Extracting Ancient Maya Structures from Aerial LiDAR Data using Deep Learning
Extracting Ancient Maya Structures from Aerial LiDAR Data using Deep Learning
复制标题
DOI：
10.1109/southeastcon51012.2023.10115095
复制
发表时间：
2023-04
期刊：
影响因子：
--
通讯作者：
Fatema Jannat;Jincheng Zhang;A. Willis;William M Ringle
中科院分区：
文献类型：
--
作者：
Fatema Jannat;Jincheng Zhang;A. Willis;William M Ringle
激光雷达技术的出现极大地扩大了古代景观的覆盖范围，从而增加了古代地表特征的数量，对考古勘探产生了革命性的影响。然而，由专家进行手工分析需要大量的时间和金钱投入。本文描述了一种用于分割的深度学习模型，即标签，感兴趣对象的语义，作为增强或取代激光雷达数据手动标记的手段。U-Net深度学习模型构成了系统的支柱，该系统已成功地在类似的激光雷达数据集上提供准确的输出。经过训练的U-Net模型集成到推理管道中，将扩展的LiDAR数据集转换为标记的输出图像。工作重点是两种语义类型的分类：(1)平台和(2)环形结构，其属性，例如位置，形状和分布，在提高我们对古玛雅文明的理解中起着重要作用。本文提供了一个基于深度学习的系统，可以有效地提取这些结构。将cnn生成的推断与专家标记的特征进行比较，以衡量算法的性能。对479平方公里的激光雷达调查结果。研究表明，CNN对环形结构和平台的IoU性能分别为0.82和0.74。讨论进一步分析了借据性能如何与这种方法作为人工标签的辅助或替代品的可行性相关。
The advent of LiDAR technology has had a revolutionary impact on archaeological prospection by vastly enlarging the coverage of ancient landscapes and consequently the number of ancient surface features. However, manual analysis by experts requires a significant time and money investment. This paper describes a deep learning model developed to segment, i.e., label, the semantics of objects of interest as a means to augment or supplant manual labeling of LiDAR data. The U-Net deep learning model forms the backbone of the system which has shown success in providing accurate outputs on similar LiDAR data set. The trained U-Net model is integrated into an inference pipeline to transform expansive LiDAR datasets into labeled output images. Work focuses on the classification of two semantic types: (1) platforms and (2) annular structures whose attributes, e.g., location, shape, and distribution, play an important role in improving our understanding of ancient Maya civilizations. This article provides a deep learning-based system that efficiently extracted these structures. CNN-generated inferences were compared against expert-labeled features to measure algorithm performance. Results for a LiDAR survey of 479 sq. km. indicate that the CNN provides an IoU performance of 0.82 and 0.74 for annular structures and platforms respectively. The discussion further analyzes how IoU performance relates to the viability of this approach as an aid or substitute for manual labeling.