Detection of Buried Roman Wall Remains in Ground‐penetrating Radar Data using Template Matching
Detection of Buried Roman Wall Remains in Ground‐penetrating Radar Data using Template Matching
复制标题
使用模板匹配检测探地雷达数据中埋藏的罗马城墙遗迹
DOI：
--
复制
发表时间：
2016
期刊：
影响因子：
--
通讯作者：
L. Verdonck
中科院分区：
文献类型：
--
作者：
L. Verdonck
虽然在过去的几十年中，考古探地雷达（GPR）数据的采集和处理已经变得成熟，但解释仍然具有挑战性。在三维中手动描绘是耗时的，并且通常等值面值的确定并不简单。本文提出了一种基于模板匹配的地下线性特征提取方法。首先，将三维（3D）GPR数据立方体合成为二维（2D）切片。为了实现这一点，基于足够大的时间窗口的能量切片通常可能是合适的，尽管在本研究中，与其他属性的组合，例如基于相位对称性，使得弱异常更加明显。在下一步中，我们计算复合2D切片和一些尺寸与数据集中的墙壁相似的模板的2D归一化互相关。在生成的相关矩阵中，如果每个像素的相关系数超过某个阈值，则保留最高相关系数。以这种方式，成功地映射了墙基础，但也产生了许多错误的检测。通过使用大小阈值和丢弃孤立的特征，后者在数量上大大减少。其余区域被包围在边界框中，在垂直拉伸后，边界框可用作墙结构的简化3D表示，并用于创建过滤的等值面。为了评价我们的结果，使用了手动解释。在2D的情况下（即，当比较自动映射结构的总面积与手动描绘的结构时），检测率和正确率都是~ 77%。在3D情况下（即比较体积）获得的比率略低（~71%）。我们的方法被应用到探地雷达测量的罗马别墅在肯特，英国。版权所有© 2016约翰威利父子有限公司.
Whereas in the last decades the acquisition and processing of archaeological ground‐penetrating radar (GPR) data have become mature, the interpretation is still challenging. Manual delineation in three dimensions is time consuming, and often the determination of an isosurface value is not straightforward. This paper presents a method designed specifically for the extraction of buried linear features such as wall foundations, based on template matching. First, the three‐dimensional (3D) GPR data cube is synthesized into a two‐dimensional (2D) slice. To achieve this, an energy slice based on a sufficiently large time window may often be appropriate, although in this study a combination with other attributes, for example based on phase symmetry, made weak anomalies more distinct. In the next step, we compute the 2D normalized cross‐correlation of the composite 2D slice and a number of templates with dimensions similar to the walls in the data set. Of the resulting correlation matrices, the highest correlation coefficient is kept for each pixel, if it exceeds a certain threshold. In this way, wall foundations are successfully mapped, but also many false detections are produced. The latter are greatly reduced in number by using a size threshold and discarding isolated features. The remaining regions are enclosed in bounding boxes, which after vertical extrusion can be used as a simplified 3D representation of the wall structures, and for the creation of a filtered isosurface. For the evaluation of our results, a manual interpretation was used. In the 2D case (i.e. when comparing the total area of the automatically mapped structures versus the manually delineated ones), both the detection rate and the correctness were ~77%. Slightly lower rates (~71%) were obtained in the 3D case (i.e. comparing volumes). Our method was applied to the GPR survey of a Roman villa in Kent, UK. Copyright © 2016 John Wiley & Sons, Ltd.