Deep Learning for the Segmentation of Large-Scale Surveys of Historic Masonry: A New Tool for Building Archaeology Applied at the Basilica of St Anthony in Padua

Abstract

In the last decade, the documentation of historical buildings has made tremendous progress in generalising the use of high-precision laser scanning and drone photogrammetry. Yet the potential of digital surveying is not fully exploited due to difficulties in manually analysing large amounts of collected data. Machine learning offers immense potential as a game-changer in building archaeology, especially for the documentation of structures composed of millions of units. This paper presents the first segmentation of large-scale surveys of historic masonry using machine learning, using the thirteenth-century Basilica of St Anthony (Padua, Italy) as a case study. Based on a drone survey of the north façade of the building (110 × 70 m), a state-of-the-art non-learning segmentation approach is described and its limitations for historical structures are illustrated. Then, a new workflow based on convolutional neural networks (CNN) is presented. The result is a precise mapping of about 300,000 individual bricks showing a large variety of formats and bonds. The automatic surveys are analysed using visual programming language (VPL), enabling a rapid and feature-based identification of building phases and repair interventions. The outcome demonstrates the validity of machine learning for the analysis of historical structures and its potential in the field of heritage.