Flood trends and river engineering on the Mississippi River system



[1] Along >4000 km of the Mississippi River system, we document that climate, land-use change, and river engineering have contributed to statistically significant increases in flooding over the past 100–150 years. Trends were tested using a database of >8 million hydrological measurements. A geospatial database of historical engineering construction was used to quantify the response of flood levels to each unit of engineering infrastructure. Significant climate- and/or land use-driven increases in flow were detected, but the largest and most pervasive contributors to increased flooding on the Mississippi River system were wing dikes and related navigational structures, followed by progressive levee construction. In the area of the 2008 Upper Mississippi flood, for example, about 2 m of the flood crest is linked to navigational and flood-control engineering. Systemwide, large increases in flood levels were documented at locations and at times of wing-dike and levee construction.

1. Introduction

[2] Large floods worldwide in recent years have led to repeated suggestions that human activities have caused, or at least worsened, these events. Identifying specific mechanisms driving flood trends and quantifying their contributions, however, has been difficult [Mudelsee et al., 2003; Bronstert et al., 2007]. Previous studies have focused on climate change [Barnett et al., 2008; Kay et al., 2006], land-use change [O'Connell et al., 2007; Quilbé et al., 2008], or both [Blöschl et al., 2007; Ward et al., 2008] as potential contributors to shifts in flood occurrence. Other studies have examined the effects of local “instream” modifications of river channels and floodplains [Pinter et al., 2006].

[3] The present study was conducted on >4000 km of the Mississippi River System (Figure 1), including the Lower Mississippi River (LMR), Middle Mississippi (MMR), Upper Mississippi (UMR), and the Lower Missouri River (LMoR). Both the Mississippi and Missouri Rivers have been extensively modified, primarily to facilitate river navigation and for flood control. In the contributing basin, large-scale changes in land use and land cover have occurred, especially during the past century [Knox, 2006; Kusky, 2008]. In addition, several studies have suggested climate-driven increases in flood-producing precipitation on this system [Ya et al., 2004; Olsen et al., 1999].

Figure 1.

Mississippi and Lower Missouri River study area, including the 66 stations analyzed here. Stations in red are rated gages; green are stage-only stations for which synthetic discharges were generated. Full station list is in Table S1.

[4] Two large river-related databases were constructed to test for trends in flood magnitudes over time on the Mississippi and Missouri Rivers and assess the contributions from channel and floodplain engineering. Our hydrologic database consists of >8 million discharge and stage values, including new synthetic discharges generated for 41 stage-only stations. The hydrologic database was used to test for significant trends in discharges, stages, and “specific stages.” Our geospatial database consists of the locations, emplacement dates, and physical characteristics of over 15,000 structural features constructed along the study rivers over the past 100-150 years. Together these databases were used to generate reach-scale statistical models of hydrologic response. These models quantify changes in flood levels at each station in response to construction of wing dikes, bendway weirs, meander cutoffs, navigational dams, bridges, and other river modifications.

2. Methods and Data Sources

[5] To test for long-term changes in flood magnitudes and frequencies, we constructed a hydrologic database consisting of data from 26 rated stations (with both stage and discharge measurements) and 40 stage-only stations. For each stage-only record, we constructed a parallel record of synthetic discharges based on interpolation of daily flow from nearby rated stations [Jemberie et al., 2008]. Synthetic discharges were generated only for stations with an adjacent rated gage not separated by a significant tributary and only for years with parallel stage measurements. Stage data were adjusted to common vertical datums, and stages and discharges at each station were aliased to a common temporal extent. Two stations (Hulbert and Washington) were excluded from the trend analyses (only) because of gaps within the records. Annual maxima were extracted for the remaining 64 stations, and these records were assessed for temporal autocorrelation. For stations with significant autocorrelation, first-order autoregressive (AR1) models were fit to the annual maxima; Ordinary Least Squares (OLS) were used for non-autocorrelated records (see Table S1 of the auxiliary material). The annual maximum stage and discharge models at each station were then tested for a null hypothesis of “no trend” over time, with significant trends identified at the 95% confidence level or higher.

[6] Specific-gage analysis [Pinter and Heine, 2005; Biedenharn and Watson, 1997] quantifies changes in a river's stage-discharge relationship at a given cross section for which long time series of systematically measured stage and discharge data are available. “Specific stage” is the gage height corresponding to a pre-specified discharge value. Annual rating curves (relationships between stage and discharge) were generated for each station and for each year of complete record, as follows:

equation image

where H is the stage, Q is discharge, and a, b and c are regression coefficients. If any rating relationship had an R2 value less than 0.90, that year of data was excluded from subsequent analyses. Using these annual regressions, specific stages were computed for each year of record for each station in the study area. Because average flows vary as much as 3300% through this study area, we calculated specific stages for multiples of mean daily flow (MDF) at each site: 300% and 400% of MDF on the MMR, UMR, and LMoR and 225% and 300% of MDF on the LMR.

[7] For our geospatial database, 78 reach-scale map sets consisting of 4602 individual map sheets were assembled, scanned (if not already in digital form), and 45 map sets were rigorously rectified. For this study, 53 map sets were utilized (Table S1). GIS data layers were generated to document the progressive emplacement of wing dikes, bendway weirs, levees, meander cutoffs, bridges, navigational dams, and channel constriction (Table S2). These data layers were converted into a standardized coordinate system and re-projected for analysis. Full metadata and a user interface for access to these databases are now being completed by the U.S. Geological Survey [Remo et al., 2008]. These historical geospatial sources were used to parameterize the explanatory variables in our multivariate models of flood response over time.

[8] The response variables in our models were defined from our hydrologic data as change in specific stage at each of the 66 stations (ht) relative to baseline conditions at each location in each year t:

equation image

where Ht is specific stage in year t, and H1 is specific stage for the same discharge in the first year of record. Each geospatial explanatory variable was also re-coded as change over time following equation (2). As a result, our models were designed to identify the response in flood stages to each incremental addition or change in river engineering infrastructure.

[9] Ten models were developed, two each (the two flood conditions) for the four river reaches and for the entire system (“systemwide” model), as follows:

equation image

where y is the dependent variable, t indicates year, β0 is the intercept, βi for i = 1, 2, … k are model parameters, and ɛt is a residual error term. An OLS algorithm was used to solve each system. Stepwise selection also was used to detect the order of significance of each explanatory variable. Model parameters were estimated for all models, and explanatory variables significant at 95% confidence level were identified (Table S3). Data from 2–3 stations from each river reach were withheld for model validation, and comparison of in-sample and out-of-sample errors indicate robust model construction (Table S4).

3. Trend Analyses

[10] Looking at discharge, 11 stations showed trends in flow significant at the 95% level over the duration of record, including at 5 rated gages (Table S1). At the 90% confidence level, one additional station had a significant trend. All of significant discharge trends were positive, and all of these stations were located on the UMR. The lack of any significant negative trends is surprising given the large mainstem dams constructed on the Missouri River during the period of record analyzed here. Reductions in peak flows due to dam retention apparently have been counterbalanced by increases in runoff due to climate change and/or land-use shifts. Only on the UMR, above the Missouri confluence and with minimal dam retention on its own tributaries, did climate and land use result in statistically significant flow increases.

[11] Stage records documented trends significant at the 95% level at 19 stations (Table S1), and again all those significant stage trends were positive. Among these 19 stations, 13 did not exhibit corresponding discharge trends, suggesting that local changes – rather than climate or other upstream controls – must be driving shifts in net flood occurrence (stage) at those sites. For example on the LMoR, with a history of intense navigational engineering linked to conveyance loss [Pinter and Heine, 2005; Hathaway, 1933], six stations exhibited significant stage increases, including at one station (Rulo, NE) with a negative slope (non-significant) in discharge over time. Using a 90% confidence threshold, 29 stations systemwide had significant trends in stage, including three negative trends on the LMR. Stage trends at the remaining 26 stations on the MMR, UMR, and LMoR were all positive – i.e., significant increases in flood levels over time.

[12] The relative contributions of flow trends (discharge) and instream changes to total net change in flood occurrence (stage trends) can be quantified using specific stage. We calculated specific stages for two flow conditions: a minor flood and a moderately large flood (see Text S1) at each site. Specific stages for these flows changed dramatically over time, including increases and decreases >6 m at some sites. Trends in specific stages correlate well with trends in the annual maximum stages (Figure 2b). Because specific stages isolate the effects of instream mechanisms, this strong correlation suggests that the instream river changes have strongly influenced total flood magnitudes over time.

Figure 2.

Slopes of trendlines in specific stage (for 3 * MDF) correlate (a) poorly with discharge trends but (b) well with slopes of maximum annual stages. Maximum annual stages respond to all climate- and land-use-driven changes in flow plus instream conveyance changes. The good correlation in Figure 2b and the poor correlation in Figure 2a suggests that the instream changes had a greater effect upon net flood magnitudes over time on this river system.

[13] Figure 2a illustrates a positive co-variation between discharge trends and total stage change, as expected, but the correlation is quite weak. The poor fit suggests that other factors – other than those that directly control flow volume – may have played a larger role in changing flood levels on these rivers. Discharge increased an average of 0.2% per year (relative to mean annual max. flow), too little to explain the large stage changes at most stations. At the stations with the largest increases and decreases in stage, discharge increases and decreases of >200% would be needed to drive all of the observed stage change. For example, to return flood stages (4 * MDF [mean daily flow] = 23,672 m3/sec) at Grand Tower, IL to levels caused by the same discharge in the 1880s would require discharge to be reduced by almost 60% to ∼10,000 m3/sec. In reality, peak discharge at Grand Tower decreased by just 21.7 m3/sec/year with a non-significant trend. Climate change and upstream basin land use explain some of the large increases in flood levels, but local modifications appear to dominate flood response on the heavily modified Mississippi system.

4. Geospatial Modeling

[14] In addition to the hydrologic data above, we compiled a geospatial database of the precise locations and timing of river construction over the past 100–150 years along >4000 km of the study rivers. This database was used to test local correlations between historical channel and floodplain modifications and changes in flood levels (parameterized as change in specific stages; see Text S1). The most ubiquitous signal in our results was associated with wing dikes. Wing dikes constructed downstream of a location were associated with increases in stage, consistent with backwater effects upstream of these structures. These backwater effects were clearly distinguishable from the effects of upstream dikes, which triggered simultaneous incision and conveyance loss at sites downstream. On the UMR, for example, stages increased 10.8 ± 0.6 cm for each 1000 m of wing dike built within 20 RM (river miles) downstream. These values represent parameter estimates and associated uncertainties for relationships significant at the 95% level in each reach-scale model (see Table S1). At Dubuque, Iowa (Figure 3), for example, roughly 14,000 linear meters of downstream wing dikes were constructed between 1892 and 1928, which were associated with a 1.52 ± 0.08 m increase in stage for 4 * MDF. Wing dikes are used on navigable rivers to maintain minimum channel depths for barge traffic during low-flow conditions. The behavior of these structures during flood flows, however, has been controversial [Stevens et al., 1975; Pinter et al., 2001; Criss and Shock, 2001], and the Army Corps of Engineers continues to build new wing dikes, bendway weirs, and related “river training” structures [McGuire, 2008].

Figure 3.

Hydrologic response at Dubuque, Iowa to wing-dike construction (∼1880–1940) and emplacement of navigational dams (∼1940). Wing-dike index (green) shows cumulative length of wing dikes (in km) within 20 RM downstream. Specific stages increase with wing-dike construction, step up coincident with dam completion (and a last pulse of wing dikes), and subsequently gently decline due to mild incision.

[15] Like wing dikes, levees were present on all four study reaches. The strongest levee signal was on the LMoR. For a flood flow of 4 * MDF, each 1% increase in leveed area downstream was associated with a 2.2 ± 0.2 cm increase in stages. On other river reaches, however, levees resulted in local decreases as well as local increases in stage. Declining stages due to levees are interpreted as the result of incision and flow confinement, whereas increasing stages are interpreted as: 1) loss of floodplain storage, and 2) loss of overbank flow conveyance. These two mechanisms appear to vary in importance depending on local levee geometry and history. On the MMR and LMR, where much levee construction was synchronous, storage losses due to upstream levees were consistently the dominant mechanism. On the LMoR and UMR, where levees were interspersed with un-leveed floodplain during one or more periods, obstruction of overbank flood flow by downstream levees and the associated backwater effects upstream were dominant.

[16] The effects of other river structures were evident either locally and/or at lower levels of significance. On the UMR, navigational dams were built to raise water levels within the pools upstream during low-flow conditions, but during floods the dams are retracted. In the 70-80 years since construction of the navigational dams, flood levels (specific stages for flood flows) have been static or declined slightly over time. This benign hydrologic response occurs primarily, it appears, because permanent inundation of much of the natural floodplain has prevented levee encroachment along most of the UMR. On the LMR, large declines in flood stages occurred over the period of record and were strongly correlated with the locations and timing of meander cutoffs. The strongest correlations were with cutoffs downstream of each station and with time since cutoff emplacement, consistent with progressive incision following completion of each cutoff [Biedenharn and Watson, 1997]. Incision due to individual meander cutoffs has coalesced, resulting in reach-scale shortening and incision as seen at Vicksburg, MS, where three man-made cutoffs downstream in 1929, 1933, and 1936 triggered incision and flow acceleration that reduced stages by up to 4 m, followed by a gradual rebound of 1–1.5 m in the subsequent decades.

5. Conclusions

[17] Damages from floods worldwide have risen dramatically over the past 100 years [Munich Re Group, 2007]. Much of this increase is due to economic development in floodplains [Pinter, 2005; Pielke, 1999], but flooding itself has physically increased in magnitude and frequency on many rivers [Ward et al., 2008; Pinter et al., 2006; Helms et al., 2002]. Discharge records from 66 stations on the Mississippi River system confirm a pattern of increasing flows, although significant trends were detected only on the UMR, upstream of most of the reservoir capacity on the system. These discharge trends contribute to increases in total flood levels documented on the study rivers, but total change in flood levels – including trends significant at the 90% level at 29 stations – correlate more strongly with instream river modifications such as navigational dikes and levees. Previous studies have shown that climate and land-use change may significantly impact flooding, and those results are echoed here. Our results further quantify the cumulative long-term impacts of navigation engineering, flood control, and other local river structures and activities. The navigable rivers of the Mississippi system have been intensively engineered, and some of these modifications are associated with large decreases in the rivers' capacity to convey flood flows. We suggest that past hydrologic response to river engineering represents a cumulative and empirical measure of hydrologic response that can be used to balance the local benefits of river engineering against the potential for large-scale flood magnification.


[18] The research reported here was supported by NSF grants 0229578 and 0552364.