Extracting meaning from grain shape: a 3D deep learning workflow

Hèctor A. Orengo (ICREA Research Professor at ICAC) is presenting the methodology and preliminary results obtained for the identification of past agricultural strategies from grain shape, in the AEA Virtual Spring Conference (Association for Environmental Archaeology), which is taking place online on 13 May 2023.

Extracting meaning from grain shape: a 3D deep learning workflow

Orengo, H.A.; Livarda, A.; Kriti, A.; Berganzo, I.; Aliende, P.; Herranz Rodrigo, D.; Mylonas, I. & Ninou, E.

During the last years the increase in computational power driven by GPU-based processing has been paralleled by a multiplication of Machine Learning (ML) and Deep Learning (DL)-based methods. At the same time, 3D scanning has become cheaper, while considerably increasing resolution and accuracy with the wider availability of structured light-based approaches.

This presentation will showcase recent advances developed under the project ‘DarkRevisited’. After several seasons of experimental barley cultivations (21 landraces with 3 replicas under 9 different cultivation regimes), we have achieved an important reference collection of barley grains. The project has scanned thousands of these seeds at a very high resolution using a specifically developed workflow. The digital grains were then automatedly oriented, and 30 different types of measurements extracted. These served to train an ML algorithm which can identify barley varieties with a high degree of precision. However, our objectives included the identification of the environmental conditions / or agricultural management strategies during the growth of the cereals, which required the use of DL techniques directly applied to the 3D shape of the grains and specifically developed data augmentation techniques.

In this paper we will present the methods developed and preliminary results obtained for the identification of past agricultural strategies from grain shape, and the potential of these approaches for the discipline in general.

Funding:

Tags: , , ,