Measuring terrain distances through extracted channel networks

This paper initiates a forensic analysis of the causes of levee failures by analyzing and extracting information from a sequence of elevation data. This is a crucial step in bettering the design and construction of levees and dams. (Fully diagnosing failures usually requires knowledge beyond the geometry of the levee, such as weather conditions and material properties). We use results from computer simulations of levee overtopping for training data. The simulations use smoothed particle hydrodynamics coupled with a well-known erodibility model. Using the sequential nature of our data, we extract important channel networks that form as the soil is scoured away. We present a series of metrics to measure the distance between channel networks to assist in determining the critical threshold value used to extract important channels from the flow network. Methods for determining this "ideal" threshold have gone mainly unexplored, and so we present a comparison of various threshold values and how closely they identify matching channel networks on sequential terrains. These threshold values allow us to identify important properties of the terrain that form its "fingerprint," a way of characterizing the geometry of the terrain. Our method for fingerprinting terrain is an important step toward the diagnosis of levee failure from digital elevation data.