Duration: 0:00

27 views

0

Martin Olivier

This paper introduces the application of convolutional neural networks (CNNs) for feature detection in magnetostatic archaeological surveys. The study is based on a relatively small but high-resolution dataset consisting of approximately one hundred geomagnetic surveys conducted in France (as part of the Briquetage de la Seille project, Lorraine, and the National Museum of Archaeology, Saint-Germain-en-Laye) and in Britain (Department of Archaeology, University of Oxford).

Using a simple CNN architecture, the team first developed a classifier to detect archaeological features in segmented 96x96 pixel images. The model achieved an accuracy of 85% with relatively short training times, even on consumer-grade GPUs. Following this, the researchers implemented YOLO (You Only Look Once), a real-time object detection framework. Unlike the basic classifier, YOLO enabled the detection of archaeological targets across full survey images and produced bounding boxes, providing both spatial location and size data for each identified object.

With YOLO, the model reached approximately 95% accuracy in identifying object classes, though it was less precise in determining exact object positions and sizes. The paper concludes by suggesting potential improvements for the system and reflects on the broader value of deep learning approaches for object detection in remote sensing data within archaeological contexts.