A secondary assessment of sediment trapping effectiveness by vegetated buffers

Abstract

Vegetated buffers and filter strips are a widely used Best Management Practice (BMP) for enhancing streamside ecosystem quality and water quality improvement through nonpoint source pollutant removal. Most existing studies are either site-specific, rely on limited data points, or evaluate buffer width and slope as the only design variables for predicting sediment reduction, not considering other parameters such as soil texture, vegetation types, and runoff loads that can significantly influence the buffer efficiency. In this paper, we carry out a meta-analysis of published studies and fit regression models to explore the sediment removal capacity of riparian buffers. We compiled 905 data points from over 90 studies (including data from an online BMP database) documenting sediment trapping by vegetated buffers and recorded data regarding buffer characteristics such as buffer width, slope, area, vegetation type, sediment loading, water flow rates, and sediment removal efficiency. We found that an exponential regression model describing the relationship between sediment removal efficiency by the buffer and water inflow/outflow volume ratio explained 44% of the variance. Adding the square root of roughness increased the R² to 0.50. The model performance was compared with other sediment reduction regression models reported in the literature. The results point towards the importance of considering flow parameters in vegetative buffer design. The improved empirical relationships derived here can be used at local scales to understand sediment trapping potential by vegetated buffers for water quality mitigation purposes and can be built into extant hydrologic models for improved watershed-scale assessments.

1. Introduction

Naturally occurring riparian forests and streamside vegetation play a critical role in intercepting and purifying pollutant laden runoff, but their degradation by humans coupled with nonpoint source pollutant export has contributed to the deterioration of over 50% of stream and river lengths in the US (Sweeney and Newbold, 2014). Sediment pollution can clog waterways, cause flooding, and reduce water quality for domestic uses (drinking, cooking), recreational uses and/or municipal and industrial uses (Ribaudo et al., 1999). Sediment–laden water also negatively affects aquatic biodiversity and can destroy aquatic habitat by decreasing light-penetration in water, increasing water temperatures, clogging fish gills, and covering spawning areas and smothering aquatic biota (Cooper, 1993; Ribaudo et al., 1999). The US EPA ranks siltation as the leading cause of pollution of streams and rivers in the United States (US Environmental Protection Agency, 1998; Ribaudo et al., 1999).

Nonpoint source pollutants can either be managed at the source or intercepted to filter out nutrients and sediment before they reach surface waters (Dillaha et al., 1989; Ribaudo et al., 2001). The establishment and maintenance of vegetative filter strips (VFS) and riparian buffers have gained immense popularity as a cost-effective interception strategy for mitigating water quality and improving riparian ecosystem quality by nonpoint source pollutant removal (Lowrance et al., 1997; Webber et al., 2010; Sweeney and Newbold, 2014; Cole et al., 2020). Vegetative filter strips (VFS) are bands or areas of closely grown vegetation that receive and purify runoff from upslope areas such as croplands or pastures or other pollutant source areas (Dillaha et al., 1988). Vegetative filter strips and buffers perform a wide array of functions - they filter out sediments and nutrients from runoff by promoting processes such as infiltration, adsorption, plant uptake, sedimentation, and pollutant degradation through numerous biogeochemical processes (Webber et al., 2010; Pinho et al., 2008; Rahman et al., 2014). They prevent streambank erosion and improve habitat and biodiversity (Sweeney and Newbold, 2014; Lind et al., 2019; Cole et al., 2020). Several studies have documented the effectiveness of vegetated filter strips for sediment trapping. Le Bissonnais et al. (2004) reported as much as 98% decrease in sediment loads from a field using a 6 m grass strip. Duchemin and Hogue (2009) reported total suspended solid load reductions of 87% using grass strips and 85% using mixed grass and tree buffer strips. Lee et al. (1999) documented 77% and 66% sediment load reduction from adjacent crop fields using 6 m and 3 m grass buffers, respectively. Since vegetated buffers form an integral part of watersheds, either as on-site mitigation features in the form of grass hedges/vegetated filter strips at field (or other pollutant producing sites) edges or as end-of-pipe features such as riparian buffers, field and watershed scale models often need algorithms to better simulate hydrology and/or water quality through buffer systems. Hence there is a need to assess buffer effectiveness and/or load reductions from these systems through quantitative methods.

Regression (statistical) models are useful tools for water quality prediction and for making management decisions regarding buffer maintenance and pollutant attenuation (Mayer et al., 2007). Published literature has identified various factors affecting the load mitigation performance of vegetated buffers including buffer width, slope, area ratio (pollutant source area: buffer area), and hydrological flow conditions (Arora et al., 1996; Abu-Zreig et al., 2004; Boyd et al., 2003; Barfield et al., 1998; Daniels and Gilliam, 1996; Dillaha et al., 1989; Dosskey et al., 2008, 2011; Duchemin and Hogue, 2009; Gharabaghi et al., 2006; Lee et al., 1999, 2000; Deletic and Fletcher, 2006; Alemu et al., 2017; Saleh et al., 2017). A few studies have conducted meta-analysis assessments of sediment trapping-effectiveness by vegetated buffers. Liu et al. (2008) evaluated data from 31 studies and concluded that regardless of the area ratio of buffer to agricultural field, optimum sediment trapping was obtained when buffer width was 10 m and had a slope of 9%. In a meta-analysis study by Zhang et al. (2010), buffer width alone captured 37% of the total variance in sediment removal efficiency. In that study, a 30 m buffer with a slope ≈ 10% removed >85% of the sediment. These studies primarily evaluated buffer width and slope as design variables for predicting sediment reduction. While these are important design variables to consider for sediment reduction as well as for estimating costs related to buffer installation and maintenance (Dosskey et al., 2008), there is a need for evaluating buffer width and slope impacts in light of various site conditions such as soil textures, vegetation types, and runoff loads that can significantly influence buffer efficiency.

Conducting secondary analysis studies on buffer efficiency can be quite daunting because of the variability in buffer parameterization across different studies. Quantifying loads and runoff with consistent dimensions and interpretation across experimental studies poses a significant challenge because of the large variations in the objectives and site conditions in which the vegetative filter strip trials are tested. Sediment trapping efficiency (or sediment reduction) is represented by the following equation:

$$R_m=\left\{\frac{M_{in}-M_{out}}{M_{in}}\right\}\times 100\text{\%}$$ (1)

where R_m is the percent removal efficiency, M_in is the sediment mass entering the buffer, and M_out is the sediment mass leaving the buffer. However, there is a large variability in quantifying M_in and M_out across different studies. Some studies such as Dillaha et al. (1988, 1989) and Lee et al. (1989, 2000, 2003) compare a control erosion plot with a buffered erosion plot, where the dimensions of the erosion plots are the same in both cases. Other studies such as Uusi-Kämppä and Jauhiainen (2010), Tingle et al. (1998), and Thayer et al. (2012) compare different area ratios of erosion plot and buffer with each other to determine % removal. Studies such as Abu-Zreig et al. (2004) strictly evaluate loads entering and leaving the buffer area. Consequently, the units in which loads have been calculated in the various studies differ. Most studies quantify loads as load per unit area. i.e., kg/ha, kg/m², tons/ha/yr, etc. However, a significant portion of them does not specify the area over which the load was calculated. For some studies such as Lee et al. (2000) and Uusi-Kämppä and Jauhiainen (2010), the reported loads were divided by the entire area of the plot which included the erosion plot and the buffer. For other studies such as Abu-Zreig et al. (2004), the loads reported were divided over the buffer area only. This can cause inconsistencies in the reporting of the loads and calculation of sediment trapping efficiency. Moreover, many experimental approaches are used to generate runoff required for testing the effects of the buffer. One approach is to let natural rainfall generate runoff (Lee et al., 2003; Arora et al., 1996; Duchemin and Hogue, 2009; Daniels and Gilliam, 1996); which in many cases did not produce enough water for this purpose (Hay et al., 2006). Other approaches include simulated rain events (Barfield et al., 1998; Dillaha et al., 1989; Coyne et al., 1995; Chaubey et al., 1994), simulating inflows (Van Dijk et al., 1996; Deng et al., 2011; White et al., 2007), or both (Schmitt et al., 1999). These create challenges in the quantification of inflows, runoff, and rainfall, if measured at all. Meta-analysis studies have mostly overlooked these inconsistencies, which can have implications on the structure of the developed model.

Stand-alone models such as the process-based model vegetative strip model (VFSMOD; Muñoz-Carpena and Parsons, 2004) and Riparian Ecosystems Management Model (REMM; Lowrance et al., 2000) have been used to evaluate sediment reduction for different site designs and vice versa. For instance, Dosskey et al. (2008, 2011) used VFSMOD to develop graphical design aids for width and area ratio to achieve specific sediment reduction targets under a broad range of agricultural site conditions. However, VFSMOD algorithms are complex and require detailed inputs and significant computing resources to run the models and interpret results, and as such, are not used in site planning (Dosskey et al., 2008). Simpler mathematical models for buffer impacts based on theoretical equations, simplified mathematical abstractions, or regressions have been used in commonly utilized watershed models such as the Soil and Watershed Assessment Tool (SWAT). Earlier versions of SWAT considered a vegetated buffer model where trapping efficiency was solely a function of filter strip width (Neitsch et al., 2002). SWAT ver. 2012 considers an improved sediment reduction model where sediment trapping is a function of sediment loading to the buffer and runoff reduction actuated by the buffer. This model was developed by White and Arnold (2009) at the field scale using data from published literature, supplemented with data from VFSMOD simulations to deal with lack of inflow and runoff data. VFSMOD simulations were used to develop an empirical runoff reduction model in which runoff reduction was calculated as a function of runoff loading to the buffer and saturated hydraulic conductivity of the soils. A sediment reduction model was formulated based on measured data from 61 entries, and sediment reduction was quantified as a function of sediment loading to the buffer and runoff reduction. They observed that sediment loading to the buffer alone accounted for 41% of the variability in buffer sediment trapping, which increased to 64% when runoff reduction was added to the model. While predictions by VSMOD, like other physically-based models, are an inevitable manifestation of our limited understanding of reality and thus at best approximations, observed flow and sediment data are almost undoubtedly influenced by other equally important factors or processes not accounted for in the model.

Most studies on buffer sediment removal efficiency are limited to small-scale evaluation of filter strips and/or site-specific assessment of riparian areas. A comprehensive dataset can provide insights into generalizations about factors that are crucial for improving sediment removal and can inform best management practices in areas where data are scarce. This study aimed to compile a comprehensive dataset and develop a regression model while addressing the challenges in data quantification explained above. The objective was to understand if this exercise points towards similar sensitive parameters for buffer sediment trapping efficiency, as identified in the literature, while using a larger dataset.

We conducted a detailed analysis of published studies on sediment trapping, evaluated the inflow and runoff conditions using different assumptions to glean realistic field relationships between buffer characteristics and trapping efficiency. Our specific objectives were to 1) construct a database from published literature and online databases with detailed site-specific buffer characteristics such as width, slope, flow volumes/rates, sediment loading, and sediment reduction, 2) develop a sediment reduction model using multi-regression analysis as a function of various design characteristics, and, 3) compare and assess model performance to other published sediment reduction regression models for vegetated buffers. The overall objective was to obtain improved relationships that can be used at local scales to understand sediment trapping potential by vegetated buffers for water quality mitigation purposes and potentially be built into extant hydrologic models for improved watershed-scale assessments.

2. Methods

2.1. Database compilation

We searched for peer-reviewed literature using keywords filter strip, vegetated buffer, riparian buffer, vegetated filter, etc., alone or in combination to populate a Microsoft Excel® database with detailed information pertaining to sediment removal. Data was also obtained from an online stormwater BMP database (http://www.bmpdatabase.org/). We recorded detailed information about authors, buffer vegetation type, average slope, width, area, inflow and outflow loads, location, inflow and outflow volumes, and inflow and outflow rates. In addition to these attributes, we also recorded percentage sediment reduction (from Eq. (1)) and the source-buffer area ratio. We calculated percent sediment removal effectiveness in two ways depending on how data was provided: (1) as the percentage difference in loads between influent into and effluent out of the buffer, or (2) as the percentage difference in loads between the edge of a cropland with no buffer and that with a test buffer (Fig. 1). We considered data from both plot and field-scale systems; however, most entries were from experimental plots.

Fig. 1.

Percent sediment removal effectiveness was calculated in two ways depending on how data was provided. If load characteristics were reported for the buffer as presented in (a), then % removal (R_m) was calculated as the difference in loads between influent into (M₁) and effluent out of the buffer (M₂). If the study compared performances of control versus buffered sites and reported outflow load characteristics for these sites as presented in (b), then % removal (R_m) was calculated as the percentage difference in loads between edge of a cropland with no buffer (M₁) and that with a test buffer (M₂).

We collected data from >90 studies with ~905 data entries. Broadly, data were categorized as being “event-based”, where the data reported was measured for short-term events such as individual storm events or simulated rainfall events, or as “long-term”, where the measured data were reported as annual/multi-event/multi-year sums or averages. Hereafter, the terms ‘vegetated buffer’ or simply ‘buffer’ has been used to describe vegetated buffers, riparian buffers, vegetative filter strips (VFS), and vegetated hedges.

2.2. Measured variables influencing buffer sediment trapping performance

2.2.1. Buffer width, length, area

Many studies have evaluated the effect of buffer width on sediment trapping. The terms ‘width’ and ‘length’ of the buffer have been used interchangeably in the literature to describe buffer dimensions. In this study, buffer width is defined as the distance of the buffer parallel to runoff flow and buffer length as the buffer distance perpendicular to runoff flow. Since most studies included in this database are rectangular experimental plots, the buffer area is the product of the buffer length and width.

Intuitively, sediment trapping increases with an increase in buffer width. For instance, Dillaha et al. (1989) observed that increasing the buffer width from 4.6 m to 9.1 m increased sediment trapping efficiency by 14%. Coarser particles are easily trapped in the upper part of the buffer, while finer particles are harder to trap and are retained along the width of the buffer. However, at a certain buffer width, most of the sediment is effectively removed beyond which additional buffer width makes little difference (Zhang et al., 2010). A one-size-fits-all buffer width for optimum sediment trapping does not exist since buffer efficacy for trapping is influenced by multiple synergistic factors. For instance, buffers work better under the influence of shallow uniform flow than under concentrated flow conditions (Liu et al., 2008). Runoff velocity can also significantly influence trapping efficacy. Higher runoff velocities can reduce residence time within the buffer and cause erosion within the buffer, causing more sediment at the buffer outflows, thus decreasing trapping efficiency.

2.2.2. Buffer slope

Sediment trapping efficiency is also affected by the slope of the buffer. As buffer slope increases, the runoff velocity increases, potentially decreasing residence time and decreasing buffer efficacy. Gentle slopes, however, have been shown to increase trapping efficacy by facilitating laminar runoff flow through the buffer (Zhang et al., 2010). At steeper slopes, the effect of decreased residence time and increasing flow channelization dominates, causing efficiency to decrease. Dillaha et al. (1989) found that for the same buffer width, sediment trapping efficiency by the buffer increased as slope is decreased from 16% to 5–11%. Zhang et al. (2010) identified a critical slope of 10% above which buffer efficacy begins to decrease. Liu et al. (2008) observed that a polynomial regression relationship best described the influence of slope on buffer trapping efficiency. In another study by Yuan and Locke (2009), analysis of buffer efficiency plots against buffer width revealed that buffers were less effective for slopes that were > 5% than for slopes ≤5%.

2.2.3. Vegetation characteristics

The height and density of vegetation can influence buffer efficiency in trapping sediment. Dense vegetation can increase sediment deposition by decreasing the velocity of runoff. In this study, we categorized buffer vegetation as grass buffers (1), woody buffers (2), and mixed vegetation buffers (3). Grass buffers included buffers and filter strips that had mostly grasses, or stiff grass hedges, or crops (herbaceous vegetation). Woody buffers included filter strips and buffers that comprised of woody shrubs and/or trees. Mixed buffers included buffers and filter strips comprising a mix of herbaceous and woody vegetation, such as riparian buffer systems.

2.2.4. Residence time and roughness

Residence time and roughness are critical parameters that influence sediment retention within the buffer. Roughness, as used in this study, indicates above-ground obstacles to runoff and sediment flow contributed by vegetation density or by features that can hinder flow and increase sediment deposition. Increased residence time and roughness facilitate sediment deposition, thus increasing sediment retention. However, most studies did not report residence time or roughness, and very few studies measure roughness indicators such as vegetation height or vegetation density. Hence, we included proxy terms to account for roughness and residence time, as described below.

Expressions for residence time and roughness were determined by modifying the Manning’s equation as described below.

Residence time can be expressed as.

$$t=\frac{w}{u}=\frac{wLh}{Q}$$ (2)

where t is the residence time, w is the buffer width, u is the average runoff velocity, L is the buffer length perpendicular to flow, h is the depth of flow, and Q is the average flow rate through the entire buffer length.

From Manning’s equation for overland flow, we have.

$$Q=\frac{(Lh)h^{\frac{2}{3}}\surd s}{n}=\frac{1}{n}L{h}^{\frac{5}{3}}\surd s$$ (3)

where s is the buffer slope, and n is the Manning’s coefficient.

Substituting for h from Eqs. (3) in (2) we get an expression for residence time as.

$$t=\frac{w^{0.6}{n}^{0.6}}{q^{0.4}{s}^{0.3}};\text{where}q=\frac{Q}{wL}$$

$$t=n^{0.6}\left(\frac{w^{0.6}}{q^{0.4}{s}^{0.3}}\right)$$

$$\text{t}\propto \left(\frac{w^{0.6}}{q^{0.4}{s}^{0.3}}\right)$$ (4)

Substituting for Q from Eqs. (3) in (2), we get

$$t=\frac{nw}{h^{\frac{2}{3}}\surd s}$$

$$\text{t}\propto \frac{w}{\surd s}$$ (5)

Flow velocity can also be represented as a function of flow depth h, gravitational constant g, bed slope s, and channel roughness commonly described by the Darcy-Weisbach equation as.

$$u=\sqrt{\left(\frac{8ghs}{f}\right)}$$ (6)

where u is the average flow velocity and f is the Darcy-Weisbach friction factor.

For laminar flow f = 64/Re, where Re = uh/ν is the Reynold’s number (and ν is the kinematic viscosity). Substituting for f and Re in Eq. (6), we get

$$u=\frac{g{h}^{2}s}{8\nu }$$

Substituting for u in Eq. (2), we get an expression for t as

$$t=\frac{8w\nu }{g{h}^{2}s}$$

$$t\propto \frac{w}{s}$$ (7)

Substituting for u = Q/hL from Eq. (2), we get an expression for h as.

$$h=\sqrt[3]{\frac{8\nu Q}{gsL}}$$ (8)

Substituting Eqs. (8) in (2) we get another expression for t as

$$t=\frac{wL}{Q}\sqrt[3]{\frac{8\nu Q}{gsL}}$$

$$t\propto \frac{w{L}^{2/3}}{s^{1/3}{Q}^{2/3}}$$

$$t\propto \frac{w^{0.3}}{s^{0.3}{q}^{0.6}}$$ (9)

where $q=\frac{Q}{wL}$

Right-hand-side expressions for residence time from Eqs. (4), (5), (7) and (9) were used as proxies for residence time. Eqs. (4) and (5) are applicable to turbulent flow conditions (since Manning’s equation assumes turbulent flow conditions), and Eqs. (7) and (9) are for laminar flow conditions.

Manning’s roughness $n=\frac{h^{\frac{5}{3}}L\surd s}{Q}=\frac{h^{\frac{5}{3}}\surd s}{qw.}$

$$n^{*}\propto \frac{\surd s}{qw}$$ (10)

From here we used $\frac{\surd s}{qw}$ as a surrogate variable for roughness.

2.2.5. Area ratio

Research has shown that area ratio, which is the ratio of the upland contributing area to the area of the vegetated buffer, significantly impacts sediment retention by the buffer (Dosskey et al., 2011; Webber et al., 2009, 2010; Boyd et al., 2003). The contributing area is a surrogate for the size of the runoff load and the buffer area is a surrogate for trapping effectiveness of the buffer (Dosskey et al., 2011). As the area ratio increases, sediment trapping efficiency of the buffer decreases. Boyd et al. (2003) reported higher sediment reduction for 15:1 plots than 45:1 plots. However, site conditions such as soil texture, slope, and runoff rate can influence the area ratio-trapping efficiency relationship.

2.2.6. Sediment loads in runoff

While many studies reported loads in mass per unit area (kg/ha, kg/m², tons/ha), oftentimes, they failed to report the area over which loads were expressed, or that information lacked clarity. Sometimes this metric was reported as mass divided by VFS area, and other times as mass divided by the sum of erosion plot and buffer areas (Fig. 2a and b). For instance, Lee et al. (2000) compared the sediment retention of a switchgrass buffer with that of a mixed switchgrass-woody buffer for different simulated rainfall intensities and durations. Sediment loads were reported in kg/ha and included the loads transported from the bare cropland source area with either no buffer, a 4.1 m switchgrass buffer, or a 16.3 m wide switchgrass-woody buffer with the collectors located at the lower ends of these plots. Sediment loads from the bare cropland area were assumed to be loading into the buffers. Runoff volumes were converted to depth over the entire plot area (source plot with no buffer, or source plot + buffer as the case may be) (Fig. 2a). In this case, we assumed sediment loading into the buffer (L_in) as the mass of sediment contributed from the source plot area, and sediment transported from the buffer (L_out) as the mass of sediment contributed by the total plot area (source plot + buffer).

Fig. 2.

(a). Experimental plot design in study by Lee et al. (2000) where sediment loads from the bare cropland area is assumed to be loading into the buffers and runoff volumes were converted to depth over the entire plot area (source plot with no buffer, or source plot + buffer as the case may be). (b). Experimental plot design in study by Uusi-Kämppä and Jauhiainen (2010) where the dimensions of the buffered and non-buffered plots are different.

In some other studies such as Uusi-Kämppä and Jauhiainen (2010) (Fig. 2b) and Tingle et al. (1998), the dimensions of the buffered and non-buffered plots are different. For studies such as these, the sediment loads per unit area leaving the non-buffered plot were multiplied with the source area to calculate the total load. This value was then scaled for the source areas of buffered plots to get a more accurate representation of sediment loading into the buffer. Similarly, the sediment loads exiting the buffer were calculated over the entire plot area (buffer + source area).

When concentrations were reported instead of loads, sediment loads were calculated by multiplying concentration and runoff volumes. Sediment trapping efficacy by the buffer was calculated as represented in Eq. (1).

2.2.7. Inflow and outflow volumes, flow rates

Water runoff loading to the buffer combines several aspects of source area hydrology, precipitation, and area ratio (White and Arnold, 2009). Similar problems were encountered with reported runoff volumes as with sediment loads. In most cases, runoff volumes were expressed as depth over the plot area, but it was often unclear if that included the source area and the buffer or just the buffer. We calculated runoff volumes entering the buffer as the runoff volume produced from the contributing source area and runoff volume exiting the buffer was the volume contributed by the entire plot (buffer + source area).

Most meta-analysis studies do not include the effect of flow rate on the buffer’s sediment retention capacity. Evaluating the role of flow rate is critical to understand the functioning of a vegetated buffer system. Understandably so, these data are the hardest to come by. Several studies did not report runoff depths or volumes, and if they did, then flow rates were not reported. Studies that created runoff conditions by simulating rain events or evaluated buffer functioning during natural rain events, some provide the hydrograph, while others report the time over which rainfall was simulated. In this study, we tried to extract information on runoff volumes and flow rates based on the quantitative and qualitative descriptions of the study. If hydrographs were provided, they were used to estimate average inflow and outflow rates. Else, if the study employed rainfall simulators, the duration of rainfall simulation was assumed to be the duration of inflow and outflow. Runoff loading to the buffer did not include rainfall falling directly onto the buffer, but included inflows, simulated or from rainfall, occurring upslope of the buffer. However, runoff leaving the buffer included flows from rainfall occurring on the buffer as it is not possible to separate these components from the buffer outflows.

2.3. Statistical analyses

The compiled data was used to conduct statistical analyses and discern relationships between different factors influencing sediment removal efficiency and to develop a regression model predicting sediment removal by the buffers. Several procedures and assumptions were adopted for conducting statistical analyses. Studies were grouped by vegetation type and buffer width categories (0–5 m, 5–10 m, 10–20 m, and >20 m). Box plots were created to visualize the distribution of buffer sediment trapping efficiency with vegetation types and width categories. Kruskal Wallis and Wilcoxon rank-sum tests were used to inform significant differences in sediment trapping efficiency between groups; while the former test can test if there are significant differences between groups, the latter allows for pairwise comparisons of groups. Linear and nonlinear relationships between different buffer design characteristics and removal effectiveness were fitted with regression models and evaluated. Univariate regression relationships between flow variables (inflow and outflow volumes and flow rates) were also explored for potentially reconstructing missing flow data to avoid losing valuable information during model building.

Manual stepwise regression (forward) approach was utilized using linear regression fitting to construct a sediment reduction model using variables that we identified to be most relevant to buffer installation and maintenance. At every step, variables (or their transformations) were added or discarded based on their goodness-of-fit measures such as coefficient of determination (R²), Akaike’s Information Criterion (AIC), and normalized root mean square error (NRMSE). We did not consider the products of predictor variables (interaction terms), but only included the sum of variables or their transformations for model building. All analyses were performed using R version 3.1.3 (R Core Team, 2013).

3. Results and discussion

3.1. General efficacy

The literature used for this analysis is summarized in Appendix: Table 1A; it consists of 342 data entries from 53 studies and considers only entries associated with calculated positive sediment concentration and mass reduction. The table includes parameters related to buffer characteristics such as width, area, vegetation type, slope, area ratio of contributing source to buffer, sediment loading (L), flow volumes (V), and flow rates (Q). Overall, sediment removal efficiency varied from 0 to 100%, with a mean removal of 75% and median removal of 82%. For data entries categorized as long-term, efficiency ranged from 0 to 100% with mean and median efficiencies of 73 and 79%, respectively. For the event-based category, the efficacies varied from 1.3 to 100%, with mean and median efficacies of 76 and 82%, respectively.

3.2. Effect of vegetation type

The effect of vegetation type on buffer’s sediment removal efficiency were statistically significant (Kruskal Wallis H = 8.0, df = 2, p = 0.02). From the boxplot in Fig. 3, sediment trapping efficiency is significantly higher for grass buffers and mixed grass-woody vegetation than for woody vegetation-only buffers (p < 0.05; Wilcoxon rank-sum test). However, of the studies considered, only six entries had specifically woody vegetation and all of which came from a single study with a 9.1 m wide buffer along a 5% slope. Most of the data entry (293 entries) came from grasses-only buffers with widths ranging from 0.3 m to 91 m along slopes ranging from 0.5% to 50%, followed by entries from mixed vegetation buffers (43 entries) with widths ranging from 4.3 m to 23 m along slopes ranging from 1% to 21%. More data from woody and mixed buffers would have been useful to get an effective comparison of their performance with grass buffers. Median trapping efficacy for herbaceous buffers was similar to that of mixed vegetation buffers. These results are similar to those in Yuan and Locke (2009), who, in their literature review and analysis, found both forested and grassy vegetation buffers to have similar sediment trapping efficiencies. However, a common observation is that much more information exists for grass buffers than mixed buffers, and there is a need for more detailed studies on woody-only and mixed vegetation buffers.

Fig. 3.

Boxplot of % sediment reduction for different vegetation types (grass, woody, mixed). The lower and upper boundary indicate the 25th and 75th percentile, while the bold line within the box indicates the median sediment removal for each vegetation type. The bars above and below the box represent the 90th and 10th percentiles of sediment reduction respectively. The circles indicate values that lie more than one and a half times the length of the box from either end of the box. Sediment reduction is higher for grass and mixed vegetation buffers and lower for woody vegetation buffers (Kruskal Wallis p = 0.02).

3.3. Effect of buffer width and slope

Fig. 4 shows that buffer width significantly influences the buffer’s sediment reduction efficiency (Kruskal-Wallis H = 31.9, df = 3, p < 0.001), with the 10–20 m buffer width achieving statistically higher sediment reduction than either the 0–5 m, 5–10 m or > 20 m width categories (Wilcoxon rank-sum test p < 0.01). The decreased removal for larger buffer widths in some studies such as Young et al. (1980) could have resulted from the interaction between the buffers’ physical characteristics and flows, leading to increased erosion potential from the buffer.

Fig. 4.

Boxplot of % sediment reduction for different width categories. The lower and upper boundary indicate the 25th and 75th percentile, while the bold line within the box indicates the median sediment removal for each width category. The bars above and below the box represent the 90th and 10th percentiles of sediment reduction respectively. The circles indicate values that lie more than one and a half times the length of the box from either end of the box. Sediment reduction is highest for the 10 m – 20 m width category and significantly lower for the 0–5 m, 5–10 m and > 20 m width categories (Kruskal-Wallis p < 0.001).

Plotting sediment removal versus buffer width on a regression plot using all the data resulted in a very scattered view with no clear trend. Hence, for greater visual clarity, we plotted the average sediment removal for each value of buffer width instead, to discern relationships. These trends were assessed for buffer widths 1, 2, …, 40 m, which included 95% of the data. Fig. 5a shows the relationship between buffer sediment trapping efficiency and buffer width after pooling both event-based and longer-term data and considering all vegetation types together. A logarithmic regression model best describes this relationship (R² ₌ 0.36) after comparing it with linear, logarithmic, polynomial, and exponential model fits. Applying similar comparisons, Fig. 5b shows that logarithmic models best describe relationships between % sediment removal and buffer width for grass buffers (R² ₌ 0.36) and mixed buffers (R² ₌ 0.23). Exploring these trends in compiled data is part of the exploratory analyses and was not intended for model development.

Fig. 5.

Relationship between buffer sediment trapping efficiency and buffer width after pooling both event-based and longer-term data for buffers widths ≥1 m: (a) grass and mixed vegetation buffers combined, (b) grass and mixed vegetation buffers separated, (c) same with (a) but only for slopes >5% (similar relationship was obtained for grass-only buffers with slopes >5%).

*Each data point represents the average sediment reduction for each value of buffer width.

Yuan and Locke (2009) observed that buffers were more effective for slopes less than 5% (see page 10) than for those with slope ≤ 5%. In our analysis, we observed no clear relationship between width and removal for slopes ≤5%. Fig. 5c shows the relationship between buffer width and average sediment removal for slopes >5%, considering all vegetation types. This relationship is best described using a logarithmic relationship (R² = 0.50). A similar relationship was obtained for grass-only buffers with slopes >5%. Zhang et al. (2010) observed that a critical slope of 10% changed the relationship between slope and sediment reduction from positive to negative. In this study, we could not clearly identify any critical slope which influenced the direction of impact on sediment reduction by the buffer.

3.4. Volumes and flow rates

Volume ratio (V_r), calculated as inflow volume divided by outflow volume, was used to evaluate the effect of runoff reduction on sediment reduction. Fig. 6 shows the relationship between volume ratio and sediment removal by the buffer. The relationship is best explained using an exponential regression model (R² = 0.37). We found that volume ratio, also a proxy for flow rate ratio in this study (since the volume and flow rate ratios were 99.9% correlated), is the best predictor for sediment removal yet in this study. These results are consistent with those derived by White and Arnold (2009), where they observed that runoff reduction explained around 20% of the variability in their measured data.

Fig. 6.

Relationship between volume ratio and sediment removal by the buffer. The relationship is best explained using an exponential regression model (R² = 0.37).

Other variables, including flow rates and volumes, area ratio, roughness, residence time parameters (from Eqs. (4), (5), (7) and (9)), and sediment loading into the buffer, did not present clear trends on plots against sediment removal.

3.5. Sediment removal multiple regression model

In developing a sediment reduction model, only event-based data were considered, which included 273 entries. Of these, flow rates could be calculated for fewer than 200 entries. To avoid losing valuable information for model development and to enable the inclusion of flow rate as a predictor in model construction, we attempted to reconstruct missing flow data based on univariate regression relationships with a good fit (high R²) between relevant variables. This included regression relationships between outflow rate and outflow volume (Q_out = 14.4 * V_out; R² = 0.80; N = 170) and that between ratios of flow rates and flow volumes (1:1 relationship with R² = 1, N = 170) (Appendix: Table 1B). Of these, we considered data for which buffer widths were < 40 m, slopes under 40%, and studies where inflow volume was no more than 60 times the outflow volume and for which outflow volumes were no more than 3 times the inflow volumes, i.e., 1/3 ≤ V_in/V_out ≤ 60. These bounds were considered since most of the event-based data entries fell within these boundaries. Omitting entries with missing values and applying the conditions above resulted in a total of 194 entries which were used to construct the sediment reduction model.

Factors considered for the model building process were buffer width (w), slope (s), sediment load per unit buffer area (L_inA), inflow rate over unit buffer area (Q_inA), average flow rate over the buffer (Q_av; calculated as 0.5*(Q_in + Q_out)), ratio of flow volumes (V_in/V_out or V_r), roughness factor (n*), residence time factors (t₁, t₂, t₃ and t₄ from Eqs. (4), (5), (7) and (9) respectively), and their square, log and exponential transformations, where A = wL is buffer area. Predictors (or their transformations) were added in a stepwise forward manner to the multiple regression model based on the significance of their coefficients, overall R², AIC and NRMSE values (Table 1). R² varies from 0 to 1, and better model performance is indicated by a higher R². AIC is an indicator of model parsimony, and a lower AIC relative to other models represents a more parsimonious model. Care was taken to avoid highly correlated variables in the same model. NRMSE varies between 0 and 1, with values closer to 0 indicating better model fit.

Table 1.

Model building using forward stepwise regression considering event based data for all vegetation types together (for the data presented in Appendix: Table 1B). Only models with significant coefficients are included in this table. Better model performance is indicated by a higher R² and lower NRMSE both of which vary between 0 and 1. AIC is an indicator of model parsimony, and a lower AIC relative to other models is representative of a more parsimonious model. The highlighted model has the highest R² and lowest AIC, and was chosen as the final model describing sediment removal by vegetated buffers.

Models	R2	AIC	NRMSE
$R_m\sim \text{a}+\text{b}\ast \left(V_r\right)$	0.10	1738.2	0.94
$R_m\sim \text{a}+\text{b}\ast \log \left(V_r\right)$	0.34	1679.7	0.81
$R_m\sim \text{a}+\text{b}\ast \left(V_r^2\right)$	0.03	1753.5	0.98
$R_m\sim \text{a}+\text{b}\ast \left({\surd V}_r\right)$	0.21	1714.2	0.89
$R_m\sim \text{a}+\text{b}\ast \exp .\left({\text{c}}^{\ast }{V}_r\right)$	0.44	1650.5	0.75
$R_m\sim \text{a}+\text{b}\ast \exp .\left(\text{c}\ast V_r\right)+\text{d}\ast \log (w)$	0.45	1647.3	0.74
$R_m\sim \text{a}+\text{b}\ast \exp .\left(\text{c}\ast V_r\right)+\text{d}\ast \exp .(w)$	0.46	1645.9	0.74
$R_m\sim \text{a}+\text{b}\ast \exp .\left(\text{c}\ast V_r\right)+\text{d}\ast (w^2)$	0.45	1649.5	0.74
$R_m\sim \text{a}+\text{b}\ast \exp .\left(\text{c}\ast V_r\right)+\text{d}\ast \left(\surd w\right)$	0.44	1650.8	0.75
$R_m\sim a+b\ast \exp .\left(c\ast V_r\right)+d\ast \left(t_1^2\right)$	0.45	1649.6	0.74
$R_m\sim \text{a}+\text{b}\ast \exp .\left(\text{c}\ast V_r\right)+\text{d}\ast \log \left(t_2\right)$	0.46	1643.2	0.73
$R_m\sim \text{a}+\text{b}\ast \exp .\left(\text{c}\ast V_r\right)+\text{d}\ast \left(\surd t_2\right)$	0.45	1647.9	0.74
$R_m\sim \text{a}+\text{b}\ast \exp .\left(\text{c}\ast V_r\right)+\text{d}\ast \left(n^{\ast }\right)$	0.50	1630.4	0.71
$R_m\sim \text{a}+\text{b}\ast \exp .\left(\text{c}\ast V_r\right)+\text{d}\ast \left(n^{\ast 2}\right)$	0.48	1636.9	0.72
$R_m\sim \text{a}+\text{b}\ast \exp .\left(\text{c}\ast V_r\right)+\text{d}\ast \left(\surd n^{\ast }\right)$	0.50	1629.7	0.71
$R_m\sim \text{a}+\text{b}\ast \exp \left({\text{c}}^{\ast }{V}_r\right)+\text{d}\ast \log \left(n^{\ast }\right)$	0.48	1635.7	0.72
$R_m\sim \text{a}+\text{b}\ast \exp .\left({\text{c}}^{\ast }{V}_r\right)+\text{d}\ast \exp .\left(n^{\ast }\right)$	0.49	1632.8	0.71
$R_m\sim \text{a}+\text{b}\ast \exp .\left({\text{c}}^{\ast }{V}_r\right)+\text{d}\ast \left(t_4\right)$	0.46	1646.0	0.74
$R_m\sim \text{a}+\text{b}\ast \exp .\left(\text{c}\ast V_r\right)+\text{d}\ast \left(t_4^2\right)$	0.46	1643.5	0.73
$R_m\sim \text{a}+\text{b}\ast \exp .\left(\text{c}\ast V_r\right)+\text{d}\ast \left({\surd t}_4\right)$	0.46	1646.2	0.74
$R_m\sim \text{a}+\text{b}\ast \exp .\left(\text{c}\ast V_r\right)+\text{d}\ast \log \left(t_4\right)$	0.47	1642.3	0.73

The ratio of flow volumes (V_r) alone accounted for 44% of the variability in the observed data. Adding the square root of the surrogate variable of roughness factor (n*) increased the R² to 50%. The final model for % sediment reduction is as follows (Table 2):

$$R_m=107.2-{82.8}^{\ast }{\text{e}}^{\left(-{0.8}^{\ast }{V}_r\right)}-{35.5}^{\ast }{\surd n}^{\ast }$$ (11)

where $V_r=V_{\mathit{\text{in}}}/V_{\mathit{\text{out}}}$, and $n^{\ast }=\left(\frac{\surd s}{qw}\right)$ from Eqn 10, where q is in L/min/m² of the buffer and w is in m. From the sediment reduction plot in Fig. 7, the model over-predicts for smaller values of observed reduction, but predicts higher values with greater precision.

Table 2.

Multiple regression models for predicting % sediment reduction for the data presented in Appendix: Table 1B. Results are also presented for the application of other meta-analysis models to this data.

	This study	This study (considering Vr alone)	White and Arnold (2009)	Liu et al. (2008)	Zhang et al. (2010)
Full model	a + b * exp. (c * Vr) + d * (√n*)	a + b * exp. (c * Vr)	a + b * LinA + c * (Rr%)	a + b * (w) + c * (s) − d*(s2)	k * (1 − exp. (− b * w))
	a = 107.22* (100.05, 114.38)	a = 97.81* (91.83, 103.79)	a = 67.22* (64.23, 70.20)	a = 71.05* (62.50, 79.60)	k = 77.01* (73.73, 80.28)
	b = − 82.80* (− 100.14, − 65.45)	b = − 86.13* (− 108.20, − 64.06)	b = − 0.23 (− 0.58, 0.12)	b = − 0.01 (− 0.62, 0.60)	b = 1.79* (1.00, 2.58)
	c = − 0.78* (− 1.04, − 0.51)	c = − 0.92* (− 1.25, − 0.59)	c = 0.32* (0.26, 0.38)	c = 279.36* (80.18, 476.54
	d = − 35.45* (− 49.76, − 21.14)			d = 1963.32* (939.34, 2987.29)
R2	0.50	0.44	0.41	0.10	0.02
AIC	1629.74	1650.46	1658.68	1743.31	1755.78
NRMSE	0.71	0.75	0.76	0.95	0.99
N	194	194	194	194	194
Grasses-only	a + b * exp. (c * Vr) + d * (√n*)	a + b * exp. (c * Vr)	a + b * LinA + c * (Rr%)	a + b * (w) + c * (s) − d*(s2)	k * (1 − exp. (− b * w))
Model	a = 107.11* (100.29, 113.92)	a = 97.48* (91.71, 103.26)	a = 68.24* (65.14, 71.33)	a = 70.80* (62.03, 79.57)	k = 77.76* (74.04, 81.48)
	b = − 80.01 * (− 96.88, − 63.15)	b = − 84.88* (− 107.49, 62.27)	b = − 0.27 (− 0.60, 0.07)	b = − 0.23 (− 0.82, 0.37)	b = 1.72* (0.98, 2.46)
	c = − 0.76* (− 1.03, − 0.50)	c = −0.94* (−1.28, −0.59)	c = 0.310* (0.26, 0.36)	c = 429.21* (224.82, 635.60)
	d = − 36.59* (− 50.54, − 22.63)			d = 3226.20* (2153.23, 4299.17)
R2	0.57	0.49	0.47	0.26	0.03
AIC	1273.42	1296.34	1301.40	1355.25	1393.61
NRMSE	0.66	0.71	0.72	0.86	0.98
N	154	154	154	154	154
Mixed-only	a + b * exp. (QinA) + c * log (Vr)		a + b * LinA + c * (Rr%)	a + b * (w) + c * (s) − d*(s2)	k * (1 − exp. (− b * w))
Model	a = 72.97* (65.75, 80.19)		a = 58.90* (40.10, 77.69)	a = 74.86* (57.77, 91.94)	a = 79.85* (73.20, 86.49)
	b = − 0.01* (− 0.01, − 0.004)		b = 9.90 (− 13.00, 33.81)	b = 0.58 (− 1.25, 2.40)	b = 1.00 (− 1.64, 3.64)
	c = 20.43 (6.92, 33.94)		c = 0.51* (0.19, 0.82)	c = − 70.84 (− 428.75, 287.06) d = − 507.72 (− 2323.22, 1307.78)
R2	0.48		0.25	0.03	0.00
AIC	279.66		292.34	303.61	300.60
NRMSE	0.71		0.85	0.97	0.99
N	35		35	35	35

^*Indicates significance at 0.05 level.

Fig. 7.

Plot of predicted sediment reduction (using model R_m = 107.2–82.8*e ^{(− 0.8_*V_r)} − 35.5*√n*) versus observed sediment reduction.

Model performance was compared with that derived from applying other sediment reduction regression models developed in the literature, such as those of White and Arnold (2009), Liu et al. (2008), and Zhang et al. (2010), to data entries in this study (Table 2). Of these, only the model presented by White and Arnold (2009) had a comparable performance with the next highest R², with the Liu et al. (2008) and Zhang et al. (2010) models having much lower performances. This could be because of the inclusion of runoff reduction in the White and Arnold (2009) model (calculated as [(V_in– V_out) *100]/V_in)). This was not the case with the models presented in Liu et al. (2008), which only considers buffer slope and width, and Zhang et al. (2010), which considers only buffer width While the model in the White and Arnold (2009) study had a significant correlation between sediment loading and sediment reduction with their data (R² = 41%), our study did not present any significant relationship with sediment loading. This could be because of the larger number of entries considered for this study (194 entries) as opposed to 61 entries used in the White and Arnold (2009) paper. Another reason for this difference could be how we accounted for sediment loads and runoff volumes in this study.

For grass buffers (N = 154), V_r alone accounted for 49% of the variability in the observed data (Table 1). Adding the square root of n further increased the R² to 57%. The only other comparable model was the White and Arnold (2009) model (R² = 47%). The final sediment reduction model for mixed buffers (N = 35) included inflow rate and the volume ratio with an R² of 48% (Table 2), which was higher than the other models.

The results of this study point towards the importance of considering flow in buffer design. This is evident from the better fit obtained for models that considered flows versus those that did not. However, the overall model accounted only for 50% of the total variance (and 57% of the variance for grass buffers). Several other factors may be responsible for the large percentage of unaccounted variation. This study only considers sediment removal under uniform sheet flow conditions. However, it is not possible to ensure complete sheet flow conditions in these systems. Surface structure variations due to surface peaks and depressions in the flow path, differential infiltration capacities along the flow path, presence of vegetation and other organic matter, rainfall on the buffer, sediment deposition, and erosion can cause convergence or divergence of flow. Moreover, sediment accumulation can change the microtopography of the buffer aiding flow concentration (Gharabaghi et al., 2006). Studies such as Helmers et al. (2005) observed a decline in buffer’s sediment reduction with an increase in flow convergence. Adjustments for flow concentration or divergence was not considered in this study. Factors such as vegetation height, density, shape, and resilience can greatly influence the sediment deposition within the buffer. The denser the vegetation, the more sediment can be trapped by the buffer. However, vegetation density could not be accounted for here due to the lack of reporting in underlying studies. Further research is needed as more information becomes available.

This model has many limitations. While an effort was made to include data with uniform sheet flow, it is possible that studied buffers had unintended flow concentration that went unnoticed or unreported since it can be challenging to maintain these conditions. For instance, buffer slopes intuitively influence runoff velocity and consequently flow conditions. Gradual slopes promote laminar sheet flow while steeper slopes can cause flow to concentrate, leading to lesser sediment reduction. Moreover, most of the data in the study come from experimental plots, which many researchers suggest overestimate the real effectiveness of vegetated buffers (White and Arnold, 2009). Experimental plots are typically tested over the short-term, and do not account for long-term sediment accumulation in the buffer that can decrease its sediment trapping efficiency (Sweeney and Newbold, 2014; Liu et al., 2017). Another overlooked factor is rainfall on the buffer. This was deliberately overlooked in an attempt to conduct a systematic comparison since studies showed much variation in how they initiated runoff, ranging from natural rain events on source area and buffer, to rain events only on the contributing source area, to rainfall on entire areas combined with sediment-laden inflows, and so on. The intensity and duration of rainfall on the filters and antecedent soil moisture conditions can have a significant influence (Rahman et al., 2013) on the functional capacity of these buffer systems. Most experimental plots consider sediment delivery from smaller contributing areas or small design storms (in case of simulated rainfall), when it is the storms with >10 year return interval that deliver most sediment over the long term (Sweeney and Newbold, 2014).

The science of implementation, management, and restoration of various types of vegetated buffers for a multitude of uses is continually evolving with many new studies being published since the data for this study was compiled (such as Haukos et al., 2016, Saleh et al., 2017, Alemu et al., 2017, and Luo et al., 2020). Since the data collected and analyzed for our study are quite extensive, we feel that a reanalysis including these newer data points will not necessarily result in a significant difference in the results, and hence these studies have not been included. However, it will be worth revisiting in the future once significant additional data become available to add to the data points in this study.

4. Conclusions

Sediment reduction by vegetated buffers are a consequence of synergistic influences of the physical dimensions and characteristics of the buffer as well as their site-specific hydrological responses to local runoff/storm events. Very few models consider hydrological responses, understandably, because of the lack of detailed information reported in the literature. This study compiled a comprehensive database of 53 studies consisting of 342 entries, which were used for evaluating the influence of various factors, including the buffer’s physical and hydrological characteristics on sediment reduction capacity. After applying constraints to the database, 194 entries were eventually selected to construct a sediment reduction model using a stepwise forward linear regression approach. A regression equation considering the hydrological response of the buffer, included as the exponential transformation of ratio of flow volumes V_in/V_out, was observed to be the most influential factor in predicting sediment reduction capacity accounting for 44% of the variance in measured sediment reduction. The addition of terms considering the roughness factor improved the performance to account for 50% of the variability in measured sediment reduction. The model developed here outperformed existing models which considered fewer variables and/or which did not consider hydrological responses when applied to the data in this study – R² was slightly higher than a previously published sediment reduction model which included runoff reduction (White and Arnold, 2009), and significantly higher than models which considered only physical characteristics such as buffer width and slope (Liu et al., 2008; Zhang et al., 2010). The sediment reduction model from this study may be used in conjunction with other models when necessary. For instance, where runoff characteristics are lacking, models such as the VFSMOD derived runoff reduction equation from White and Arnold (2009) may be used to calculate runoff reduction and determine the buffer’s sediment reduction efficiency.

A high level of uncertainty still exists regarding the required conditions for the installation of the optimum buffer. Nevertheless, this study provides a valuable insight into the relevant variables of vegetated buffer systems through the extensive data compiled, including the importance of hydrological considerations. The improved sediment trapping efficiency regression model may be built into existing watershed models, such as SWAT, to evaluate sediment control at basin scales. Relationships derived from this study can be highly valuable when appropriately quality-controlled data are limited, field-monitored results are lacking, and cost of sediment-traps and associated analytical costs are too high. Overall, the results of this study can prove useful in assessing the effectiveness of site-specific buffers to inform management decisions regarding buffers and sediment control.