High-performance 3D morphometrics via deep learning and tabular foundation models: a case study on complex cereal grain classification

Abstract

This study explores the integration of advanced 3D morphometric techniques and machine/deep learning (ML/DL) for the analysis of complex shapes. In this study, cereal grains were employed to develop a series of complex 3D classification tasks, aiming to improve previous 2D-based classifications of barley grain origins, type, and landrace. Traditional geometric morphometric methods in archaeobotany, typically reliant on 2D data, are expanded here using high-resolution 3D models, spherical harmonics (SH) capturing complex shape variations, and different classification models. Using the Northern European Barley Dataset (NEBD), this work tests multiple ML/DL approaches, including gradient boosting machines, multilayer perceptron (MLP), MeshCNN and Tabular Foundation Models (TFM), to determine optimal classification methods across various attributes. Results indicate that SH-based coefficients combined with MLP and TFM achieved the highest classification accuracies, with over 90% accuracy in binary tasks and over 80% accuracy in multi-class landrace classification. While MeshCNN showed potential, performance was limited by computational constraints resulting in the use of lower mesh resolutions. While MLP classification of SH-based 3D shape representation achieved similar results, TFM allowed the direct use of 3D grains measures within a simple workflow. Our results demonstrate that these methods allow for significant shape analysis advances in archaeobotanical studies and beyond. This approach can enable identification and differentiation of, up to now, non-identifiable grain attributes, underscoring its potential for broad application in morphometrics.

Plant phenotyping relevance

3D穀粒形状の取得・表現と機械学習による形態分類が研究の中心であり、植物器官の形態形質を抽出・分類する方法論的研究である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.