Modeling wet headwater stream networks across multiple flow conditions in the Appalachian Highlands
Jensen, Carrie K.
McGuire, Kevin J.
Dolloff, C. Andrew
MetadataShow full item record
Despite the advancement of remote sensing and geospatial technology in recent decades, maps of headwater streams continue to have high uncertainty and fail to adequately characterize temporary streams that expand and contract in the wet length. However, watershed management and policy increasingly require information regarding the spatial and temporal variability of flow along streams. We used extensive field data on wet stream length at different flows to create logistic regression models of stream network dynamics for four physiographic provinces of the Appalachian Highlands: New England, Appalachian Plateau, Valley and Ridge, and Blue Ridge. The topographic wetness index (TWI) was the most important parameter in all four models, and the topographic position index (TPI) further improved model performance in the Appalachian Plateau, Valley and Ridge, and Blue Ridge. We included stream runoff at the catchment outlet as a model predictor to represent the wetness state of the catchment, but adjustment of the probability threshold defining wet stream presence/absence to high values for low flows was the primary mechanism for approximating network extent at multiple flow conditions. Classification accuracy was high overall (> 0.90), and McFadden's pseudo R2 values ranged from 0.69 for the New England model to 0.79 in the Appalachian Plateau. More notable errors included an overestimation of wet stream length in wide valleys and inaccurate reach locations amid boulder deposits and along headwardly eroding tributaries. Logistic regression was generally successful for modeling headwater streams at high and low flows with only a few simple terrain metrics. Modification and application of this modeling approach to other regions or larger areas would be relatively easy and provide a more accurate portrayal of temporary headwaters than existing datasets.