Under Pressure: Predicting pressure on micro CT-scans of archaeological soil samples using convolutional neural networks

Abstract

In order to protect buried archaeological remains from the pressure of build sites, it is important to assess how this pressure affects the remains. To aid this assessment we answer the question: Can we use convolutional neural networks (CNN) to predict the pressure that was applied to archaeological soil samples scanned by a micro CT-scanner? The dataset used in this study was created by repeatedly scanning a single, artifact rich soil sample. With each scan, an increasing amount of pressure is applied to the sample, damaging the artifacts. The soil sample was scanned by a micro CT-scanner. The amount pressure applied serves as the label, in the machine learning process while the images (slices from the 3D scan) from the various samples serve as input. A convolutional neural network making use of transfer learning, tries to predict the pressure belonging to the images when the image is fed as input. The test results on 261 unseen images after training the model show a 99.61% correct prediction rate. The results are very promising, but since the model was trained on a very specific dataset, they are not representative for a more general prediction of pressure applied to archaeological soil samples.