Twenty thousand leagues under plant biominerals: a deep learning implementation for automatic phytolith classification

Abstract

Abstract Phytoliths constitute microscopic SiO 2 -rich biominerals formed in the cellular system of many living plants and are often preserved in soils, sediments and artefacts. Their analysis contributes significantly to the identification and study of botanical remains in (paleo)ecological and archaeological contexts. Traditional identification and classification of phytoliths rely on human experience, and as such, an emerging challenge is to automatically classify them to enhance data homogeneity among researchers worldwide and facilitate reliable comparisons. In the present study, a deep artificial neural network (NN) is implemented under the objective to detect and classify phytoliths, extracted from modern wheat ( Triticum spp.). The proposed methodology is able to recognise four phytolith morphotypes: (a) Stoma, (b) Rondel, (c) Papillate, and (d) Elongate dendritic. For the learning process, a dataset of phytolith photomicrographs was created and allocated to training, validation and testing data groups. Due to the limited size and low diversity of the dataset, an end-to-end encoder-decoder NN architecture is proposed, based on a pre-trained MobileNetV2, utilised for the encoder part and U-net, used for the segmentation stage. After the parameterisation, training and fine-tuning of the proposed architecture, it is capable to classify and localise the four classes of phytoliths in unknown images with high unbiased accuracy, exceeding 90%. The proposed methodology and corresponding dataset are quite promising for building up the capacity of phytolith classification within unfamiliar (geo)archaeological datasets, demonstrating remarkable potential towards automatic phytolith analysis.

Plant phenotyping relevance

植物由来の植物珪酸体を画像から検出・分類する深層学習手法とデータセットの開発が研究の中心であり、植物形態情報の取得・抽出に該当する。

Code and data availability

The paper's phytolith photomicrograph dataset (annotated images of four morphotypes from modern wheat) is explicitly stated to be publicly available on Kaggle. Code and trained NNs are mentioned as contributions but no public repository URL is provided, so only the dataset qualifies.