Bank process observations
Bank processes were observed at a total of 68 locations across the three study streams (two banks observed in 34 reaches, each 100 m long). Sites were chosen by dividing each of the study watersheds into sub-watersheds using ArcGIS (ArcHydro extension) (Release 9.3.1, ESRI, Redlands, CA, USA) and digital elevation models of the watersheds, and this information was used to identify sub-watersheds for sampling.
Thorne (1998) was used as a guide for identifying bank failure types and making bank observations. In this study, field observations of bank processes focused on the main types of bank failures extensively described and photographed by Thorne (1998). These include shallow slides, referred to as slumps (which are common in low-cohesion banks and often occur after rotational slips and slab failures), rotational slips (which occur in highly cohesive materials), slab-type block failure, referred to as slab failures (associated with steep bank angles and cohesive materials), cantilever failure (failure of overhanging material, which usually occurs in banks with layers of material that alternate from cohesive to non-cohesive), pop-out failure (which results from bank seepage processes in a cohesive, steep bank), and piping failure (collapse of a portion of a bank because of groundwater flow rates and pressures).
Bank characteristics observed in the field were also developed using guidelines discussed in Thorne (1998). An average bank angle was calculated for each bank by averaging measurements of the upper, middle, and lower bank slopes made with an Abney level. Bank angles used to calculate an average bank angle for each site were taken with the Abney level at 25% of the slope for upper bank measurements, 50% of the slope for middle bank measurements, and 75% of the slope for lower bank measurements. Bank characteristics were examined at each site. Characteristics used to model bank failure status using statistical analyses included: bank material composition, bank failure location, bank-face vegetation characteristics, bank failure status, and bar association. Bank material composition was determined in the field by hand from bulk samples pulled from each part of the bank face—upper, middle, and lower and described using USDA textural classes (silty clay, silty clay loam, silt loam, silt, sandy clay, sandy clay loam, sandy loam, loamy sand, sand, clay loam, and loam). Subsequently, the texture of each bank was classified into one of three groups for statistical analysis based on the dominant texture of each bank. Dominant texture was based on the first term or the single term in some instances (clay, silt, sand, and loam) of the USDA textural classification for soil separates for upper, middle, and lower banks at each site. If a textural term was used as the first term or the only term for two or more sections of the bank face, then it was chosen as the overall textural characterization for the bank. After reviewing the bank textural classifications made in the field, it was determined that the overall texture of the banks fit into one of three groups: sandy, loamy, or composite (sand and/or loam with a clay unit present). These categories represent varying degrees of bank resistance to erosion and mass failure, with sandy being the most failure prone and composite the least. Bank failure location was noted as either inside bend, outside bend, or straight, and was noted to determine whether the propagation of bends through the system as part of increased lateral migration play a role in determining occurrence of bank failures. Bank-face vegetation was characterized as being herbaceous cover, tree cover, or no vegetation present to determine any effects that vegetation may have on stabilizing bank faces.
Bank failure status was rated as active, dormant, or none. Active bank failures showed evidence of recent failure activity, including uprooted, damaged, or dead vegetation, exposed roots, convex or vertical bank faces, fresh (unweathered) bank material exposures, and unworked deposits of failed bank material at the base of banks. Dormant bank failures exhibited evidence that failures had occurred in the past but had since ceased. Such evidence included visible, weathered failure scars with one more of the following: mature trees with bent trucks (the presence of reaction wood) that ultimately straighten, concave/‛relaxed bank slopes, and bank material vegetated with mature trees or saplings. If no evidence could be found to support either active or dormant bank failure, the bank was rated as not having failures present (‘none’) and was excluded from further analyses. The presence of bars and their location relative to banks (opposite bank, attached to bank, or no bar present) was also noted to determine whether bars attached to banks initiate bank stabilization (bank failure dormancy) by protecting the bank from shear stresses experienced during storm flows. Additionally, an average channel w/d ratio was calculated for each reach using cross-sectional measurements of average channel width and average depth (from top of banks) made using a laser level with detector with a manufacturer specified accuracy of 1 cm.
Modeling bank failure status
With exception of average bank angle and channel w/d ratio, most of the data collected in the field yielded categorical data. The classification of bank status as either active or dormant meant that the dependent variable was binary. As a result of both of these factors, logistic regression was chosen to model the effects of different bank characteristics on determining bank failure status.
Logistic regression is commonly used in the social sciences and, in recent years, has become increasingly prevalent in the physical sciences. In geomorphology, logistic regression has been used to model and predict the location of landslides (Eeckhaut Van Den et al., 2006; Chang et al., 2007), to assess shallow earthflow susceptibility (Tolga et al., 2005), and to predict changes in channel pattern (Bledsoe and Watson, 2001). It is primarily used for two purposes: (i) to determine which independent variables (predictors) are important to the occurrence or non-occurrence of the dependent variable and (ii) to create a parsimonious equation to be used to predict the occurrence of the dependent variable. In this study, multiple logistic regression was used to determine which independent variables (average bank angle, channel w/d ratio, bank material composition, bank-face vegetation, failure location, and bar association) are important to determining the likelihood that banks will have active or dormant bank failure (the dependent variable), and in so doing, highlight the factors that contribute to bank stabilization over time.
For this research, a six parameter multiple logistic regression equation was used in the form of the following:
where x1 is average bank angle, x2 is channel w/d ratio, x3 is bank material composition, x4 is bar association, x5 is failure location, and x6 is bank-face vegetation. Because having a large range of values can make interpreting regression coefficients difficult, the measurements for average bank angle and channel w/d ratio were converted to z-scores and normalized prior to being put into the logistic regression model. Categorical independent variables were coded prior to analyses. Categories for each of the independent variables are summarized in Table 1.
Table 1. Summary of categories for independent and dependent variables used in the logistic regression model
|Bank angle||W/D ratio||Vegetation||Erosion location||Bar association||Bank material||Failure status|
|No categories||No categories||None||Absent||Absent||Composite||Dormant|
| || ||Herbaceous||Outside meander||Behind bar||Loamy||Active|
| || ||Trees||Inside meander||Opposite bar||Sandy|| |
In cases where a variable had two categories, the number 0 was assigned to one category and the number 1 to the other. For example, bank status had two categories: dormant = 0 and active = 1. In the case of independent variables with three categories (the maximum possible in this study), values 0–2 were assigned, and these were re-coded by spss (Release 19.0.0, IBM SPSS Statistics for Macs, Armonk, NY, USA) to be binary. In such instances, spss treats one category, specified by the user, as a reference value, and codes the other two categories as either 0 or 1, which are then compared with the reference value to determine their relative importance within the group before being added to the model. For example, the variable ‘bank composition’ had three categories: sandy, loamy, and composite (having unconsolidated layers with a cohesive clay unit). It was coded by spss as follows: composite = reference, 0 = loamy, and 1 = sandy. When doing logistic regression, it is advisable to run the model in a variety of ways (entering all independent variables at once or one at a time) to discern which method results in the most robust model overall (Elliott and Woodward, 2007). For this study, the logistic regression was run in spss in two ways. Firstly, by having all independent variables included at the start; and secondly, by adding independent variables one at a time to see if the model improved with each addition. Both methods resulted in the same variables being statistically significant/insignificant, but the first method (all independent variables entered at once) yielded the more robust model overall according to standard model diagnostics (model summary output, Hosmer and Lemeshow Test, and classification table output, all of which are discussed in more detail later). As a result, output from the first model was chosen for reporting.