Using machine learning to identify karst sinkholes from LiDAR-derived topographic depressions in the Bluegrass Region of Kentucky
Using machine learning to identify karst sinkholes from LiDAR-derived topographic depressions in the Bluegrass Region of Kentucky
复制标题
DOI：
10.1016/j.jhydrol.2020.125049
复制
发表时间：
2020-09
影响因子：
6.4
通讯作者：
Junfeng Zhu;Adam M. Nolte;Nathan Jacobs;M. Ye
中科院分区：
文献类型：
--
作者：
Junfeng Zhu;Adam M. Nolte;Nathan Jacobs;M. Ye
现有天坑的分布和特征的信息是了解岩溶含水层系统和评估天坑灾害的关键。激光雷达提供了准确和高分辨率的地形信息，并已被用于改善许多岩溶地区的天坑划定。然而，激光雷达数据也揭示了许多地形凹陷，通过人工目视检查从这些凹陷中识别天坑可能是缓慢而费力的。为了提高识别过程的效率，我们将六种机器学习方法（逻辑回归，朴素贝叶斯，神经网络，随机森林，RUSBoost和支持向量机）应用于LiDAR衍生地形凹陷的形态特征数据集。来自肯塔基州的蓝草地区的波旁县、伍德福德县和埃克塞特县的天坑数据被用来导出用于训练和测试机器学习方法的数据集。数据集由22，884条记录组成，每条记录有10个变量。对于每种方法，80%的记录的随机子集用于训练，剩余的20%用于测试。受试者工作特征曲线显示，所有6种方法均适用于数据集，如所有曲线下面积（AUC）均大于0.87所示。神经网络是表现最好的方法，AUC为0.95，测试平均准确度为0.85。为了进一步改进天坑映射过程，我们随后开发了一个两步过程，将经过训练的神经网络分类器和手动目视检查相结合，并将该过程应用于同样位于蓝草地区的斯科特县。我们能够通过人工检查仅27%的地形凹陷来定位该县97%的天坑，神经网络将其归类为具有相对较高的天坑概率。该研究表明，机器学习是一种很有前途的方法，以提高识别效率的岩溶地区，其中高分辨率的地形信息。
Information about the distribution and characteristics of existing sinkholes is critical for understanding karst aquifer systems and evaluating sinkhole hazards. LiDAR provides accurate and high-resolution topographic information and has been used to improve delineation of sinkholes in many karst regions. LiDAR data also reveal many topographic depressions, however, and identifying sinkholes from these depressions through manual visual inspection can be slow and laborious. To improve the efficiency of the identification process, we applied six machine learning methods (logistic regression, naive Bayes, neural network, random forests, RUSBoost, and support vector machine) to a dataset of morphometric characteristics of LiDAR-derived topographic depressions. Sinkhole data from Bourbon, Woodford, and Jessamine Counties in the Bluegrass Region of Kentucky were used to derive the dataset for training and testing the machine learning methods. The dataset consisted of 22,884 records with 10 variables for each record. For each method, a random subset of 80% of the records was used for training and the remaining 20% was used for testing. The test receiver operating characteristic curves showed that all six methods were applicable to the dataset, as demonstrated by all area under the curves (AUCs) being greater than 0.87. Neural network emerged as the method that performed best, with an AUC of 0.95 and a testing average accuracy of 0.85. To further improve the sinkhole mapping process, we subsequently developed a two-step process that combined the trained neural network classifier and manual visual inspection and applied the process to Scott County, also in the Bluegrass region. We were able to locate 97% of the sinkholes in the county by manually inspecting only 27% of the topographic depressions the neural network classified as having relatively high probabilities of being sinkholes. This study showed that machine learning is a promising method for improving sinkhole identification efficiency in karst areas in which high-resolution topographic information is available.