Date of Award:


Document Type:


Degree Name:

Master of Science (MS)


Civil and Environmental Engineering

Committee Chair(s)

Brian Mark Crookston


Brian Mark Crookston


Jeffery S. Horsburgh


Steven L. Barfuss


Low Head Dams are man-made structures that span the width of a river and allow water to flow over them. Although these structures vary in age and original purpose, some pose a life-threatening danger when recreationalists are entrapped in the reverse current at the toe. Despite these dangerous structures causing many drowning deaths in the USA each year, most low head dams do not have safety regulations and remain unknown to those charged with dam and public safety. Thus, to improve public safety at these structures, the first step is to fully inventory low head dams, so that subsequent efforts (i.e., safety signage, hazard analysis, rehabilitation, removal, etc.) may be undertaken. A joint task committee between the American Society of Civil Engineers (ASCE), Environmental and Water Resources Institute (EWRI), Association of State Dam Safety Officials (ASDSO), and United States Society on Dams (USSD) was formed. This committee relies on volunteers to: 1) share known locations and existing data sets and 2) crowdsourcing to scan aerial imagery to identify low head dams manually. To support these efforts, this study was undertaken based in ArcGIS Pro. After developing a custom image preparation Python code, a deep learning model was developed that can scan high resolution aerial imagery to locate unknown low head dams for review and confirmation by trained experts, often licensed professional engineers.

Development and training of the model included the preparation of images of known low head dams in Indiana using an ArcGIS Pro geoprocessing tool. Another tool trained the deep learning model and a third tool detected dams. It was found that the quality of the training images had a significant impact on the quality of the deep learning model and the accuracy of detected dams. The quality of training images was improved through the development of Python code to supplement the first tool used.

In total, four iterations of the tools were used to develop the final deep learning model. The final model was tested on another set of low head dams in Indiana that were not used in training. The model was able to find 89% of the highly visible dams, characterized as those low head dams that have visible white water on the downstream side and are not blocked by vegetation. It is able to detect other not-as-visible dams with variable levels of accuracy. The model is considered to perform comparably to humans manually scanning imagery for these structures.

The final model and supporting data are available open source so that others may use the results of this research to locate low head dams in other locations or adapt this tool for other feature identification efforts. This supporting data includes the Python code and a detailed methodology. Since performance is highly dependent upon dam visibility in an image and image resolution, it is anticipated that this model could undergo adaptation and additional training for other regions (e.g., different topography, landcover, etc.) or identification of other man-made structures.