79 Th e Universal Soil Loss Equation (USLE) and its derivatives are widely used for identifying watersheds with a high potential for degrading stream water quality. We compared sediment yields estimated from regional application of the USLE, the automated revised RUSLE2, and ? ve sediment delivery ratio algorithms to measured annual average sediment delivery in 78 catchments of the Chesapeake Bay watershed. We did the same comparisons for another 23 catchments monitored by the USGS. Predictions exceeded observed sediment yields by more than 100% and were highly correlated with USLE erosion predictions (Pearson r range, 0.73?0.92; p < 0.001). RUSLE2- erosion estimates were highly correlated with USLE estimates (r = 0.87; p < 001), so the method of implementing the USLE model did not change the results. In ranked comparisons between observed and predicted sediment yields, the models failed to identify catchments with higher yields (r range, ?0.28?0.00; p > 0.14). In a multiple regression analysis, soil erodibility, log (stream ? ow), basin shape (topographic relief ratio), the square-root transformed proportion of forest, and occurrence in the Appalachian Plateau province explained 55% of the observed variance in measured suspended sediment loads, but the model performed poorly (r2 = 0.06) at predicting loads in the 23 USGS watersheds not used in ? tting the model. Th e use of USLE or multiple regression models to predict sediment yields is not advisable despite their present widespread application. Integrated watershed models based on the USLE may also be unsuitable for making management decisions. Empirical Models Based on the Universal Soil Loss Equation Fail to Predict Sediment Discharges from Chesapeake Bay Catchments Kathleen B. Boomer,* Donald E. Weller, and Thomas E. Jordan Smithsonian Environmental Research Center Reduced soil fertility and sharp declines in aquatic resources have intensi? ed worldwide e? orts to limit erosion and reduce pollution in streams and rivers. In the Chesapeake Bay region, excess sediment and sediment-sorbed phosphorus have degraded shallow estuarine habitats, especially as human development has intensi? ed during the past 50 yr (Kemp et al., 2005). Despite reduced point source contributions, sediment delivery remains high, and estuarine habitat quality remains impaired (e.g., Stankelis et al., 2003). Although these trends indicate the in? uence of nonpoint pollution sources (Boesch et al., 2001), identifying the most in? uential nonpoint sources remains di? cult, especially for sediment and phosphorus (Weller et al., 2003). Empirical and simulation models provide important tools for integrating our understanding of processes controlling water and material discharges, characterizing human interactions in the landscape, and predicting discharges from ungauged basins (Carpenter, 1996). Models derived from the Universal Soil Loss Equation (USLE) are some of the most widely applied tools for predicting sediment yield from whole catchments (Kinnell, 2004a; Yoder et al., 2004). Th ese models are based on the original USLE devel- oped to help farmers minimize topsoil loss on agricultural ? elds (Wischmeier and Smith, 1978): A = R K LS CP where A is the estimated long-term annual soil loss (Mg soil loss ha?1 yr?1), R is a rainfall and runo? factor representing the summed erosive potential of all rainfall events in a year (MJ mm ha?1 h?1 yr?1), L and S are topographic factors that describe slope length and steepness (dimensionless), K is the soil erodibility factor representing units of soil loss per unit of rainfall erosivity (Mg ha h ha?1 MJ?1 mm?1), and CP characterizes land cover and conservation management practices (dimensionless). Subsequent enhancements, notably the Revised USLE (RUSLE1 and RUSLE2), incorporate a broader set of land cover classes and attempt to capture deposition in complex terrains (Re- nard et al., 1997). Th e core USLE factors were re? ned by enabling users to characterize additional subfactors (Yoder et al., 2004). Process-based equations for transport capacity and deposition Abbreviations: CN, Curve Number; CP, USLE cover-practice management factor; K, soil erodibility; LS, USLE length-slope topographic factor; NLCD, National Land Cover Database; R, rainfall erosivity; RUSLE, revised-Universal Soil Loss Equation; SDR, sediment delivery ratio; SERC, Smithsonian Environmental Research Center; USLE, Universal Soil Loss Equation. Smithsonian Environmental Research Center, Edgewater, MD 21037-0028. Copyright ? 2008 by the American Society of Agronomy, Crop Science Society of America, and Soil Science Society of America. All rights reserved. No part of this periodical may be reproduced or transmitted in any form or by any means, electronic or mechanical, including pho- tocopying, recording, or any information storage and retrieval system, without permission in writing from the publisher. Published in J. Environ. Qual. 37:79?89 (2008). doi:10.2134/jeq2007.0094 Received 20 Feb. 2007. *Corresponding author (boomerk@si.edu). ? ASA, CSSA, SSSA 677 S. Segoe Rd., Madison, WI 53711 USA TECHNICAL REPORTS: LANDSCAPE & WATERSHED PROCESSES 80 Journal of Environmental Quality ? Volume 37 ? January?February 2008 were incorporated to compute deposition on concave slopes, at dense vegetative strips, in terrace channels, and in sediment basins (Foster et al., 2003). Th e most recent release (RUSLE2) estimates annual erosion rates by summing the products of fac- tor values for each day, rather than using annual average values, and is considered the best tool for estimating rill and interrill erosion rates for conservation planning (Foster et al., 2003). Because the USLE and RULSE predict ?edge-of-? eld? erosion and do not account for the interaction among adjacent ? eld plots, catchment erosion estimates often are adjusted downward by a sed- iment delivery ratio (SDR). Most published SDRs are developed empirically by relating the ratio of observed sediment delivery rates and USLE predicted erosion rates to landscape characteristics such as watershed size (Vanoni, 1975), watershed shape (Maner, 1958), or the slope of the main stream channel (Williams and Berndt, 1972). More complex algorithms include distributed models that estimate SDRs from topographic variation and ? ow path length (Yagow et al., 1998) or additional surface characteristics including ? ow-path roughness and gradient, slope shape, and soil moisture, such as SEDMOD (Spatially Explicit Delivery MODel; Fraser et al., 1998). Th e USLE-SDR predictions remain widely used for estimating annual soil loss at the catchment scale (Trimble and Crosson, 2000a) in ungauged drainage basins (e.g., Angima et al., 2003; Martin et al., 2003; Lu et al., 2004; Boellstor? and Benito 2005; Fu et al., 2005; Onyando et al., 2005; Wang et al., 2005). A number of watershed models also rely on the USLE estimates for model parameterization. Examples include GWLF (General- ized Watershed Loading Function; Haith and Shoemaker, 1987), AGNPS (AGricultural Non-Point Source; Young et al., 1989), SWAT (Soil & Water Assessment Tool; Arnold and Allen, 1992), applications of HSPF (Hydrological Simulation Program-Fortran; Bicknell et al., 1993), and SEDD (Sediment Delivery Distributed model; Ferro and Porto, 2000). Although the USLE framework is used extensively for catch- ment-scale empirical or simulation models (Kinnell, 2004a; Yoder et al., 2004), it is not well veri? ed at that scale (Trimble and Crosson, 2000b). Watershed modelers often cite the USLE?s extensive ? eld validation based on more than 10,000 plot-years of data (Nearing et al., 2000), but that validation is for gross erosion rather than sediment transport to down-gradient areas. Previous validation e? orts indicate the USLE reliably estimates erosion from individual land units (e.g., Risse et al., 1993; Ali and Sharda, 2005), and prediction of stream sediment yield has been deemed successful in small catchments where ? eld observa- tions of catchment geography, rather than regional spatial data, are used in the USLE calculation (e.g., Angima et al., 2003). In contrast, validation of soil erosion estimates using regional data from 98 catchments across Europe indicated the USLE-based empirical models provided poor predictions of observed stream sediment delivery (Van Rompaey et al., 2003). In addition, the study incorporated ? eld measurements to derive basin-speci? c SDRs. Th is method limits the utility of USLE applications to monitored basins and precludes meeting the important goal of predicting loads from ungauged catchments. Although many researchers, including the developers, have cautioned against ap- plying the plot-scale USLE at the catchment scale, (Wischmeier and Smith 1978; Risse et al., 1993; Kinnell, 2004a), widespread application continues, suggesting there is a clear need for valida- tion at whole watershed scale. We tested the ability of USLE-based models to predict whole catchment sediment discharges using two independent data sets. Observed annual average sediment yields were com- pared with estimates from Geographic Information System implementations of the original USLE, the RUSLE2.0, and the erosion models enhanced with sediment delivery ratios to explore whether the base model or any of its derivatives e? ec- tively predicted whole catchment discharges. Based on their widespread acceptance, we expected that some USLE-based models could provide good predictions of sediment yield from watersheds. We further expected that more recent ver- sions that account for sediment deposition (SDRs) or include other improvements would give better predictions than earlier versions. Even if yield predictions were not quantitatively accurate, we expected all the USLE implementations to cor- rectly separate erosion-prone watersheds from those with little erosion. Th at is, we expected that across many watersheds, the rank correlation of observed sediment yield with sediment yield predicted by any of the USLE-based models should be strong and statistically signi? cant. Materials and Methods Study Location Th e Smithsonian Environmental Research Center (SERC) established continuous monitoring stations in 78 drainage basins across the 166,000 km2 Chesapeake Bay watershed (Jordan et al., 1997a; Liu et al., 2000). Catchment sizes ranged between 5 and 91,126 ha. Th e basins were arranged in clusters throughout the Chesapeake watershed (Fig. 1; Table 1), which extends over six physiographic provinces (Langland et al., 1995): the Coastal Plain (n = 45 study watersheds), Piedmont (n = 10), Mesozoic Lowland (n = 7), Appalachian Mountain (n = 9), and Appalachian Plateau (n = 7). Basins were selected with di? ering proportions of agricultural and nonagricultural land cover and no reservoirs or point sources to observe the e? ects of land cover on water quality in dif- ferent physiographic provinces. Land cover ranged from 2 to 100% forest, 0 to 39% agriculture, and 0 to 82% residential and commercial development (1992 National Land Cover Database [NLCD]; Vogelmann et al., 2001). Percent impervi- ous area ranged from 0 to 40% (Regional Earth Science Ap- plication Center [RESAC]; Goetz et al., 2003). Data from 23 additional drainage basins in the Chesapeake Bay watershed where USGS monitoring programs provided annual mean estimates of sediment delivery were evaluated (Table 2) (Gellis et al., 2005; Langland et al., 1995). Esti- mates of annual average sediment yield were based on data collected daily or determined from the ESTIMATOR model (Cohn et al., 1992). Catchment sizes ranged between 101 and 90,530 ha. Land cover ranged from 5 to 100% forest, 0 to 40% agriculture, and 0 to 30% development. Catchments with reservoirs were not included in our analyses. Boomer et al.: USLE-based Empirical Models Fail to Predict Sediment Discharges 81 Water Monitoring Data In the 78 SERC study catchments, automated samplers were used to moni- tor stream depth continuously and to collect ? ow-weighted water samples com- posited weekly for at least 1 yr between 1974 and 2004. Th is approach e? ectively quanti? ed dissolved and suspended mate- rials, including those transported episodi- cally during storm ? ows (Jordan et al., 1997a). Samples were analyzed for a suite of constituents including total suspended solids (Jordan et al., 1997a). Total sus- pended solids in unpreserved composite samples were collected on weighed 0.45-?m membrane ? lters, rinsed with distilled water to remove salts, dried in a vacuum desiccator, and reweighed. An- nual mean ? ow rates and ? ow-weighted mean concentrations were multiplied to estimate annual average loads and di- vided by catchment area to produce yield estimates (Mg ha?1 yr?1). Digital Data Sources Digital spatial data sets describing topography, land cover, and soil character- istics were acquired from public websites. Watershed delineations (Baker et al., 2006a) and topographic variables, includ- ing slope and slope variation, were derived from the USGS National Elevation Da- tabase (EROS Data Center, 1999; source resolution: 27.78 m pixels). Land cover estimates were obtained from the 1992 NLCD (Vogelmann et al., 2001) (source resolution, 30 m pixels). Th e 21 land cover classes were consolidated into six classes: cropland, grassland, development, forest, wetland, and barren areas. Surface soil erodibility was derived from the USDA-NRCS STATSGO soils database (USDA-NRCS, 1995) (scale: 1:250,000) converted to a grid with 30-m resolution. Th e proportion of impervious area for each catchment was derived from the RESAC dataset (Goetz et al., 2003) (source resolution, 30 m). Physiographic province was determined from sur? cial geology maps (Langland et al., 1995) (scale: 1:500,000). Monthly average precipitation amounts (1971?2000) were provided by the Spatial Climate Analysis Service (2002) (scale, 1:250,000). Data Analysis Grid-based USLE Analysis Average erosion rates for each catchment were calculated from an erosion grid (resolution, 30 m), which was a product of USLE-factor grids. Slope length (L) equaled the pixel width of the National Elevation Dataset (NED) (27.78 m). Slope steep- ness (S) was de? ned as the maximum change in elevation of the NED within a 3 ? 3 grid cell neighborhood. Th e resulting LS grid was resampled to a 30-m resolution using cubic interpola- tion (ArcGIS 9.1). Rainfall erosivity (R) was derived from linear interpolation of the national iso-erodent map (Wischmeier and Smith, 1978). Cover management factors (C) were assigned based on land cover derived from the consolidated NLCD classes. We did not di? erentiate erosion control practices, and the support practice factor (P) was set to one. Surface soil erodibility (K) was extracted and rasterized from the STATSGO database. Fig. 1. Locations of study catchments within physiographic provinces of the Chesapeake Bay watershed. Black shading indicates basins monitored by the Smithsonian Environmental Research Center (SERC); striped shading indicates basins monitored daily by USGS. Inset shows the location of the Chesapeake Bay watershed in the eastern USA. 82 Journal of Environmental Quality ? Volume 37 ? January?February 2008 RUSLE Analysis Gross erosion rates were estimated using the automated ver- sion of the Revised-USLE2 (Renard et al., 1997; USEPA, 2004). Th e application identi? es potential sediment transport routes, us- ing raster grid cumulation and maximum downhill slope meth- ods (Van Remortal et al., 2001), and depositional zones based on signi? cant changes (>50%) in slope (Hickey et al., 1994). Slope length is subsequently de? ned as the distance from the origin of an overland ? ow path to a point where deposition occurs or where the ? ow path converges with others to form a de? ned channel (Van Remortal et al., 2001). Cover and practice (CP) values were calculated from the RUSLE database, which incor- porated a wider range of land use and land cover characteristics than our USLE application by using additional information from county-level agricultural censuses (Yoder et al., 2004). Applied CP values also incorporated climatic e? ects (Foster et al., 2003). Sediment Delivery Ratios Five variations of SDR models were implemented, includ- ing three lumped-parameter models that estimate the SDR from watershed area (in square miles) or watershed slope: SDR = 0.42 ? Area?0.125 (Vanoni, 1975) SDR = 0.417762 ? Area?0.134958 ? 0.127097 (USDA-NRCS, 1983) Log(SDR) = 2.943 ? 0.824 log(L/R) (Maner, 1958) where L = maximum length of watershed, and R = watershed relief, represented by the di? erence between average elevation Table 1. Regional landscape characteristics of drainage basins within the Chesapeake Bay watershed that were monitored by the Smithsonian Environmental Research Center (SERC). Landscape metric/predictor? Appalachian Plateau Appalachian Mountain Mesozoic Lowland Piedmont Upland Coastal Plain Number of watersheds 7 9 7 10 45 Area (ha) 990?32,672 700?28,637 533?7873 45.3?3241 5.2?91,126 Mean annual R? (MJ mm ha?1 h?1 yr?1) 1370?1400 1885?1970 2400?2550 2550?2550 2800?3425 Mean USLE (LS) (dimensionless) 1.76?3.70 0.51?5.06 0.59?2.91 0.98?2.12 0.07?1.72 Mean K (Mg ha h ha?1 MJ?1 mm?1) 0.03 0.02?0.04 0.04 0.04 0.02?0.06 Mean USLE CP factor (dimensionless) 0.08?0.10 0.09?0.13 0.08?0.17 0.07?0.11 0.05?0.26 Mean A (Mg soil loss ha?1 yr?1) 6000?16,000 3300?22,000 5800?26,700 8500?25,500 400?22,500 Mean 30-m pixel diff erence in elevation (m) 8?12.5 3?15 3?11 5.5?9.5 0?9 Basin relief ratio 0.02?0.07 0.04?0.18 0.01?0.24 0.02?0.27 0.01?1.43 % Cropland 0?1 0?31 0?39 0?22 0?61 % Development 0?1 0?1 0?5 0?1 0?82 % Impervious area <1 0?2 0?6 0?1 0?39 % Forest 2?98 73?95 4?99 11?92 9?100 Weighted CN 68?74 60?79 69?74 60?68 57?81 Annual runoff (cm yr?1) 23?32 21?46 31?38 15?30 14?52 Mean annual stream fl ow (cm yr?1) 49?80 54?80 49?82 23?46 10?104 Mean annual sediment yield (Mg ha?1 yr?1) 0.09?0.5 0.02?1.2 0.3?0.9 0.05?0.5 0.01?1.2 Mean annual total P yield (kg ha?1 yr?1) 0.3?0.8 0.1?3.0 0.6?3.5 0.2?0.8 0.0?4.6 ? A, erosion rate; CN, curve number; CP, cover-practice; K, soil erodibility; LS, length-slope; R, rainfall erosivity; USLE, Universal Soil Loss Equation. Table 2. Regional landscape characteristics of selected drainage basins within the Chesapeake Bay watershed that were monitored by USGS. Landscape metric/predictor Appalachian Plateau Appalachian Mountain Mesozoic Lowland Piedmont Upland Coastal Plain Number of watersheds 5 2 5 7 4 Area (ha) 2607?11,981 3874?8854 101?2894 1510?90,527 963?29,062 Annual R? (MJ mm ha?1 h?1 yr?1) 1480?1600 2090?2320 2160?2720 2710?3080 3240?3360 Mean USLE LS (dimensionless) 2.72?6.36 1.81?2.84 0.12?3.12 0.62?1.40 0.10?1.23 Mean K (Mg ha h1 ha?1 MJ?1 mm?1) 0.03 0.03 0.03?0.04 0.04 0.03?0.05 Mean USLE CP factor (dimensionless) 0.01 0.01?0.02 0.01?0.05 0.02?0.06 0.02?0.05 Mean A (Mg soil loss ha?1 yr?1) 8700?25,600 11,400?13,500 500?19,000 6400?12,900 1300?15,900 Mean 30 m pixel diff erence in elevation (m) 10.1?16.1 7.9?10.3 0.7?10.5 4.1?8.1 0.5?6.8 Basin relief ratio 0.16?0.51 0.09?0.13 0.06?0.19 0.07?0.30 0.03?0.05 % Cropland 0?4 5?12 0?14 1?12 3?39 % Development 0?1 0?1 1?30 1?23 1?19 % Impervious area 0?1 0?1 2?19 1?10 2?14 % Forest 48?99 48?59 6?64 30?38 28?73 Weighted CN 61?76 65?71 68?78 64?69 60?68 Annual runoff (cm yr?1) 18?31 20?40 32?43 21?35 17?31 Mean stream fl ow (cm yr?1) 35?70 44?59 36?55 36?45 35?42 Mean annual sediment yield (Mg ha?1 yr?1) 0.03?1.0 0.2?0.6 0.6?2.8 0.1?3.7 0.05?0.3 ? A, erosion rate; CN, curve number; CP, cover-practice; K, soil erodibility; LS, length-slope; R, rainfall erosivity; USLE, Universal Soil Loss Equation. Boomer et al.: USLE-based Empirical Models Fail to Predict Sediment Discharges 83 of the watershed divide and the watershed outlet. We also used two distributed models that calculate a SDR for each pixel and estimate the proportion of eroded sediment that is transported from each cell to the stream channel. Th e ? rst model weights erosion estimates by ? ow path distance and by the relief of the potential sediment source above the stream: SDR = exp(?0.4233 ? ? ow path length [m] ? slope factor) where slope factor = exp{?16.1 ? (relief to stream/? ow path length + 0.057)} ? 0.6 (Yagow et al., 1998). Th e second distributed model was provided by USEPA in combination with the automated RUSLE2 program (USEPA, 2004). SEDMOD provides pixel-based SDR estimates based on ? ow-path slope gradient and shape, vegetation surface roughness, stream proximity, and soil texture and moisture content (Fraser et al., 1998). For each catchment, ? ve predic- tions of sediment delivery resulted by multiplying erosion estimates by the ? ve SDR algorithms. Annual Runoff Th e Curve Number (CN) method (USDA Soil Conserva- tion Service, 1986) was used to estimate runo? potential and annual runo? . STATSGO data describing hydrologic soil groups were combined with the land use data and assigned a curve number. Monthly runo? was calculated from monthly average precipitation data (Spatial Climate Analysis Service, 2002) and summed to derive annual average runo? . Correlation and Multivariate Regression Analyses Zonal statistics (ArcGIS 9.1) were used to derive sum- mary values of landscape characteristics for each catchment. Spearman rank correlation coe? cients were used to assess how well the observed sediment yields corresponded with pre- dictions from the USLE-based models. Pearson correlations were used to compare results among the USLE-based models and to compare USLE-predicted erosion rates with input factors (i.e., LS, R, CP, and K). For the multiple regression analysis, additional predictor variables considered included physiographic province, watershed size, variation in terrain complexity (determined by comparing slopes in adjacent grid cells), topographic relief ratio, land cover proportions, percent impervious area, runo? potential, and annual aver- age runo? (determined using the CN method). Continuous variables except rainfall erosivity (R), mean annual erosion, and basin topographic relief ratio were log10 transformed, and percentage variables were square-root transformed to improve normality and reduce heteroscedascity before analysis. We used Pearson correlation coe? cients to detect and eliminate redundant variables. Univariate linear regressions were used to detect which of the independent variables best predicted annual sediment yield. Best subsets multiple regressions were subsequently used to model the relationship between sedi- ment yield and catchment landscape features. Results Correlation of Observed Sediment Yields with USLE- based Predictions Th e USLE and RUSLE erosion estimates were strongly cor- related and not statistically di? erent for SERC drainage basins (Pearson r = 0.87; p < 0.001) (Fig. 2) or USGS drainage basins (Pearson r = 0.95; p < 0.001). Variation between USLE and RUSLE predictions di? ered by physiographic province; RUSLE estimates were consistently higher than USLE for catchments in the Appalachian Plateau and Appalachian Mountain regions and lower for catchments in the Coastal Plain and Piedmont regions. Di? erences between the two models were larger in basins predict- ed to have higher erosion rates by both models. Because of the strong correlation between the USLE and RUSLE predictions, we focus on the USLE results for subsequent analyses, including implementation of the SDR equations and the statistical analyses. When erosion predictions were compared with measured sediment yields from the SERC or the USGS monitoring sites, rank correlations were numerically low (Spearman r range, ?0.22 to ?0.02) and not statistically signi? cant (p range, 0.21? 0.69). Many of the highest observed sediment yields occurred in the Coastal Plain where low erosion rates were predicted. Adjusting predicted erosion rates with SDR equations did not improve the rank correlations between observed and pre- dicted sediment yield across the region or within physiographic provinces of the Chesapeake Bay watershed (Table 3). Sediment delivery ratio predictions were strongly correlated with the un- adjusted USLE predictions (r range, 0.73?0.92; p < 0.001), and correlations between observed and SDR-predicted sediment yield were numerically low for both SERC drainage basins (Spearman r range, ?0.17 to ?0.08; p range, 0.14?0.50) and USGS drainage basins (Spearman r range, ?0.28 to 0.00; p range, 0.20?0.99). Th e Yagow algorithm was not implemented for USGS-moni- tored drainage basins because of the similarity in all other results and the extensive e? ort required for calculating ? ow path dis- Fig. 2. Universal soil loss equation (USLE) versus revised USLE (RUSLE2) predictions of mean annual soil erosion rates (Mg ha?1 yr?1). Symbols indicate the physiographic province in which a catchment is located. APPL, Appalachian Plateau; APMN, Appalachian Mountain; MELO, Mesozoic Lowland; PDUP, Piedmont Upland; CP, Coastal Plain. 84 Journal of Environmental Quality ? Volume 37 ? January?February 2008 tance and ? ow path relief across each drainage basin. Estimates from RUSLE applications yielded similar negative correlations between observed and predicted sediment yields (Fig. 3). We examined which of the component USLE factors most strongly in? uenced SERC and USGS basin erosion estimates and found that these corresponded primarily with the aver- age LS factor (Table 4). Th e importance of the LS factor was emphasized by the signi? cant but negative correlations of erosion estimates with other USLE input factors known to promote erosion. For example, higher annual rainfall erosiv- ity (R) counterintuitively corresponded with lower estimated erosion rates due to the predominant in? uence of topography on rainfall distribution (Pearson r = ?0.79; p < 0.001). Land cover patterns were also associated with di? erences in basin topography. Land cover classes presumed to be more sus- ceptible to erosion, including developed land and cropland, occupied a greater proportion of area in drainage basins with lower mean LS factors (Pearson r = ?0.30; p = 0.002 for LS and developed land and r = ?0.45; p < 0.001 for LS and crop- land), whereas forest land cover predominated in catchments with steeper slopes (r = 0.56; p < 0.001 for LS and forest). As a result, there was signi? cant negative correlation between the CP factor and the LS factor (r = ?0.56; p < 0.001). Lower C values, presumed to represent land uses that promote soil stability and limit erosion, were also counterintuitively associ- ated with higher erosion rates (r = ?0.45; p < 0.001). Relating Sediment Discharges to Geographic Factors Using data from all 78 SERC catchments, correlation anal- yses of annual sediment yield with the independent variables indicated seven signi? cant univariate relationships between observed stream water quality and catchment characteristics (Table 5). Before the best subsets multiple regression analysis, estimates for percent impervious and cropland areas were re- moved because of signi? cant covariance with urban and forest areas, respectively. Th e best univariate predictors of sediment delivery included the USLE input factors K and CP, percent development, percent forest cover, annual average runo? , and mean observed stream ? ow. Th e multiple regression analysis identi? ed a set of six variables that together explained 55% of the observed variance in annual sediment yield (F[5,72] = 17.31; p < 0.001) (Table 6). Th e six variables chosen were soil erodibility (K), log(stream ? ow), basin shape (log basin topo- graphic relief ratio), square-root?transformed values of per- cent forest, and occurrence in the Appalachian Plateau. How- ever, the best subsets multiple regression model performed poorly when the USGS observations of sediment yield were used as a veri? cation data set (Pearson r = ?0.01; p = 0.95) (Fig. 4). Sediment delivery generally was underpredicted. Table 3. Summary of Spearman rank correlations of Universal Soil Loss Equation?based predictions and annual average sediment yields observed by the Smithsonian Environmental Research Center (SERC) and USGS. Model SERC observed sediment yield (n = 78) USGS observed sediment yield (n = 23) r p r p Erosion model? USLE ?0.02 0.69 ?0.22 0.32 RUSLE2 ?0.12 0.21 ?0.01 0.96 Sediment delivery model Vanoni, 1975 ?0.11 0.34 ?0.07 0.72 USDA Soil Conservation Service, 1986 ?0.10 0.36 ?0.02 0.94 Maner, 1958 ?0.08 0.50 ?0.28 0.20 Yagow, 1998 ?0.13 0.28 na? na Fraser, 1998 (RUSLE2) ?0.17 0.14 0.00 0.99 ? RUSLE, revised Universal Soil Loss Equation; SCS, Soil Conservation Service; USLE, Universal Soil Loss Equation. ? This correlation was not assessed. Fig. 3. Revised Universal Soil Loss Equation (RUSLE2)-based predicted erosion, RUSLE2 area-based sediment yield (SY) (Vanoni, 1975), and SEDMOD (Fraser et al., 1998) fl owpath-based SY versus observed SY in basins monitored by the Smithsonian Environmental Research Center (SERC) and USGS. Spearman rank correlation coeffi cients between the observed sediment yields and predicted estimates are indicated for each dataset. Boomer et al.: USLE-based Empirical Models Fail to Predict Sediment Discharges 85 Discussion Results from this study demonstrate the limitations of USLE- based predictions for whole catchments and reinforce previous arguments against using these models for watershed management (Kinnell, 2004a; Trimble and Crosson, 2000b). Th e USLE and RULSE2-based predictions, used in conjunction with SDRs, did not adequately predict observed sediment yield or the rankings of yields from whole catchments as measured in two independent datasets collected by SERC and by the USGS (Gellis et al., 2005; Jordan et al., 1997a; Jordan et al., 1997b; Langland et al., 1995). Instead, estimates from the SDR applications were strongly cor- related with the unmodi? ed USLE estimates of edge-of-? eld ero- sion. Th ese results demonstrate the inadequacy of extrapolating USLE erosion estimates to the whole catchment by incorporating current algorithms to account for hillslope or catchment trans- port processes. Th e lack of correlation of USLE-based predictions with measured sediment yields also suggests that integrated hy- drologic models that rely on USLE-SDR estimates as valid input or calibration data and treat the estimates as observed data (Chen and Mackay, 2004; Kinnell 2004a) may also be poor tools for predicting sediment yields. Such models include HSPF applica- tions (e.g., Donigian and Bicknell, 2006), GWLF (Haith and Shoemaker, 1987), and SWAT (Arnold and Allen, 1992). One of the di? culties in assuming that the plot-scale erosion model can be applied at the regional scale is that it is not possible to incorporate the plot-scale complexities and details prescribed for each of the USLE factors. For example, plot-based applications require users to assign cover management (C) values based on de- tailed characterizations of crop coverage (including crop rotation schedule, crop type, density of ? ne roots, density of ground cover, soil roughness, soil consolidation potential, and antecedent mois- ture conditions). Assigning plot-scale management factors (P) has similar complexities. In contrast, catchment-scale assignments of C and P factors are necessarily much less detailed. For example, as in other reported analyses, we aggregated land use classes into ? ve categories and did not distinguish di? erent forms of agriculture or di? erent housing densities (e.g., Ali and Sharda, 2005; Boellstor? and Benito, 2005), nor could we incorporate the e? ects of di? er- ent conservation practices (i.e., we set the P factor to one across the entire drainage basin). Using broad land cover classes disregards the signi? cant e? ects that spatial heterogeneity in agricultural practices Table 4. Pearson correlation coeffi cients between mean basin Universal Soil Loss Equation (USLE) input factors and mean basin USLE erosion estimates. USLE input factor USLE-based erosion estimate SERC basins (n = 78) USGS basins (n = 23) r p r p Topographic length-slope 0.66 <0.001 0.85 <0.001 Rainfall erosivity ?0.27 0.01 ?0.47 0.02 Surface soil erodibility ?0.04 0.71 ?0.40 0.05 Land use cover-practice ?0.43 <0.001 ?0.55 0.01 Table 5. Summary of univariate regression analyses relating log-transformed observed annual sediment yield (kg ha-1 yr-1) to catchment landscape features. Landscape metric/predictor? Chesapeake Bay watershed (n = 78) Appalachian Plateau (7) Appalachian Mountain (9) Mesozoic Lowland (7) Piedmont Upland (10) Coastal Plain (45) Area (ha) 0.01 0.16 0.16 0.25 0.00 0.02 R MJ mm ha?1 h?1 yr?1 0.00 0.00 0.84*** 0.08 0.02 0.04 USLE LS factor (dimensionless) 0.00 0.48 0.69** 0.04 0.01 0.11* K Mg ha h ha?1 MJ?1 mm?1 0.19*** na? 0.97*** na na 0.21** USLE CP factor (dimensionless) 0.09** na 0.59* 0.10 0.02 0.06 Annual A (Mg soil loss ha?1 yr?1) 0.00 0.46 0.82*** 0.09 0.01 0.15** Variation in terrain complexity 0.00 0.11 0.06 0.04 0.10 0.01 Basin topographic relief ratio 0.05 0.00 0.15 0.00 0.01 0.21** % Cropland 0.02 0.00 0.87*** 0.04 0.02 0.00 % Development 0.13*** 0.05 0.12 0.39 0.51* 0.15** % Impervious area 0.08** 0.17 0.86*** 0.10 0.08 0.06 % Forest 0.14*** 0.20 0.70** 0.03 0.01 0.15** Weighted CN 0.04 0.02 0.15 0.04 0.02 0.01 Annual runoff (cm yr?1) 0.18*** 0.00 0.62** 0.21 0.01 0.10* Mean stream fl ow (cm yr?1) 0.13*** 0.02 0.07 0.04 0.09 0.36*** * Signifi cant at the 0.05 probability level. ** Signifi cant at the 0.01 probability level. *** Signifi cant at the 0.001 level. ? A, erosion rate; CN, curve number; CP, cover-practice; K, soil erodibility; LS, length-slope; R, rainfall erosivity; USLE, Universal Soil Loss Equation. ? The relationship was not analyzed because there was no variation in the landscape characteristic throughout the physiographic province. Table 6. Percentages of variance explained for multiple regression models relating average annual sediment yield to geographic variables. Each percentage of variance explained is for a model including the term on the line and all terms on previous lines. Log(Sediment Yield in kg ha?1 yr?1) Factor Coeffi cient % Variance explained Constant ?0.09 Soil erodibility 30.71** 19 Log (stream fl ow in cm yr?1) 0.86** 32 Basin topographic relief ratio ?0.23* 42 Sqrt (% forest) ?0.68** 52 Appalachian Plateau Province 0.33* 55 * Signifi cant at the 0.05 probability level. ** Signifi cant at the 0.001 probability level. 86 Journal of Environmental Quality ? Volume 37 ? January?February 2008 and residential and livestock densities can have on nutrient load- ing and export (Johnes, 1996). In addition, land cover constants assigned for the USLE calculations ignore interactions with other geographic factors, such as variations in climatic patterns and plant growth along latitudinal gradients (Risse et al., 1993). Discrepan- cies in the model scale also a? ect the reliability of characterizing soil erodibility (K). At the plot scale, the K factor is determined from ? eld measurements of soil texture, structure, organic matter con- tent, and permeability (Wischmeier and Smith, 1978), but these properties become more variable and di? cult to parameterize with increasing scale (Zeleke and Si, 2005). Regional analyses rely on generalized parameter estimates from broad-scale soil maps such as STATSGO (USDA-NRCS, 1995). Additionally, the USLE-based sediment delivery models do not account for complexities at the landscape level that in? uence sediment transport and delivery. Predictions are based on plot-scale mechanisms, namely the e? ects of rainfall energy on soil detach- ment given the plot?s surface characteristics. When the scale is in- creased from a plot to a hillslope or catchment, additional processes such as overland ? ow and in? ltration hydraulics become increas- ingly important to controlling sediment transport and delivery (Slaymaker, 2006; Yoder et al., 2004). We tried to capture some of the landscape-level e? ects by implementing SDR algorithms that incorporate watershed ? ow path characteristics (e.g, Fraser et al., 1998; Yagow et al., 1998). Despite the strong potential for improv- ing predictive capability by accounting for ? ow path characteristics (Baker et al., 2006b), this approach did not improve the reliability of the USLE-based predictions. Because USLE estimates at the plot scale are well veri? ed (Risse et al., 1993), our results suggest that the e? ects of sediment transport processes are predominantly more important than the plot-scale erosion rates, which is counter to the assumptions underlying the USLE-SDR approach. Hydrologic processes that control sediment transport and deliv- ery at the catchment scale are at best crudely accounted for in vari- ous USLE-SDR models (Slaymaker, 2006). One of the key pro- cesses a? ecting sediment delivery is catchment runo? generation (Sheridan and Hubbard, 1987; Smith et al., 2005), suggesting the potential utility of combining USLE erosion estimates with runo? calculations, such as the CN method (e.g., Haith and Shoemaker, 1987). We tested whether incorporating annual average runo? estimates improved our ability to predict sediment discharge. We expected that runo? estimates would be an adequate proxy for the e? ects of catchment relief and development and that observed sediment yield would increase with estimated annual runo? . Th e logic and approach is similar to the MUSLE (Modi? ed USLE; Williams, 1975), which transforms the USLE into an event- based model by replacing the long-term annual rainfall erosivity with runo? estimates for each storm. Incorporating hydrologic characteristics at an annual timescale showed some promise, but results were variable, as demonstrated by the signi? cant correlation between estimated runo? and observed sediment yields from the SERC study watersheds and the lack of a similar correlation for the USGS watersheds. Th is inconsistency also may re? ect limitations of the CN method (Garen and Moore, 2005). Th e importance of hydrologic processes to sediment transport processes, which are driven by precipitation events, suggests that the timescale of the USLE model may be inappropriate for predict- ing catchment sediment discharge. Th e USLE predicts long-term (>20 yr) annual average soil erosion rates, a much longer time than the brief extreme events that often dominate sediment discharge events (e.g., Jordan et al., 1997a). Although the MUSLE is event based, our results indicate potential limitations with this approach because it assumes the landscape factors (i.e., soil erodibility, length-slope, cover, and management) remain relatively static on a year-to-year basis, representing an unchanging erosion potential. We found weak correlations between any of the USLE factors and the observed sediment yields. Results likely re? ect interactive e? ects of weather patterns on USLE input factors, resulting in erosion and sediment delivery rates that vary inconsistently with the amount and intensity of rainfall events (e.g., Lenzi et al., 2003; Kinnell, 2004b). Successful empirical models will require a focus on short-term precipitation data and more time relevant land use and cover data and incorporating nonlinear relationships between sediment production and rainfall erosivity. Perhaps the most signi? cant limitation of using USLE-based watershed models is that they do not account for sediment gen- eration by processes other than the e? ects of overland ? ow on interrill and rill erosion. In particular, gully erosion (Prosser et al., 2001) and in-stream processes, especially bank erosion and resuspension of sediment materials, can contribute signi? cantly to observed sediment loads (Trimble, 1997). Stream bank ero- Fig. 4. Observed annual average sediment yield (SY in kg ha?1 yr?1) versus sediment yield predicted by linear regression on log- or square-root transformed landscape factors. Plot A depicts calibration with SERC data; Plot B depicts attempted validation with USGS data. Points along the diagonal 1:1 line indicate perfect agreement between predictions and measurements. Symbols indicate the physiographic province in which a catchment is located: Appalachian Plateau (APPL); Appalachian Mountain (APMN); Mesozoic Lowland (MELO); Piedmont Upland (PDUP); or Coastal Plain (CP). Boomer et al.: USLE-based Empirical Models Fail to Predict Sediment Discharges 87 sion could be predominantly important to the sediment budget where strati? ed and unconsolidated deposits enhance the poten- tial for stream incisement (Campo and Desloges, 1994; Nagle et al., 2007), as in the Coastal Plain physiographic province of the Chesapeake Bay watershed (Markewich et al., 1990). Hu- man alterations to stream networks, such as the construction and subsequent failure of mill dams (Downward and Skinner, 2005) or enhanced peak ? ow due to land use change (Po? et al., 2006), also can enhance the impact of in-stream processes. Disparities between the predicted and observed sediment yields might arise from data and computational limitations, particularly measurement error and model implementation. For example, signi? cant di? erences in the LS factor can arise from variation in spatial analysis methods, including selection of the input data, vari- able de? nition, and implementation of GIS routines. User-selected GIS procedures also can in? uence predictions signi? cantly. For example, simple USLE calculations summarize LS values across an entire catchment, whereas more sophisticated applications incor- porate ? ow direction and exclude areas with steep slope declines where deposition is more likely than erosion (Hickey et al., 1994). Th e choice of GIS algorithms used for slope calculations (Dunn and Hickey, 1998), drainage basin delineations (Baker et al., 2006a), and ? ow routing (Desmet and Govers, 1996) also might alter the results. Th e strong correspondence between predictions from the automated RUSLE2 (USEPA, 2004) and our own USLE calculations (Fig. 2), however, suggests that computational di? er- ences in the implementation may have minor e? ects, so that the failure to correctly predict catchment discharges is more likely due to the more fundamental conceptual problems. Because the USLE-based models did not work well for whole catchments, we explored the utility of the input factors, together with additional landscape features, to improve sediment yield predictions in a multiple regression model. Soil erodibility (K), stream ? ow, topographic relief ratio, percent forest cover, and phys- iographic province together explained over half of the variability in sediment yield among SERC watersheds, but a large portion of the variance (45%) remained unexplained. Previous empirical studies, which included catchment land cover and physical characteristics as potential predictors, have reported similarly signi? cant ? ndings, but the importance of landscape versus physiographic features was inconsistent among the e? orts (Table 7). For example, some studies identi? ed land cover features as the most signi? cant factors (Jones et al., 2001; Basnyat et al., 1999), whereas others identi? ed physical catchment features, such as physiographic province or drainage area, as more important (Verstraeten and Poesen, 2001; Restrepo et al., 2006). We also found that an empirical model calibrated with sediment yield data from the SERC watersheds did not perform well in a validation test with yield data from the USGS watersheds. Th is suggests that it may be dangerous to rely on other empirical models completed without the bene? t of valida- tion with an independent dataset. Variation among the empirical models (Table 7) could re? ect the di? culty of using annual average observations to model elevated sediment loads, which occur mainly in response to short-term weather events (Jordan et al., 1997a). In addition, static models like these do not capture dynamic interac- tions among the input factors, which change in response to short- term and interannual weather ? uctuations (Lenzi et al., 2003). Th ese trends collectively suggest that scientists and managers have not captured the linkages between the catchment landscape setting and the physical mechanisms that regulate erosion and sediment transport processes. Conclusions We implemented seven variations of USLE-based models to estimate erosion and sediment delivery, but none provided a reliable tool for assessing sediment discharge from 101 catchments where stream water quality was monitored con- tinuously for at least 1 yr. Our results reinforce previous argu- ments that USLE-based sediment delivery models provide an inadequate framework for managing land and water resources Table 7. Summary of published statistical models relating observed stream sediment yield (SY) or total suspended solids (TSS) to catchment landscape characteristics. Model R2 (%) This study (78 catchments in the Chesapeake Bay watershed, USA): log(SY in kg ha?1 yr?1) = 30.71 (mean soil erodibility)? + 0.86 log[stream fl ow in cm yr?1]) ? 0.23 (basin relief ratio) ? 0.68 (sqrt[% forest area]) + 0.33(Appalachian Plateau) 55 (Restrepo et al., 2006) (32 catchments in the Magdalena River watershed, Colombian Andes in South America): log(SY in Mg km?2 yr?1) = ?0.8838 + 0.8140 log(runoff in mm yr?1) ? 0.3906 log(average maximum discharge in m3 s?1)? 58 (Weller et al., 2003) (23 catchments in the Patuxent River watershed, MD): TSS (mg L?1) = 1.0(%cropland) + 0.6(% development) + 0.7 (Coastal Plain) + 11.5(Week)? + 17.2(% cropland ? week) + 7.8 (% development ? week) + 11.8 (Coastal Plain ? week) + 6.9(% cropland ? Coastal Plain ? week) 58 (Verstraeten and Poesen, 2001) (26 catchments in Central Belgium): log(SY in Mg ha?1 yr?1) = 3.72 ? 0.72 log (area in ha) ? 0.84 log (hypsometric integral)? + 0.11 log (drainage length in m) 76 (Jones et al., 2001) (17 catchments in the Chesapeake Bay watershed): log(SY in kg ha?1 yr?1) = 8.472 + 0.079(% development) ? 0.116(% wetlands) ? 0.038(riparian forest)# 79 (Basnyat et al., 1999) (21 catchments in the Fish River watershed, Al, USA): log(TSS in mg L?1) = 3.7(% forest) + 17.33(% development) ? 20.7(% orchard) + 11.47(% cropland) + 17.66(% pasture) 76 ? Universal Soil Loss Equation parameter (Mg hr MJ-1 mm-1). ? Long-term (1975?1995) average maximum water discharge. ? Because of the week factor, model is specifi c to the time period monitored (August 1997?August 1999). ? The diff erence in the mean and minimum catchment elevations relative to the diff erence in the maximum and minimum catchment elevations. # Percent of watershed with forest land cover adjacent to stream edge, defi ned by land cover in adjacent 30-m pixels. 88 Journal of Environmental Quality ? Volume 37 ? January?February 2008 at the catchment scale (Kinnell, 2004a; Trimble and Crosson, 2000a, b). Th e USLE was not intended to predict e? ects on stream water quality, yet the models continue to be widely applied at the catchment scale by scientists (e.g., Boellstor? and Benito, 2005; Fu et al., 2005; Kim et al., 2005; Onyando et al., 2005; Wang et al., 2005), policymakers (e.g., Donigian and Bicknell, 2006; USEPA, 2005), and watershed modelers. Our review of published statistical models and the poor perfor- mance of our own empirical model in a validation attempt with independent sediment yield data also suggest that many other non-USLE empirical models developed to predict annual sediment yield (Table 7) may be unreliable. First, a comparison of published statistical models revealed contradictions in the attribution of sedi- ment delivery to land cover versus physiographic factors. Second, the disappointing performance of our model in the validation with independent data highlights the danger of relying on empirical models that have not been tested with a validation dataset. Our ? ndings also suggest some directions for future research on predicting sediment discharge in ungauged drainage basins: (i) Identify potential predictor variables that conceptually link landscape and stream characteristics to ? ow velocity, stream power, and the ability to transport sediment; (ii) incorporate metrics to indicate potential sediment sources within streams, including bank erosion and legacy sediments; and (iii) develop predictions for temporal scales ? ner than the long-term annual average time frame. Consistent and veri? able results from addi- tional empirical studies will also help reduce the uncertainty in the predictions of process-based, integrated simulation models. Acknowledgments Th is research funded by the National Oceanic and Atmospheric Administration Coastal Oceans Program (grant numbers NA66RG0129 and NA03NOS4780008), National Science Foundation (grant numbers BSR-9085219 and DEB-9317968), and the Smithsonian Institution Environmental Sciences Program. Matthew Baker, Kevin Sigwart, and Kathryn Sullivan provided valuable assistance with the analysis of watershed geography. References Ali, S., and V.N. Sharda. 2005. Evaluation of the Universal Soil Loss Equation (USLE) in semi-arid and sub-humid climates of India. Appl. Eng. Agric. 21:217?225. Angima, S.D., D.E. Stott, M.K. O?Neill, C.K. Ong, and G.A. Weesies. 2003. Soil erosion prediction using RUSLE for central Kenyan highland conditions. Agric. Ecosyst. Environ. 97:295?308. Arnold, J.G., and P.M. Allen. 1992. A comprehensive surface-groundwater ? ow model. J. Hydrol. 142:47?69. Baker, M.E., D.E. Weller, and T.E. Jordan. 2006a. Comparison of automated watershed delineations: E? ects on land cover areas, percentages, and relationships to nutrient discharge. Photogramm. Eng. Remote Sens. 72:159?168. Baker, M.E., D.E. Weller, and T.E. Jordan. 2006b. Improved methods for quantifying potential nutrient interception by riparian bu? ers. Landscape Ecol. 21:1327?1345. Basnyat, P., L.D. Teeter, K.M. Flynn, and B.G. Lockaby. 1999. Relationships between landscape characteristics and nonpoint source pollution inputs to coastal estuaries. Environ. Manage. 23:539?549. Bicknell, B.R., J.C. Imho? , J.L. Kittle, A.S. Donigian, Jr., and R.C. Johanson. 1993. Hydrologic Simulation Program- FORTRAN (HSPF): User?s manual for release 10 EPA/600/R-93/174. USEPA Environmental Research Lab., Athens, GA. Boellstor? , D., and G. Benito. 2005. Impacts of set-aside policy on the risk of soil erosion in central Spain. Agric. Ecosyst. Environ. 107:231?243. Boesch, D.F., R.B. Brins? eld, and R.E. Magnien. 2001. Chesapeake Bay eutrophication: Scienti? c understanding, ecosystem restoration, and challenges for agriculture. J. Environ. Qual. 30:303?320. Campo, S.H., and J.R. Desloges. 1994. Sediment yield conditioned by glaciation in a rural agricultural basin of southern Ontario, Canada. Phys. Geogr. 15:495?515. Carpenter, S.R. 1996. Microcosm experiments have limited relevance for community and ecosystem ecology. Ecology 77:677?680. Chen, E., and D.S. Mackay. 2004. E? ects of distribution-based parameter aggregation on a spatially distributed agricultural nonpoint source pollution model. J. Hydrol. 295:211?224. Cohn, T.S., D.L. Caulder, E.J. Gilroy, L.D. Zynjuk, and R.M. Summers. 1992. Th e validity of a simple log-linear model for estimating ? uvial constituent loads: An empirical study involving nutrient loads entering Chesapeake Bay. Water Resour. Res. 28:2353?2364. Desmet, P.J.J., and G. Govers. 1996. Comparison of routing algorithms for digital elevation models and their implications for predicting ephemeral gullies. Int. J. Geogr. Inf. Syst. 10:311?331. Donigian, A.S., Jr., and B.R. Bicknell. 2006. Sediment parameter and calibration guidance for HSPF. BASINS Tech. Note 8. USEPA, Washington, DC. Downward, S., and K. Skinner. 2005. Working rivers: Th e geomorphological legacy of English freshwater mills. Area 37:138?147. Dunn, M., and R.A. Hickey. 1998. Th e e? ect of slope algorithms on slope estimates within a GIS. Cartography 27:9?15. EROS Data Center. 1999. National Elevation Dataset. USGS, Washington, DC. Available at http://ned.usgs.gov/ (veri? ed 7 Aug. 2007). Ferro, V., and P. Porto. 2000. Sediment delivery distributed (SEDD) model. J. Hydrol. Eng. 5:411?422. Foster, G.R., T.E. Toy, and K.G. Renard. 2003. Comparison of the USLE, RUSLE1.06c, and RUSLE2 for application to highly disturbed lands. Proc. 1st Interagency Conf. on Research in the Watersheds. K.G. Renard et al. (ed.) 27?30 Oct., Benson, AZ. p. 154?160. Fraser, R.H., P.K. Barten, and D.A.K. Pinney. 1998. Predicting stream pathogen loading from livestock using a geographical information system-based delivery model. J. Environ. Qual. 27:935?945. Fu, B.J., W.W. Zhao, L.D. Chen, Q.J. Zhang, Y.H. Lu, H. Gelinck, and J. Poesen. 2005. Assessment of soil erosion at large watershed scale using RUSLE and GIS: A case study in the Loess Plateau of China. Land Degrad. Dev. 16:73?85. Garen, D.C., and D.S. Moore. 2005. Curve number hydrology in water quality modeling: Uses, abuses, and future directions. J. Am. Water Resour. Assoc. 41:377?388. Gellis, A.C., W.S.L. Banks, M.J. Langland, and S.K. Martucci. 2005. Summary of suspended-sediment data for streams draining the Chesapeake Bay watershed, water years 1952?2002. Scienti? c Investigations Rep. 2004- 5056. U.S. Dep. of the Interior, USGS, Reston, VA. Goetz, S.J., R.K. Wright, A.J. Smith, E. Zinecker, and E. Schaub. 2003. Ikonos imagery for resource management: Tree cover, impervious surfaces, and riparian bu? er analyses in the mid-Atlantic region. Remote Sens. Environ. 88:195?208. Haith, D.A., and L.L. Shoemaker. 1987. Generalized watershed loading functions for stream nutrients. Water Resour. Bull. 23:471?478. Hickey, R.A., A. Smith, and P. Jankowski. 1994. Slope length calculations from a DEM within ARC/INFO GRID. Comput. Environ. Urban Syst. 18:365?380. Johnes, P.J. 1996. Evaluation and management of the impact of land use change on the nitrogen and phosphorus load delivered to surface waters: Th e export coe? cient modelling approach. J. Hydrol. 183:323?349. Jones, K.B., A.C. Neale, M.S. Nash, R.D.V. Remortel, J.D. Wickham, K.H. Riitters, and R.V. O?Neill. 2001. Predicting nutrient and sediment loadings to streams from landscape metrics: A multiple watershed study from the United States Mid-Atlantic Region. Landscape Ecol. 16:301?312. Jordan, T.E., D.L. Correll, and D.E. Weller. 1997a. Relating nutrient discharges from watersheds to landuse and stream? ow variability. Water Resour. Res. 33:2579?2590. Jordan, T.E., D.L. Correll, and D.E. Weller. 1997b. Nonpoint source discharges of nutrients from Piedmont watersheds of Chesapeake Bay. J. Am. Water Resour. Assoc. 33:631?645. Boomer et al.: USLE-based Empirical Models Fail to Predict Sediment Discharges 89 Kemp, M.W., W.R. Boynton, J.E. Adolf, D.F. Boesch, W.C. Boicourt, G.S. Brush, J.C. Cornwell, T.R. Fisher, P.M. Glibert, J.D. Hagy, L.W. Harding, E.D. Houde, D.G. Kimmel, W.D. Miller, R.I.E. Newell, M.R. Roman, E.M. Smith, and J.C. Stevenson. 2005. Eutrophication of Chesapeake Bay: Historical trends and ecological interactions. Mar. Ecol. Prog. Ser. 303:1?29. Kim, J.B., P. Saunders, and J.T. Finn. 2005. Rapid assessment of soil erosion in the Rio Lempa Basin, Central America, using the Universal Soil Loss Equation and geographic information systems. Environ. Manage. 36:872?885. Kinnell, P.I.A. 2004a. Sediment delivery ratios: A misaligned approach to determining sediment delivery from hillslopes. Hydrol. Processes 18:3191?3194. Kinnell, P.I.A. 2004b. Th e mathematical integrity of some Universal Soil Loss Equation variants. Soil Sci. Soc. Am. J. 68:336?337. Langland, M.J., P.L. Lietman, and S. Ho? man. 1995. Synthesis of nutrient and sediment data for watersheds within the Chesapeake Bay drainage basin. USGS, Lemoyne, PA. Lenzi, M.A., L. Mao, and F. Comiti. 2003. Interannual variation of suspended sediment load and sediment yield in an alpine catchment. Hydrol. Sci. J. 48:899?915. Liu, Z.-J., D.E. Weller, D.L. Correll, and T.E. Jordan. 2000. E? ects of land cover and geology on stream chemistry in watersheds of Chesapeake Bay. J. Am. Water Resour. Assoc. 36:1349?1365. Lu, D., G. Li, G.S. Valladares, and M. Batistella. 2004. Mapping soil erosion risk in Rondonia, Brazilian Amazonia: Using RULSE, remote sensing and GIS. Land Degrad. Dev. 15:499?512. Maner, S.B. 1958. Factors a? ecting sediment delivery rates in the Red Hills physiographic area. Trans. Am. Geophys. 39:669?675. Markewich, H.W., M.J. Pavich, and G.R. Buell. 1990. Contrasting soils and landscapes of the Piedmont and Coastal Plain, eastern United States. Geomorphology 3:417?447. Martin, A., J.T. Gunter, and J.L. Regens. 2003. Estimating erosion in a riverine watershed?Bayou Liberty-Tchefuncta River in Louisiana. Environ. Sci. Pollut. Res. 10:245?250. Nagle, G.N., T.J. Fahey, J.C. Ritchie, and P.B. Woodbury. 2007. Variations in sediment sources and yields in the Finger Lakes and Catskills regions of New York. Hydrol. Processes 21:828?838. Nearing, M.A., M.J.M. Romkens, L.D. Norton, D.E. Stott, F.E. Rhoton, J.M. La? en, D.C. Flanagan, C.V. Alonso, R.L. Binger, S.M. Dabney, O.C. Doering, C.H. Huang, K.C. McGregor, and A. Simon. 2000. Measurements and models of soil loss rates. Science 290:1300?1301. Onyando, J.O., P. Kisoyan, and M.C. Chemelil. 2005. Estimation of potential soil erosion for River Perkerra catchment in Kenya. Water Resour. Manage. 19:133?143. Po? , N.L., B.P. Bledsoe, and C.O. Cuhaciyan. 2006. Hydrologic variation with land use across the contiguous United States: Geomorphic and ecological consequences for stream ecosystems. Geomorphology 79:264?285. Prosser, I.P., I.D. Rutherfurd, J.M. Olley, W.J. Young, P.J. Wallbrink, and C.J. Moran. 2001. Large-scale patterns of erosion and sediment transport in river networks, with examples from Australia. Mar. Freshw. Res. 52:81?99. Renard, K.G., G.R. Foster, G.A. Weesies, K.K. McCool, and D.C. Yoder. 1997. Predicting soil erosion by water: A guide to conservation planning with the revised Universal Soil Loss Equation (RUSLE) Agriculture Handbook No. 703. USDA, Washington, DC. Restrepo, J.D., B. Kjerfve, M. Hermelin, and J.C. Restrepo. 2006. Factors controlling sediment yield in a major South American drainage basin: Th e Magdalena River, Colombia. J. Hydrol. 316:213?232. Risse, L.M., M.A. Nearing, A.D. Nicks, and J.M. La? en. 1993. Error assessment in the Universal Soil Loss Equation. Soil Sci. Soc. Am. J. 57:825?833. Sheridan, J.M., and R.K. Hubbard. 1987. Transport of solids in stream? ow from Coastal Plain watersheds. J. Environ. Qual. 16:131?136. Slaymaker, O. 2006. Towards the identi? cation of scaling relations in drainage basin sediment budgets. Geomorphology 80:8?19. Smith, J.A., M.L. Baeck, K.L. Meierdiercks, P.A. Nelson, A.J. Miller, and E.J. Holland. 2005. Field studies of the storm event hydrologic response in an urbanizing watershed. Water Resour. Res. 41:W10413 doi:10.1029/ 2004WR003712. Spatial Climate Analysis Service. 2002. 2.5-acrc minute 1971?2000 mean monthly precipitation grids for the conterminous United States. Oregon State University, Corvallis, OR. Available at http://prism. oregonstate.edu/ (veri? ed 7 Aug. 2007). Stankelis, R.M., M.D. Naylor, and W.R. Boynton. 2003. Submerged aquatic vegetation in the mesohaline region of the Patuxent estuary: Past, present, and future status. Estuaries 26:186?195. Trimble, S.W. 1997. Contribution of stream channel erosion to sediment yield from an urbanizing watershed. Science 278:1442?1444. Trimble, S.W., and P. Crosson. 2000a. Measurements and models of soil loss rates?Response. Science 290:1301. Trimble, S.W., and P. Crosson. 2000b. U.S. soil erosion rates?Myth and reality. Science 289:248?250. USDA-NRCS. 1983. Sediment sources, yields, and delivery ratios. National Engineering Handbook, Section 3, Sedimentation. U.S. Gov. Print. O? ce, Washington, DC. USDA-NRCS. 1995. State soil geographic (STATSGO) data base for New York, Pennsylvania, Maryland, Delaware, and Virginia. USDA, Washington, DC. Available at http://www.ncgc.nrcs.usda.gov/products/ datasets/statsgo/data/index.html (veri? ed 7 Aug. 2007). USDA Soil Conservation Service. 1986. Urban hydrology for small watersheds. NTIS #PB87101580. Tech. Release 55. USDA, Washington, DC. USEPA. 2004. Computation of soil and landform metrics: Programs and U.S. Geodatasets. Release Version 1.1. USEPA, Las Vegas, NV. USEPA. 2005. Little Juniata River Watershed, Blair County (PA). USEPA, Washington, DC. Available at http://www.epa.gov/reg3wapd/tmdl/pa_ tmdl/LittleJuniata/LittleJuniataRiverDR.pdf (veri? ed 7 Aug. 2007). Van Remortal, R.D., M.E. Hamilton, and R.A. Hickey. 2001. Estimating the LS factor for RUSLE through iterative slope length processing of digital elevation data. Cartography 30:27?35. Van Rompaey, A.J.J., V. Vielillefont, R.J.A. Jones, L. Montanarella, G. Verstraeten, P. Bazzo? , T. Dostal, J. Krasa, J. de Vente, and J. Poesen. 2003. Validation of soil erosion estimates at European scale. European Soil Bureau Research Report No.13, EUR 20827 EN, O? ce for O? cial Publications of the European Communities, Luxembourg. Vanoni, V.A. 1975. Sedimentation engineering 54. ASCE manuals and reports on engineering practices. ASCE, Reston, VA. Verstraeten, G., and J. Poesen. 2001. Factors controlling sediment yield from small intensively cultivated catchments in a temperate humid climate. Geomorphology 40:123?144. Vogelmann, J.E., S.M. Howard, L. Yang, C.R. Larson, B.K. Wylie, and N. Van Driel. 2001. Completion of the 1990?s National Land Cover Data set for the conterminous United States from Landsat Th ematic Mapper data and ancillary data sources. Photogramm. Eng. Remote Sens. 67:650?662. Wang, X.D., X.H. Zhong, and J.R. Fan. 2005. Spatial distribution of soil erosion sensitivity on the Tibet Plateau. Pedosphere 15:465?472. Weller, D.E., T.E. Jordan, D.L. Correll, and Z.J. Liu. 2003. E? ects of land- use change on nutrient discharges from the Patuxent River watershed. Estuaries 26:244?266. Williams, J.R. 1975. Sediment-yield prediction with universal equation using runo? energy factor. p. 244?252. In Present and prospective technology for predicting sediment yield and sources. ARS.S-40, U.S. Gov. Print. O? ce, Washington, DC. Williams, J.R., and H.D. Berndt. 1972. Sediment yield computed with universal equation. J. Hydraul. Div. Proc. Am. Soc. Civil Eng. 98:2087?2098. Wischmeier, W.H., and D.D. Smith. 1978. Predicting rainfall-erosion losses: A guide to conservation planning. Agriculture Handbook #507. USDA, Washington, DC. Yagow, E.R., V.O. Schanholtz, B.A. Julian, and J.M. Flagg. 1998. A water quality module for CAMPS. American Society of Agricultural Engineers Meeting Presentation Paper No. 88-2653. ASAE, St. Joseph, MI. Yoder, D.C., G.R. Foster, G.A. Weesies, K.G. Renard, K.K. McCool, and J.B. Lown. 2004. Evaluation of the RUSLE soil erosion model. p. 107?116. In J.E. Parsons et al. (ed.) Agricultural non-point source water quality models: Th eir use and application. Southern Cooperative Series Bull. No. 398. Young, R.A., C.A. Onstad, D.D. Bosch, and W.P. Anderson. 1989. AGNPS: A nonpoint-source pollution model for evaluating agricultural watersheds. J. Soil Water Conserv. 44:4522?4561. Zeleke, T.B., and B.C. Si. 2005. Scaling relationships between saturated hydraulic conductivity and soil physical properties. Soil Sci. Soc. Am. J. 69:1691?1702.