Automated large‐scale mapping and analysis of relict charcoal hearths in Connecticut (USA) using a Deep Learning YOLOv4 framework
Automated large‐scale mapping and analysis of relict charcoal hearths in Connecticut (USA) using a Deep Learning YOLOv4 framework
复制标题
DOI：
10.1002/arp.1889
复制
发表时间：
2022-12
影响因子：
1.8
通讯作者：
Wouter Verschoof‐van der Vaart;A. Bonhage;A. Schneider;W. Ouimet;T. Raab
中科院分区：
文献类型：
--
作者：
Wouter Verschoof‐van der Vaart;A. Bonhage;A. Schneider;W. Ouimet;T. Raab
在过去的十年中，许多研究使用来自欧洲和美国东北部不同森林地区的数字高程模型（DEM）数据，成功地手动绘制了数千个以前的木炭生产地点（也称为废弃木炭炉）。这些地点的存在导致土壤物理和化学性质发生重大变化，称为遗留效应，因为大量的木炭留在土壤中。在景观中发现的大量木炭炉需要使用自动化方法来绘制和分析这些地形。我们提出了一种基于开源数据和软件的新方法，在大规模LiDAR数据集中自动检测遗留木炭炉（用简单局部地形模型可视化）。此外，该方法同时提供了一般和特定领域的信息，可用于进一步研究遗留效应。不同版本的方法在康涅狄格州西北部的数据上进行了微调，随后在康涅狄格州的两个不同地区进行了测试。结果表明，尽管需要额外的后处理来处理激光雷达质量的变化，但这些性能足够，F1得分在0.21到0.76之间。经过测试，预测模型的最佳版本（平均F1得分为0.56）被应用于整个康涅狄格州。结果显示，与该州已知的木炭炉分布明显重叠，同时也发现了新的浓度。这表明了该方法在大规模数据集上的可用性，即使地形和激光雷达质量发生变化。
In the past decade, numerous studies have successfully mapped thousands of former charcoal production sites (also called relict charcoal hearths) manually using digital elevation model (DEM) data from various forested areas in Europe and the north‐eastern USA. The presence of these sites causes significant changes in the soil physical and chemical properties, referred to as legacy effects, due to high amounts of charcoal that remain in the soils. The overwhelming amount of charcoal hearths found in landscapes necessitates the use of automated methods to map and analyse these landforms. We present a novel approach based on open source data and software, to automatically detect relict charcoal hearths in large‐scale LiDAR datasets (visualized with Simple Local Relief Model). In addition, the approach simultaneously provides both general as well as domain‐specific information, which can be used to further study legacy effects. Different versions of the methodology were fine‐tuned on data from north‐western Connecticut and subsequently tested on two different areas in Connecticut. The results show that these perform adequate, with F1‐scores ranging between 0.21 and 0.76, although additional post‐processing was needed to deal with variations in LiDAR quality. After testing, the best performing version of the prediction model (with an average F1‐score of 0.56) was applied on the entire state of Connecticut. The results show a clear overlap with the known distribution of charcoal hearths in the state, while new concentrations were found as well. This shows the usability of the approach on large‐scale datasets, even when the terrain and LiDAR quality varies.