Considerations for Augmented Flood Control Infrastructure Inspection Using Convolutional Neural Networks

Earthen flood control structures such as dams and levees are vital for managing flood risk. In the United States alone, the United States Army Corps of Engineers manages approximately 14,000 mi of levees, with an average age of 59 years. Current inspection standards consist of walking or driving along the levee looking for important performance indicator signs. Manually inspecting levees for such conditions is time consuming, in certain instances possibly dangerous to inspectors, and is susceptible to human subjectivity in some instances with the sheer quantity of levees to inspect. In recent years across the broader civil infrastructure industry, remote inspection technologies, such as unmanned aerial vehicles, have begun seeing use and allowing for systems to be inspected faster and safer than traditional methods. The data obtained from these remote inspection tools can then be utilized with machine learning algorithms to classify and identify both failures, and potential future failure locations. This combination of rapid remote inspection for data collection, and machine learning for data processing, provides a powerful tool to enhance the condition monitoring of levees rapidly, efficiently, and remotely. This paper outlines some current technologies and use cases of these tools for levee inspections and discusses ongoing work related to machine vision and machine learning automation of visual imagery applied to crack detection using convolutional neural networks, where an accuracy of up to 99.76% is achieved for crack detection using a database of 20,800 images of concrete.