Experts against automation? Comparing artificial intelligence and human identifications of multi-cell cereal husk phytoliths.

Abstract

This study compares the efficacy of a deep learning-based computer vision model against manual expert identification of multi-cell husk phytoliths from wheat (Triticum boeticum/dicoccoides), barley (Hordeum spontaneum), and oats (Avena sativa). Conducted via an online survey in 2024 with twelve respondents (92% completion rate, one respondent did not finish), the survey presented 18 phytolith images selected based on: 1) deep learning performance, 2) manual diagnostic clarity, and 3) random control. The deep learning model achieved 100% classification accuracy, while manual experts averaged 44%, with Avena proving the most challenging (26.39% accuracy) compared to Hordeum (54.17%) and Triticum (48.61%). Statistical analysis confirmed significant accuracy variations (p = 0.0016), highlighting Avena’s genus identification difficulty. Experience level influenced performance, with less than five years post-PhD researchers scoring highest (72%), though completion time did not correlate with accuracy. The algorithm’s superior performance, employing and appropriately weighting features like wave pattern and papillae, underscores the potential of artificial intelligence to transform paleoethnobotany by generating reliable, large-scale datasets. This research advocates for the integration of machine learning tools within manual studies in the short-term, to enhance accuracy and efficiency in phytolith analysis, paving the way for broader applications in environmental archaeology in the long-term