Selecting geomorphic variables for automatic river segmentation: Trade-offs between information gained and effort required

Abstract

Identifying river segments with apparently distinct geomorphic characteristics but relatively homogeneous internal features may be critically helpful in designing network analysis for characterization, environmental assessment and river management. Automatic segmentation procedures using geographic tools and statistical methods provide objective and replicable results. In particular, multivariate procedures may be appropriated for different purposes such as coping with the multiple dimensionality of river systems. Although there is an increasing number of studies dealing with segmentation using different sets of morphological variables, the influence that the selected variables have on segmentation results is rarely assessed. In this context, we defined five combinations of frequently used geomorphic variables (i.e., channel slope, active channel width, valley bottom width, channel confinement and specific stream power), and compared the obtained segmentation results. We study the upper Esla River network, covering a total length of 294 km, with the largest two rivers regulated by large dams. Variables were measured at successive river sections 200 m apart. Five segmentation results were obtained in which we compared the number and characteristics of the segments, and the internal variability and the suitability of predicting river dynamics (i.e., occurrence of bare and vegetated gravel bars). The number of segments per kilometer of river and their average length were different among segmentations but varied much more across rivers than across segmentations. In general, segmentations including channel slope and active channel width performed better in predicting the occurrence of bare gravel bars than segmentations based on stream power or valley confinement. When splitting the initial data set into regulated and non-regulated segments, differences in predicting the occurrence of gravel bars were found, with better results in the case of non-regulated rivers. Channel slope and active channel width showed a reduced explanatory power for the regulated reaches. Finally, we conclude that primary geomorphic variables such as channel slope and active channel width were more efficient than secondary variables such as stream power, which may encompass more information but needed additional data and the use of empirical models, with greater effort and much uncertainty. In the case of regulated rivers, automatic segmentation including the affected variables (e.g., active channel width) may help in detecting differences in geomorphic sensitivity to river adjustments across the resulting segments downstream from the dams. Our results offer valuable insights into the selection of geomorphic variables for river segmentation analysis, in which trade-offs between the information gained and the effort required must be considered according to the respective research targets.

Basin‐wide hydromorphological analysis of ephemeral streams using machine learning algorithms‡

“…These objective methods provide an alternative to expert judgement, as they are mainly based on visual and graphical estimations, and allow system discretization based on statistical criteria to be carried out. Compared to methods using only one variable (e.g., Guertault et al, 2018; Notebaert & Piégay, 2013; Parker et al, 2012), multivariate methods offer the possibility of simultaneously considering the complementary information provided by several variables (Bizzi & Lerner, 2012; Martínez‐Fernández et al, 2019).…”

A statistical-based reach scale classification for the lower Tapajós river channel, eastern Amazonia

“…Similar approaches have been employed to assess the impacts of interventions in large river fluvial channels (Knox and Latrubesse, 2016), restoration strategies, habitat identification (Bizzi and Lerner, 2012), and detailed characterization of reaches (Orlowski et al, 1995). The advantages of multivariate approaches include greater objectivity and replicability in the segmentation stage (Martínez-Fernández et al, 2019) and cost reduction, making it possible to investigate river systems at regional (Phillips and Desloges, 2015) or even national scales (Gaucherel et al, 2017).…”