A machine learning approach to identify barriers in stream networks demonstrates high prevalence of unmapped riverine dams

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 similar to 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.