Imaging the past using methods of the future: computer-aided interpretation of ground-penetrating radar data collected at Roman towns in Italy

由于其广泛性，地球物理勘探非常适合研究罗马城市的整体特征和发展。虽然目前我们的理解依赖于一些大量挖掘的遗址（例如庞贝，奥斯蒂亚），但勘探可以帮助回答城市规划，街道网络，居住区规模，人口估计以及城市边界与城市发展之间的相互作用等问题。然而，一个关键的瓶颈是传统的地球物理数据解释（人工圈定异常）耗时的特点。最能说明这一点的是探地雷达（GPR）技术，它可以揭示罗马城镇的详细平面图，但由于其高分辨率和3d特性，它产生了大量的数据。如今，深度卷积神经网络（DCNNs）等强大的机器学习技术在勘探数据计算机辅助解释（CAI）方面的潜力日益得到证明。然而，DCNNs需要大的、人工描述的训练集。构建这些是一项努力，部分地抵消了CAI的好处。在这个项目中，我将研究从少量人工描绘中训练DCNNs的新方法（同时提供有竞争力的结果），并将其应用于意大利罗马城镇的GPR数据。此外，我将把这些和其他CAI算法结合在一个用户友好的工具箱中，可以在考古社区中传播。最后，在CAI结果的基础上，我将使用GIS、数据融合和空间语法对来自Falerii Novi（意大利中部）的GPR数据集进行考古解释。我将在一家商业软件开发商借调期间接受罗马城市主义和机器学习方面的培训，以及计算机技能方面的培训。通过非侵入性的调查，这个跨学科的项目将加强对罗马城市主义的理解，并促进这些重要文化遗产的保护。

Because of its extensive character, geophysical prospection is well-suited to investigate the overall character and development of Roman cities. Whereas currently our understanding relies on a few largely excavated sites (e.g. Pompeii, Ostia), prospection can help answering questions on city planning, street network, the size of inhabited areas, population estimations, and the interaction between city boundaries and urban development. However, a key bottleneck is the time consuming character of the traditional interpretation of geophysical data (manual delineation of anomalies). This is best illustrated by the ground-penetrating radar (GPR) technique, which reveals detailed plans of Roman towns, but produces enormous amounts of data because of its high-resolution and 3-D nature. Today, the potential of powerful machine learning techniques such as deep convolutional neural networks (DCNNs) for the computer-aided interpretation (CAI) of prospection data is increasingly demonstrated. However, DCNNs need large, manually delineated training sets. Building these is an effort that partially undoes the benefit of CAI. In this project, I will investigate novel approaches that train DCNNs from a small amount of manual delineations (while delivering competitive results), and apply them to GPR data from Roman towns in Italy. Furthermore, I will combine these and other CAI algorithms in a user-friendly toolbox that can be disseminated among the archaeological community. Finally, on the basis of the CAI results, I will provide an archaeological interpretation of the GPR dataset from Falerii Novi (Central-Italy) using GIS, data fusion and space syntax. I will be trained in Roman urbanism and machine learning, and in computing skills during a secondment at a commercial software developer. By stimulating non-invasive investigation, this interdisciplinary project will enhance the understanding of Roman urbanism, and promote the preservation of these important cultural heritage sites.