A DEM-based approach for large-scale floodplain mapping in ungauged watersheds

Abstract

Binary threshold classifiers are a simple form of supervised classification methods that can be used in floodplain mapping. In these methods, a given watershed is examined as a grid of cells with a particular morphologic value. A reference map is a grid of cells labeled as flood and non-flood from hydraulic modeling or remote sensing observations. By using the reference map, a threshold on morphologic feature is determined to label the unknown cells as flood and non-flood (binary classification). The main limitation of these methods is the threshold transferability assumption in which a homogenous geomorphological and hydrological behavior is assumed for the entire region and the same threshold derived from the reference map (training area) is used for other locations (ungauged watersheds) inside the study area. In order to overcome this limitation and consider the threshold variability inside a large region, regression modeling is used in this paper to predict the threshold by relating it to the watershed characteristics. Application of this approach for North Carolina shows that the threshold is related to main stream slope, average watershed elevation, and average watershed slope. By using the Fitness (F) and Correct (C) criteria of C>0.9 and F>0.6, results show the threshold prediction and the corresponding floodplain for 100-year design flow are comparable to that from Federal Emergency Management Agency’s (FEMA) Flood Insurance Rate Maps (FIRMs) in the region. However, the floodplains from the proposed model are underpredicted and overpredicted in the flat (average watershed slope <1%) and mountainous regions (average watershed slope >20%). Overall, the proposed approach provides an alternative way of mapping floodplain in data-scarce regions.