots) and six squares containing hillforts at > 80% confidence (some squares overlapping, source background image: DEFRA) After testing, these output files were im- ported into QGIS as so-called Delimited Text Layers for visual inspection. Figure 5 pro- vides an example of the visualisation of part of the Ingram Valley test area. Black dots represent known hillforts, while the orange squares represent (possibly overlapping) patches. The confidence score of every patch is displayed as text.
Figure 5: Three known hillforts (black dots) and six squares containing hillforts at > 80% confidence (some squares overlapping, source background image: DEFRA) After testing, these output files were im- ported into QGIS as so-called Delimited Text Layers for visual inspection. Figure 5 pro- vides an example of the visualisation of part of the Ingram Valley test area. Black dots represent known hillforts, while the orange squares represent (possibly overlapping) patches. The confidence score of every patch is displayed as text.
For this reason, and since even these remote verifications were relatively time-consum- ing, we decided to focus on the more prom- ising candidates. That is, concentrate on the potential hillforts with the highest confidence scores. (Verschoof-van der Vaart 2020, Verschoof-van der Vaart & Landauer 2021) Empirically, we found a confidence thresh- old of around 0.8 (resulting in ca.100 poten- ial hillforts) to be a good compromise be- ween the objective of finding ‘all’ unknown objects and the manual verification effort needed.
For this reason, and since even these remote verifications were relatively time-consum- ing, we decided to focus on the more prom- ising candidates. That is, concentrate on the potential hillforts with the highest confidence scores. (Verschoof-van der Vaart 2020, Verschoof-van der Vaart & Landauer 2021) Empirically, we found a confidence thresh- old of around 0.8 (resulting in ca.100 poten- ial hillforts) to be a good compromise be- ween the objective of finding ‘all’ unknown objects and the manual verification effort needed.
As an additional test, patches containing 145 hillforts, which were marked as ‘uncon- firmed’ or ‘in doubt’ (and were therefore ex- cluded from the training dataset) in the Atlas (Lock and Ralston 2017), were put through the trained CNN.
As an additional test, patches containing 145 hillforts, which were marked as ‘uncon- firmed’ or ‘in doubt’ (and were therefore ex- cluded from the training dataset) in the Atlas (Lock and Ralston 2017), were put through the trained CNN.
Related topics:
Landscape ArchaeologyHillforts and EnclosuresAutomated Detection
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved