The Impact of Dams on Flood Flows in the United States

Natural flood regimes provide a wide array of important ecological functions. Our goal is to assess the hydrologic impact of dams on flood flows throughout the United States. Regional regression models of the median annual 1-day maximum flow were developed as a function of natural watershed characteristics, dam storage, and population density. Most of the regressions have adjusted R2 values in excess of 0.80, and overall the models covered 78% of the area of the continental U.S. Alteration of flood flows is present in every region of the country, and is more severe west of the Mississippi and especially in the southern Great Plains, desert Southwest, and northern California. The percent of U.S. rivers with greater than a 25% reduction in the median annual flood is 55% for large rivers, 25% for medium rivers, and 10% for small rivers. The majority of freshwater ecoregions in the country have at least 10% of their rivers with 25% or greater alteration in all three river size classes. A simple model based on the ratio of dam storage to mean annual runoff was developed for assessing alteration in ungauged rivers, and was found to be generally useful for classifying rivers into categories of potential alteration. Overall, we document the alteration of natural flood flows across the U.S. in more detail than has been previously accomplished, and demonstrate the efficacy of multivariate regional regression models and other indicators for assessing hydrologic alteration. Copyright # 2010 John Wiley & Sons, Ltd. key words: dams; hydrologic alteration; environmental flows; floods; flood control; hydrology; ecological flows Received 18 October 2009; Revised 28 February 2010; Accepted 20 April 2010 INTRODUCTION AND BACKGROUND Introduction Natural flood regimes provide a wide array of ecological functions that are essential for the health of river, floodplain, riparian and estuarine ecosystems, as has been detailed in the literature (Junk et al., 1989; Bayley, 1995; Poff et al., 1997; Alber, 2002; Lytle and Poff, 2004; Mathews and Richter, 2007; Piazza and La Peyre, 2007). Ecological benefits of floods include providing fish and other organisms with access to floodplain habitats that can be used for feeding, spawning and rearing; maintaining and rejuvenating plant habitats in the riparian zone and floodplain; influencing the geomorphology of the streambed; importing woody debris and organic material into the river channel; refreshing water quality conditions and helping transfer nutrients and maintain salinity conditions in estuaries. High flows just below flood stage (i.e. below bankfull stage) move sediment through the channel, provide respite for organisms from stressful low-flow conditions and improve connectivity to upstream and downstream habitats. Conversely, alteration of natural flood events can have serious consequences for ecosystem health. The typical impact of dams is to reduce the magnitude of peak flood flow *Correspondence to: T. W. FitzHugh, FitzHugh Consulting, 717 NW 70th St. #103, Seattle, WA 98117, USA. E-mail: [email protected] Copyright # 2010 John Wiley & Sons, Ltd. magnitudes, quite often dramatically (Richter et al., 1998; Magilligan and Nislow, 2005; Graf, 2006), which degrades or eliminates many of the important functions described above. Reduction of flood flows in river systems can alter ecological communities and facilitate invasions by non-native species (Poff et al., 1997), and lead to a variety of negative geomorphological consequences (Magilligan et al., 2003). Given the importance of floodplain and estuarine ecosystems from the perspective of species richness, productivity and provisioning of ecosystem services (Costanza et al., 1997; Tockner and Stanford, 2002), assessing the degree and extent of alteration of flood flows in the United States and elsewhere is an important research question that has bearing on a range of environmental and water management issues. The goal of this paper is to assess the impact that existing dams have had on peak flood flows throughout the United States, in as comprehensive a fashion as is possible given available data. Previous sub-national studies have reported on the impact of dams on natural flow regimes (including flood flows) in the Colorado River basin (Richter et al., 1998), the Connecticut River basin (Magilligan and Nislow, 2001), the state of Texas (Asquith, 2001) and the Wabash River basin in Indiana (Pyron and Neumann, 2008). Magilligan and Nislow (2005), Graf (2006) and Poff et al. (2006) analysed the impacts of dams on flows for a subset of rivers across the country (21, 36 and 43, respectively). Till date, the most comprehensive study of hydrologic alteration by dams T. W. FITZHUGH AND R. M. VOGEL was by Poff et al. (2007), who analysed the impacts of dams on intermediate size (3rd–7th order) rivers across the United States, using streamflow data for 186 stations below dams and 317 stations on undammed rivers. Similarly, Gao et al. (2009) examined several indicators for their ability to reflect changes in overall hydrologic alteration for 189 rivers with dams across the United States. These two studies covered the majority of the United States, using more streamflow data than in previous national evaluations, but still did not use all available streamflow data, for reasons described below. The hydrologic impacts of dams are typically analysed by comparing various streamflow statistics from periods before and after the dam was constructed. An important constraint on applying this method for a national assessment of the alteration of flood flows is the availability of reference data on natural flows before dam construction. Typically 20 years of preand post-dam data are recommended in order to be able to reliably detect shifts in high flow statistics (Richter et al., 1997; Huh et al., 2005). These requirements make the number of stations available for use with such a standard preversus postanalysis necessarily limited. For example, of the 4859 gaging stations for which data were used in this study, only 564 had 20 years of data both before and after construction of a dam or dams. But an additional 1808 stations had at least 20 years of data after construction of upstream dams, without sufficient pre-dam data. Another concern with the preand post-method of analysis is the possibility that climate is shifting in the United States in ways that affect flood flows, as has been suggested by Hodgkins et al. (2003) and Stewart et al. (2005). Thus, instead of assuming stationarity of the flow records, we employ a method that explicitly takes into account temporal changes in both climatic and land-use factors. We employ regional multivariate regression methods to assess impacts of dams and other factors on the behaviour of flood flows. The idea is to construct regional multivariate regression models that predict flood flows as a function of climatic, physiographic and anthropogenic characteristics of the watershed contributing to each gaging station. The US Geological Survey (USGS) has a long and rich history of developing such multivariate regression models for predicting both peak flow and low flow statistics at ungauged sites across the United States and a computer program is even available for the application of the resulting models at ungauged sites (Turnipseed and Ries, 2007). Such regional statistical models have also been developed for predicting annual average streamflows (Vogel et al., 1999),and low flow statistics (Kroll et al., 2004) across the United States, and for a variety of streamflow statistics in Washington, Colorado and Oregon (Sanborn and Bledsoe, 2006). Thus the method employed here has been well tested and vetted in the literature and in practice, and can be applied to large regions by generating data on watershed characteristics using standard GIS methods. Copyright # 2010 John Wiley & Sons, Ltd. The streamflow statistic that will be analysed here is the median annual 1-day maximum flow for each decade in the 1900s, which we term as the median annual flood (MAF). We employ a nonparametric estimator of the MAF, which does not depend on the assumption of a frequency distribution. Since the MAF has a 50% chance of being exceeded in any year, it has an average return period of 2-years. This statistic is attractive from a geomorphological perspective, because in natural stream channels, the discharge necessary to reach bankfull flow occurs, on average, with a 2 year recurrence probability (Leopold et al., 1964; Magilligan et al., 2003). Magilligan et al. (2003) states that ‘the bankfull discharge has also been shown to be the dominant discharge for sediment transport and channel maintenance’, and it ‘also sets other geomorphic and ecological thresholds, because floods that exceed this discharge are capable of inundating the adjacent river floodplain’. Hence, the flow statistic considered here is closely related to bankfull discharge and has a number of critical geomorphological and ecological functions. Our primary goal is to develop regression models for hydrologic units across the United States that relate the decadal MAF to watershed characteristics. The regression models are then used to discern the impacts of dams on flood flows across the country. Statistically significant impacts are summarized by river size and according to the freshwater ecoregions developed by Abell et al. (2008). The models and analysis presented here should provide the most comprehensive picture to date of the wide extent of dam impacts on flood flows in the United States, and will also highlight the potential for restoration of flood flows that exists in many parts of the country. This study will also test the efficacy of multivariate regression modelling for assessing the significance and degree of hydrologic alteration, an approach that to our knowledge has received little attention. Use of regional regression models to evaluate influence of dam storage on flood flows While the regression approach used here is standard in many ways, there are also some important differences from previous studies. Typically regional regression models are developed using period of record flow statistics which assume a stationary historical period. Since our goal is to model changes in flood flows due to the impact of dams during the 20th century, we examine flood data by decade. Decades are used because on the one hand they allow for assessment of trends over time, yet they also average out stochastic year-to-year variability that would otherwise be difficult to account for. Watershed characteristics that change over time, such as climate, land use and dam storage, are also calculated by decade, enabling the regressions to quantify the impacts of these different factors on the MAF. Another important difference from earlier regional regression studies is that most previous studies focused on River Res. Applic. (2010) DOI: 10.1002/rra IMPACT OF DAMS ON FLOOD FLOWS reference streamflow gaging stations, i.e. those stations that are mostly free of anthropogenic influences, so that streamflow measured at these sites is primarily influenced by natural factors. Instead we use all available streamflow gaging stations, whether impacted or not. To account for anthropogenic influences on flood flows, we included dam storage and watershed population density as potential independent variables in the regression, also computed by decade. Population density provides a surrogate measure of the influence of land development and is often highly correlated with residential impervious area. Other than dam storage and population density, there were no other variables in the regressions to represent anthropogenic impacts. While we recognize that there are other potential anthropogenic impacts on flood flows, such as land-cover changes other than impervious surfaces, and water withdrawals, it was not possible to consider the impacts of these variables in the regressions since there are no datasets representing the historical evolution of these variables during the 20th century. The use of multivariate regional regression methods provides a number of important advantages over alternative approaches for testing hypotheses. Most importantly, the analysis ‘replaces time with space’. That is, by incorporating many flow gaging stations in space, we effectively increase the sample size of the regression equations. Alternatively, each hypothesis test would only be on a single flow record, over perhaps two different periods of time (i.e. altered and unaltered). By integrating all stations within a region, the analysis effectively increases the sample size available by replacing limitations on the temporal extent of data at a single site with the fact that many sites are considered, in space, thus ‘replacing time with space’. A second advantage of the multivariate statistical approach is that it does not require that one specify beforehand that a particular station is or is not impacted by human activities, since the multivariate analysis adjusts for differences in flow that are related to anthropogenic factors. A third advantage is that a typical preand post-data analysis is difficult to implement in cases where dam storage has increased gradually on a river, due to construction of multiple dams over time, yet the regression method is well equipped to handle such situations. Lastly, because climatic data are in the regressions, the regressions will adjust for temporal climatic change across decades so that such climatic trends can be taken into account when assessing the impacts of dam storage on flood flows. Limitations of approach There are numerous concerns and caveats regarding the resulting regression equations. Regressions yield average impacts of dams on the MAF across a given region, thus they may be less precise in computing impacts at a particular location than the standard preand post-data analysis Copyright # 2010 John Wiley & Sons, Ltd. methods. While our use of regional regressions yields a more comprehensive picture of the impacts of dams on flood flows than alternate methods, it comes at the expense of losing some specificity about the impacts at a particular location. Partly for this reason, the regression results are only presented as averages for the hydrologic units for which the regressions were produced. As with all regression methods, it would be dangerous to extrapolate the results of our models, thus they should only be used within the regions and for the sites considered in our analyses. DATA AND METHODOLOGY Due to space limitations the data and methods used in this study are briefly summarized here, further details can be found in a separate report (FitzHugh and Vogel, 2010) available on the internet. Databases Decadal values for the MAF were obtained from daily streamflow data for 4859 USGS streamflow stations across the United States, using the Indicators of Hydrologic Alteration (IHA) software (Richter et al., 1996; Mathews and Richter, 2007). The stations used here had to satisfy one of two criteria: (1) they had data for the most recent available decade (the 1990s) and at least one earlier decade; or (2) they were reference stations that had data for two decades or more from the 1900s to the 1980s. Reference stations used in this study are those stations identified in Slack and Landwehr (1992), Poff (1996) and Carlisle et al. (2009). The data for these stations yielded 23 228 individual decadal values of the MAF. GIS analysis was used to compute a series of watershed characteristics to use as potential independent variables in the regressions (see Table I). These characteristics were selected from a much larger initial group of possible characteristics, and variables were only used if it was possible to generate a plausible qualitative hypothesis regarding the relationship between that variable and 1-day maximum flows (see Table I). Two other sources of information compiled to aid in this research are (1) codes from the annual instantaneous peak flow database in USGS National Water Information System (USGS NWIS, 2009), which indicate whether the peak flow for each year is altered by either regulation or diversion; (2) remarks that accompany each USGS streamflow station which describe, among other things, sources of alteration of natural streamflows, such as dams, irrigation withdrawals, etc. Methods We employ ordinary least squares multivariate regression procedures which are discussed elsewhere (Helsel and River Res. Applic. (2010) DOI: 10.1002/rra Table I. Watershed characteristics used as potential independent variables Variable name Definition Logtransformed Units Hypothesized relationship with 1-day max Source DrArea Drainage area Yes Sq. km. þ USGS NWIS (2009) Slope Basin average slope Yes Per cent þ 1 km DEM from USGS Flat Per cent flat area (with <1% slope) Yes Per cent 1 km DEM from USGS Precip Median annual precipitation for each decade Yes Mm year 1 þ PRISM data (Daly et al., 2002) Nov6pre, Feb3pre, May2pre, etc. Average of median monthly precipitation for each decade, for months of high flow Yes Mm year 1 þ PRISM data (Daly et al., 2002) Jan3pre, Dec4pre, etc. Average of median monthly precipitation for each decade, for months with most snowfall Yes Mm year 1 þ (snowmelt systems only) PRISM data (Daly et al., 2002) May2tmp, Mar5tmp, etc. Average of median monthly temperatures for each decade, for months of high flow Yes Degrees Kelvin þ (snowmelt systems only) PRISM data (Daly et al., 2002) Snow Snowfall, long-term average Yes Mm year 1 þ (snowmelt systems only) National Climatic Data Center (2009) Runoff Runoff, long-term average Yes Mm year 1 þ Gebert et al. (1987) Aqperm Aquifer permeability Yes Classes 1–7 (lowest–highest) Wolock (2003) Sand Per cent sand Yes Per cent Wolock (2003) Soilthi Soil thickness Yes Mm STATSGO (Wolock, 1997) Soilawc Soil available water capacity Yes Fraction STATSGO (Wolock, 1997) Soildep Soil depth to water table Yes Mm STATSGO (Wolock, 1997) Storatio Total maximum storage capacity of all upstream dams, divided by average annual runoff (Runoff), for each decade No Years of runoff in storage Army Corps of Engineers National Inventory of Dams database from BASINS 2.0 (1999) Popdens Population density, by decade No Persons per sq. km. þ US Census Bureau (2009) T. W. FITZHUGH AND R. M. VOGEL Hirsch, 2002). Regression models were developed for each of 209 hydrologic units (HUs) that cover the bulk of the United States (except for a few areas without streamflow stations). Maps of the HU’s are given later in Section 3 and in Figure A1 in Appendix 1. The dependent variable was log-transformed prior to creating the regressions, as were all independent variables except Storatio (maximum dam storage capacity/mean annual runoff) and Popdens (population density), because use of those two variables in real space led to more precise regression coefficients. The climatic variables used as potential independent variables varied by HU, depending on the timing of flood flows and precipitation during the year and whether the flood response of the HU was dominated by rainfall or snowmelt processes. Due to the computational complexity associated with the model selection procedure, the best regression in each HU was identified automatically using an algorithm written in the R statistical package (R Development Core Team, 2006). This algorithm is described in detail in FitzHugh and Vogel (2010), so it is only briefly summarized here. The algorithm evaluates independent variables in a stepwise manner, evaluating each Copyright # 2010 John Wiley & Sons, Ltd. variable according to its p-value (must be < 0.05), its Variance Inflation Factor (VIF, must be < 5), whether its model coefficient matches the hypothesis in Table I, and whether addition of the variable both increases the adjusted Rsquared and decreases the prediction sum of squares PRESS statistic. From this procedure a series of candidate regressions are identified, and then the final regression for each HU is selected based on a comparison of values of the PRESS statistic. Residuals were evaluated using the correlation coefficient of a normal probability plot of the model residuals, and if necessary, outliers were removed either by visual assessment of this plot or using the DFITS criterion. RESULTS AND DISCUSSION Screening and evaluation of regression models Implementation of the regression selection algorithm yielded 201 HUs with a final regression that was acceptable based on the above criteria, i.e. only eight HUs ended up without a regression. Table A1 in Appendix 1 lists the final River Res. Applic. (2010) DOI: 10.1002/rra Figure 1. Boxplots of adjusted R, PRESS and the coefficients of the five most common independent variables in regressions IMPACT OF DAMS ON FLOOD FLOWS regression models for each HU. Figure 1 shows some key results for the regressions. In general, the regressions performed well, with generally high adjusted Rs and low PRESS statistics. The regression models were then used to quantify the degree to which dams are currently reducing the MAF in each HU. This was done by setting the maximum storage/ mean annual runoff variable, Storatio, to 0, and then recalculating the MAF for all stations that had data in the 1990s. The per cent difference was then computed between this value and the fitted value of MAF from the original regression, and this per cent difference was used as an estimate of the reduction in the MAF during the 1990s due to dam storage. For a few HUs where there were no sites with Storatio1⁄4 0, we set the Storatio to the minimum value in that HU, because it is dangerous to use the regressions outside the range of the data used in their development. Next we used ancillary information available from USGS to screen and evaluate the regressions. We calculated the percentage of years in the 1990s when the peak flows at each streamflow station were coded as altered by regulation or diversion. Then we computed the proportion of total estimated alteration in each HU that was assigned to stations that have no such codes in the peak flow data. One could think of this as an estimate of the proportion of alteration estimated in an HU that is likely to be erroneous. We used this proportion to examine the degree to which regressions that identified statistically significant relationships between Storatio and MAF were estimating an average alteration in the 1990s that was generally representative of conditions in that HU. All HUs where this percentage was greater than 33% were judged to have significant errors, so these 18 HUs were dropped from further analysis. The one exception was the HU for the Susquehanna mainstem, where although this percentage was 51%, there were USGS remarks for all streamflow stations of slight regulation of flows by flood control reservoirs, including those with no alteration indicated in the peak flow codes. We also eliminated eight regressions for HUs that have a high percentage of peak flows coded as altered and gage Copyright # 2010 John Wiley & Sons, Ltd. remarks of impacts of regulation, but where the coefficient for Storatio was not statistically significantly different from zero in the regression. Finally, four more regressions were eliminated because their adjusted R was below 0.5 indicating that the statistical relationship was very weak. Overall, this left 171 regressions where the estimated per cent alterations were considered to be representative enough to continue with further analysis (shown in Figure 2). These models cover 78% of the area of the continental United States. Analysis of the impact of dams on flood flows in the United States Figure 2 is striking because it shows the wide extent of alteration of natural flood flows by dams in the continental United States. The HUs where a statistically significant relationship was found between reduction of flood flows and dam storage cover about 64% of the country, but they cover 84% of the area of the HUs where good regressions were created (those shown in Figure 2). Alteration of flood flows is present in every region of the country, though less so in the mid-Atlantic, Southeast and upper Midwest. Alteration is generally more severe west of the Mississippi and especially in the southern Great Plains, desert Southwest and northern California. Using the estimated alterations for individual gauges, we further summarize these results by freshwater ecoregion (Abell et al., 2008) and river size (see Figure 3). Overall, 3453 stations are available for this analysis. Table II and Figure 4 summarize our results. One obvious and expected conclusion is that the degree of alteration of flood flows increases as the size of the river increases. In the majority of ecoregions, alteration is greater in large rivers than in medium rivers, and greater in medium rivers than small rivers. Across the country, estimated reduction of MAF for large rivers averages 29%, for medium rivers 15% and for small rivers 7%. These data indicate that a large number of rivers in the United States have had significant reduction in flood flows due to dams. To put these numbers in perspective, we compare them to some research results on River Res. Applic. (2010) DOI: 10.1002/rra Figure 2. Estimated per cent alteration (reduction) of MAF by dam storage, for 1990s. This percentage is the average of the estimated per cent alterations for stations in each hydrologic unit that have data in the 1990s T. W. FITZHUGH AND R. M. VOGEL natural variability of flood flows due to long-term climate trends and also on the relationships of flood flow reduction to ecological impacts. Long-term variation in bankfull discharges during the Holocene has been quantified in streams in southwestern Figure 3. Freshwater ecoregions from Abell et al. (2008), and river and Copyright # 2010 John Wiley & Sons, Ltd. Wisconsin (Knox, 2000) and northeastern Utah (Carson et al., 2007), and in both cases compared to modern bankfull discharges. In Wisconsin the maximum variability of Holocene bankfull discharge was 30% from modern discharges, and in Utah it was 20%. Thus, the maximum decrease streams, by size. Numbers are ecoregion ids, referenced in Table II River Res. Applic. (2010) DOI: 10.1002/rra Table II. Average per cent decrease in MAF in 1990s due to dam storage. Numbers are the average of the estimated alterations for all streamflow stations in each ecoregion and river size category. River size classes are 0–1000 km watershed (small), 1000–20 000 km (medium) and 20 000þ km (large). Numbers in bold are size classes where there were fewer than two stations per 1000 km of river length in the size class. This could occur because a portion of the ecoregion did not have a good regression, or because of a lack of stations in general. Table also shows the per cent of area in the US part of the ecoregion that is covered by hydrologic units with valid regressions Freshwater ecoregion ID Per cent of area in United States covered by hydrologic units (%) Average per cent alteration Small rivers and streams (%) Medium rivers (%) Large rivers (%) All rivers and streams (%) Alaska & Canada Pacific Coastal 103 100 10 12 — 10 Apalachicola 155 100 1 24 22 14 Appalachian Piedmont 157 100 4 9 15 6 Bonneville 127 61 10 31 — 17 Central Prairie 146 100 18 18 30 19 Chesapeake Bay 158 100 5 4 9 5 Colorado 130 92 5 15 44 12 Columbia Glaciated 120 92 0 10 30 12 Columbia Unglaciated 121 100 10 12 27 12 Cumberland 151 100 10 3 66 14 Death Valley 128 9 — — — — East Texas Gulf 140 58 17 31 42 27 English—Winnipeg Lakes 109 66 1 16 19 13 Florida Peninsula 156 65 0 0 0 0 Gila 131 100 0 10 40 11 Lahontan 126 97 7 20 34 14 Laurentian Great Lakes 116 69 5 7 — 6 Lower Mississippi 149 76 0 1 41 1 Lower Rio Grande—Bravo 135 11 — — — — Middle Missouri 143 80 7 19 38 21 Mobile Bay 153 61 1 10 0 6 Northeast US & Southeast Canada Atlantic Drainages 118 94 7 19 15 10 Oregon & Northern California Coastal 123 100 5 17 17 10 Oregon Lakes 124 92 — — — — Ouachita Highlands 145 100 1 19 40 15 Ozark Highlands 147 100 0 15 53 17 Pecos 133 0 — — — — Sabine—Galveston 141 85 10 42 37 25 Sacramento—San Joaquin 125 76 14 29 44 21 Southern California Coastal—Baja California 159 62 9 27 13 St.Lawrence 117 98 12 16 — 14 Teays—Old Ohio 150 92 7 14 14 11 Tennessee 152 55 3 10 48 6 Upper Mississippi 148 53 0 10 13 5 Upper Missouri 142 82 6 9 19 10 Upper Rio Grande—Bravo 132 13 31 17 — 24 Upper Snake 122 75 4 14 10 11 US Southern Plains 144 73 17 23 28 23 Vegas—Virgin 129 100 18 35 — 27 West Florida Gulf 154 100 0 0 — 0 West Texas Gulf 139 97 1 10 48 11 Total 78 7 15 29 12 Copyright # 2010 John Wiley & Sons, Ltd. River Res. Applic. (2010) DOI: 10.1002/rra IMPACT OF DAMS ON FLOOD FLOWS