This article presents a novel methodology to the underwater documentation of pile fields in archaeological lakeside settlement sites using Structure from Motion (SfM). Mapping the piles of such sites is an indispensable basis to the exploitation of the high resolution absolute chronological data gained through dendrochronology. In a case study at the underwater site of Ploča, Mičov Grad at Lake Ohrid, North Macedonia, nine consecutive 10 m² strips and a 6 m² excavation section were uncovered, the situation documented, and the wood piles sampled. The gained data was vectorized in a geographic information system. During two field campaigns, a total of 794 wooden elements on a surface of 96 m² could be documented three-dimensionally with a residual error of less than 2 cm. The exceptionally high number of fishes in the 5 m deep water resulted in a significant covering of potentially important information on the relevant photos. We present a machine learning approach, especially developed and successfully applied to the automatic detection and masking of these fishes in order to eliminate them from the images. The discussed documentation workflow enables an efficient, cost-effective, accurate and reproducible mapping of pile fields. So far, no other method applied to the recording of pile fields has allowed for a comparably high resolution of spatial information.

本文介绍了一种使用运动结构(SfM)对考古湖边聚落遗址中的桩场进行水下记录的新方法。绘制这些遗址的成堆地图是利用通过树木年代学获得的高分辨率绝对年代数据的不可或缺的基础。在北马其顿奥赫里德湖 Mičov Grad Ploča 水下遗址的案例研究中，发现了九个连续的 10 m ²条带和一个 6 m ^{2 的}挖掘部分，记录了情况，并对木桩进行了采样。获得的数据在地理信息系统中被矢量化。在两次野外活动中，在 96 m ²的表面上总共使用了 794 个木制构件可以进行三维记录，剩余误差小于 2 cm。5 m 深水中的鱼类数量异常多，导致相关照片上潜在的重要信息被大量覆盖。我们提出了一种机器学习方法，特别是开发并成功应用于这些鱼的自动检测和掩蔽，以从图像中消除它们。所讨论的文档工作流程可以实现高效、经济、准确和可重复的桩场映射。到目前为止，还没有其他用于记录桩场的方法能够获得相对高分辨率的空间信息。