A Micromorphological Neural Network

We are currently beginning to develop a machine learning system based on a Convolutional Neural Network (CNN) to automate the description of micromorphological thin sections. This work is carried out in collaboration with Professor Rafael Arnay from the Department of Computer Engineering at the University of La Laguna (ULL). The project is funded through a research grant led by Professor Arnay: “Neural Network Image Tagging-Based System for Micromorphological Description in Archaeology.”

Although neural network-based image recognition systems have become increasingly widespread in recent years, their applications in microstratigraphy and micromorphology remain limited. Currently, micromorphological description is a manual process that requires substantial time and involves a degree of subjectivity. Our aim is to develop a tool that can accelerate and standardize this process.

The core idea is to train a CNN model using a micromorphology image database. We are working with high-resolution scans of thin sections, which are processed using a sliding window technique. This allows us to isolate and classify individual micromorphological features such as particle types, porosity, and microstructures. In a second phase, we will develop a web-based application for image tagging, enabling external users to contribute to the database—ultimately increasing the CNN’s learning capacity and autonomy.

Future developments will focus on the automated recognition of further archaeological materials such as malacofauna and ceramics, as well as incorporaing these methodologies into the micromorphological workflow.

Funding

Publications

Arnay, R., García-Villa, P., Hernández-Aceituno, J., Rueda-Saiz, S., & Mallol, C. (2026). Improving micromorphological analysis with CNN-based segmentation of flint/obsidian, bone and charcoal. PLOS One, 21(1), e0340353. https://doi.org/10.1371/journal.pone.0340353

Arnay, R., Hernández-Aceituno, J., & Mallol, C. (2021). Soil micromorphological image classification using deep learning: The porosity parameter. Applied Soft Computing, 102, 107093. https://doi.org/10.1016/j.asoc.2021.107093