Sipilä, Ilkka Matti Veikko; (2022) Addressing subjectivity in the classification of palaeoenvironmental remains with supervised deep learning convolutional neural networks. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Thesis (with corrections) - Addressing subjectivity in the classification of palaeoenvironmental remains with supervised deep learning convolutional neural networks.pdf - Accepted Version Download (37MB) | Preview

Abstract

Archaeological object identifications have been traditionally undertaken through a comparative methodology where each artefact is identified through a subjective, interpretative act by a professional. Regarding palaeoenvironmental remains, this comparative methodology is given boundaries by using reference materials and codified sets of rules, but subjectivity is nevertheless present. The problem with this traditional archaeological methodology is that higher level of subjectivity in the identification of artefacts leads to inaccuracies, which then increases the potential for Type I and Type II errors in the testing of hypotheses. Reducing the subjectivity of archaeological identifications would improve the statistical power of archaeological analyses, which would subsequently lead to more impactful research. In this thesis, it is shown that the level of subjectivity in palaeoenvironmental research can be reduced by applying deep learning convolutional neural networks within an image recognition framework. The primary aim of the presented research is therefore to further the on-going paradigm shift in archaeology towards model-based object identifications, particularly within the realm of palaeoenvironmental remains. Although this thesis focuses on the identification of pollen grains and animal bones, with the latter being restricted to the astragalus of sheep and goats, there are wider implications for archaeology as these methods can easily be extended beyond pollen and animal remains. The previously published POLEN23E dataset is used as the pilot study of applying deep learning in pollen grain classification. In contrast, an image dataset of modern bones was compiled for the classification of sheep and goat astragali due to a complete lack of available bone image datasets and a double blind study with inexperienced and experienced zooarchaeologists was performed to have a benchmark to which image recognition models can be compared. In both classification tasks, the presented models outperform all previous formal modelling methods and only the best human analysts match the performance of the deep learning model in the sheep and goat astragalus separation task. Throughout the thesis, there is a specific focus on increasing trust in the models through the visualization of the models’ decision making and avenues of improvements to Grad-CAM are explored. This thesis makes an explicit case for the phasing out of the comparative methods in favour of a formal modelling framework within archaeology, especially in palaeoenvironmental object identification.

Download activity - last month

Export as JSONXMLCSV

Download activity - last 12 months

Export as XMLJSONCSV

Downloads by country - last 12 months

Export as CSVJSONXML

Archive Staff Only

View Item