A machine learning approach to identify barriers in stream networks demonstrates high prevalence of unmapped riverine dams.
A machine learning approach to identify barriers in stream networks demonstrates high prevalence of unmapped riverine dams.
复制标题
用于识别河流网络中障碍的机器学习方法表明，未绘制地图的河流水坝的普遍存在。
DOI：
10.1016/j.jenvman.2021.113952
复制
发表时间：
2022
影响因子：
8.7
通讯作者：
B. Rahm
中科院分区：
文献类型：
--
作者：
B. Buchanan;S. Sethi;Scott Cuppett;Megan E. Lung;George Jackman;Liam J. Zarri;E. Duvall;J. Dietrich;Patrick H. Sullivan;Alon Dominitz;J. Archibald;A. Flecker;B. Rahm
通过拆除未使用或废弃的水坝来恢复河流生态系统的完整性已成为全球流域保护的优先事项。然而，恢复连通性的努力受到精确水坝清单的限制，这些清单往往忽略了未绘制地图的较小河流水坝。在这里，我们开发并测试了一种机器学习方法，使用公开可用的地形和地理空间栖息地数据的组合来识别未映射的水坝。具体来说，我们训练了一个随机森林分类算法，使用数字工程预测变量和已知的水坝位置来识别未映射的水坝进行验证。我们将算法应用于美国哈德逊河流域的两个子流域，量化了连通性的影响，并评估了一系列预测集，以检查分类精度和模型参数化工作之间的权衡。随机森林分类器使用与河流坡度和上游生境存在相关的变量子集，在预测水坝位置方面取得了很高的准确性（真阳性率= 89%，假阳性率= 1.2%）。未测绘的水坝在两个试验流域普遍存在。事实上，现有的水坝清单低估了水坝的真实数量80-94%。考虑到以前未绘制的水坝，洄游鱼类的树突连通性指数下降了62-90%。未映射的水坝可能无处不在，可能会极大地影响流连接信息。然而，我们发现机器学习方法可以提供一种准确和可扩展的方法来识别未绘制的水坝，可以指导制定准确的水坝清单，从而为更好地管理水坝提供信息和授权。
Restoring stream ecosystem integrity by removing unused or derelict dams has become a priority for watershed conservation globally. However, efforts to restore connectivity are constrained by the availability of accurate dam inventories which often overlook smaller unmapped riverine dams. Here we develop and test a machine learning approach to identify unmapped dams using a combination of publicly available topographic and geospatial habitat data. Specifically, we trained a random forest classification algorithm to identify unmapped dams using digitally engineered predictor variables and known dam sites for validation. We applied our algorithm to two subbasins in the Hudson River watershed, USA, and quantified connectivity impacts, as well as evaluated a range of predictor sets to examine tradeoffs between classification accuracy and model parameterization effort. The random forest classifier achieved high accuracy in predicting dam sites (true positive rate = 89%, false positive rate = 1.2%) using a subset of variables related to stream slope and presence of upstream lentic habitats. Unmapped dams were prevalent throughout the two test watersheds. In fact, existing dam inventories underestimated the true number of dams by ∼80–94%. Accounting for previously unmapped dams resulted in a 62–90% decrease in dendritic connectivity indices for migratory fishes. Unmapped dams may be pervasive and can dramatically bias stream connectivity information. However, we find that machine learning approaches can provide an accurate and scalable means of identifying unmapped dams that can guide efforts to develop accurate dam inventories, thereby informing and empowering efforts to better manage them.