Review of Sediment Modeling Tools Used During Removal of the Elwha River Dams

Abstract

The rapid mobilization of sediment stored behind dams, in amounts that are large relative to mean annual sediment loads, can jumpstart river restoration but can also adversely impact habitat, infrastructure, land, and water use upstream of, within, and downstream of the former impoundment. A wide range of geomorphic and engineering assessment tools were applied to help manage sediment-related risks associated with the removal of two dams from the Elwha River in Washington State and the release of roughly 21 million m³ of sediment. Each of these tools had its strengths and weaknesses, which are explored here. The processes of sediment erosion, transport and deposition were complex. No one model was able to fully simulate all these with the accuracy necessary for predicting the magnitude and timing of coarse and fine sediment release from the reservoir. Collectively, however, the model outputs provided enough information to guide the adaptive sediment management process during dam removal. When the complexity of the morphodynamic responses to dam removal and the associated risks exceeded the capacity of any one tool to adequately assess, synoptic forecasting proved useful. The lessons learned on the Elwha have provided insights into how to use a variety of modeling techniques to address sediment management issues as dam removal scale, complexity and risk increase.

1. Introduction

In the US and Europe, the rate of dam and weir removal has increased rapidly since the 1990s and is likely to continue increasing [1], offering opportunities to reverse anthropogenic fragmentation of river systems, increase ecosystem resilience and help address the global biodiversity emergency by providing access to more, and better quality, physical habitats for aquatic organisms [2,3,4,5,6,7,8,9,10,11,12].

Multiple benefits arise from dam removal, e.g., [13,14,15,16], but the mobilization of potentially large volumes of impounded sediment risks creating unstable landscapes within former reservoir areas, channel incision along and upstream of reservoir deltas, and aggradation and increases in suspended sediment loads in the downstream river channels, e.g., [17,18,19,20,21].

These changes can result in a range of adverse ecological and socio-economic impacts. For example, the removal of the Glines Canyon Dam from the Elwha River was halted for a year because the downstream water treatment plant had been overwhelmed by fine sediment (particle size diameter less than 0.062 mm) [22], while planning for the removal of Matilija Dam has taken three decades because of concerns around fine sediment impacts to aquatic biota, the operation of the Robles Diversion Dam and water quality problems in Lake Casitas, which is fed by the Robles Diversion [23]. Similarly, the four hydroelectric dams removed from the Klamath River in 2024 contained approximately 11.9 million m³ of sediment (84% silt and clay, 16% sand and gravel), of which approximately 36–57% was predicted to be transported downstream, creating suspended sediment concentrations of 7000 to 14,000 mg/L for several months during dam removal [24]. Predicted adverse impacts included flood-related risks to infrastructure, which necessitated the planning of mitigation measures.

To date, most dam removals have involved small to moderately sized structures, and only a small proportion of the morphodynamic and ecological responses to removal have been studied, e.g., [1,25,26,27,28,29,30,31,32,33,34,35,36,37]. Few large dams have been removed [38,39,40], and only a subset of these have publicly available studies, many of which relate to the Elwha, e.g., [4,7,41,42,43,44,45,46,47,48,49,50,51,52,53].

Forecasting the nature, timing and extent of sediment erosion, transport and deposition during and following dam removal is thus a fundamental requirement in assessing the associated risks to people, property, infrastructure, species, and institutional reputations. A range of engineering and geomorphic forecasting tools exist, not all of which have been thoroughly tested in the more unusual or demanding conditions associated with large dam removals.

With reference to post-removal field data, studies undertaken over a 28-year period before, during and after the concurrent removals of the Glines Canyon and Elwha Dams from the Elwha River, WA, USA, provide a rare opportunity to review the relative strengths and weaknesses of a suite of riverine conceptual, numerical and physical sediment modeling tools in simulating different aspects of the morphodynamics of sediment movement through and out of two large impoundments and through downstream river reaches. At the start of the dam removal planning process, there was no master plan that dictated which tools or models would be used and when. Rather, tools/models were selected more organically in response to particular information requirements as and when these became apparent. Different tools/models therefore produced different outputs, not all of which are directly comparable, though some are. Their performance is thus assessed against the questions they were designed to address and, where possible, against each other. The aim of this study is thus to provide information that will hopefully be of use to other practitioners in selecting which tools, or combination of tools, could be most useful in addressing site-specific sediment management issues related to large dam removals.

2. Study Site

The Elwha River originates in the Olympic Mountains and flows north for 70 km before entering the Strait of Juan de Fuca at Angeles Point, about 10 km west of Port Angeles, WA, USA (Figure 1). A large portion of the 833 km² watershed (83%) is within the Olympic National Park and the Daniel J. Evans Wilderness Area boundaries. The average annual precipitation of the Elwha River watershed ranges from 1 m at the river mouth to 6 m on Mount Olympus [54]. The average annual discharge is 42 m³/s and the 2-year flood peak is 400 m³/s [48,54]. Floods tend to be of short duration (one or two days) and typically occur during the fall–winter storm season. Discharges during spring snowmelt are typically less than during winter floods, but longer in duration.

Past and present glacial activity in the steep mountainous watershed, combined with high precipitation, produces average annual sediment loads of 230,000 m³/yr [48]. The natural riverbed is composed of gravel, cobbles, boulders, and sand, with some exposed bedrock. The river reaches alternate between narrow bedrock canyons and wider alluvial reaches for much of its length [55,56]. Before construction of Elwha Dam and Glines Canyon Dam, the Elwha River supported anadromous fish, including all five species of Pacific salmon and four species of trout [57,58]. The dams blocked fish passage and trapped natural sediment loads, leading to the loss of salmon runs and inundating cultural sites of the Lower Elwha Klallam Tribe [57].

Impounding Lake Aldwell, Elwha Dam was completed in 1913 as a 32 m high concrete gravity dam 7.8 km from the river mouth. Glines Canyon Dam was completed in 1927 as a 64 m thin-arch concrete dam 21 km from the river mouth and formed Lake Mills. Lake Mills trapped 16.1 ± 2.4 million m³ of sediment and Lake Aldwell trapped 4.9 ± 1.4 million m³ [22,48]. The incremental cutting down of both dams occurred concurrently, beginning in September 2011. Elwha Dam was completely removed by April 2012 and Glines Canyon Dam by August 2014.

With 84 years of coarse-sediment supply (particle diameter equal to or greater than 0.062 mm) trapped within Lake Mills, the level of sediment-related risk to downstream infrastructure and property was especially high, because located downstream of one or both dams are the Altair Campground, park roads, private property, the Elwha Surface Water Intake (ESWI), the Elwha Water Treatment Plant (EWTP), the tribal fish hatchery, and tribal lands, and homes (Figure 1).

3. Review of Sediment Modeling Tools Used for Dam Removal

The suite of riverine conceptual, numerical and physical sediment modeling tools is described in the following sections. Each modeling tool was selected and applied to address specific sediment management issues (e.g., extent of reservoir sediment erosion, transport to the sea, aggradation, increased flood stage, and sediment concentration peaks and patterns). The applicable locations and time periods simulated are summarized in

Table 1. Location and time period applicable to each model or tool. * The model was run before dam removal occurred, so the modeled time periods do not correspond to an actual calendar period of time.

Model or Tool	Location	Time Period Applicable
Conceptual model	Lake Mills to Elwha River mouth	Prior to dam removal, during dam removal, and post dam removal floods.
Reservoir drawdown experiment	Lake Mills	Prior to dam removal (April 1994).
Mass balance modeling 1	Lake Mills and Lake Aldwell	Simulated dam removal (~1.5 years) and post dam removal floods (~1.83 years). *
Physical modeling	Lake Mills	Simulated dam removal (~1.5 years) and post dam removal floods. *
Mass balance modeling 2	Lake Mills and Lake Aldwell	Measured reservoir drawdown and dam removal (1.5 years from October 2010 to April 2012). Simulated dam removal (1.75 years from April 2012 to January 2014) and post dam removal floods (2.83 years from January 2014 through November 2016).
1D numerical modeling	Rica Canyon (above Lake Mills) to Elwha River mouth	Simulated historical Lake Mills sedimentation (1927 to 1994). Simulated pre dam removal (2 years), dam removal (2 years), and post dam removal (4 years). *
2D numerical modeling	Lake Mills	Simulated erosion measured in physical model.
Synoptic forecasting	Lake Mills	Simulated terrace erosion during water years 2015 and 2016.

.

3.1. Conceptual Modeling and Geomorphic Analysis

Grant et al. [59] define a conceptual model as ‘a persistent set of ideas that usefully organizes thinking…by providing representations or abstractions of complex systems that make them easier to understand’. These models can help you to understand why a river looks the way it does and what it will look like in future [59]. Developing a conceptual model thus forces one to think through the fundamental process–form–response interactions that are likely to occur, which serves to identify potential risks and processes that may need to be assessed in more detail in order to determine appropriate mitigation. Drawing on knowledge held by multiple individuals and organizations and incrementally updating it as knowledge was derived from other studies, the Elwha conceptual model synthesized existing conditions and provided a qualitative understanding of the sediment erosion and redeposition processes expected in the reservoirs and downstream channel during dam removal [60,61].

The model hypothesized that the rate of reservoir sediment erosion would primarily be a function of discharge and the rate of reservoir drawdown. Lake Mills contained a 20 m to 30 m thick coarse sediment delta at its upstream end and 4 m to 16 m thick fine sediment deposits along the reservoir bed (bottomset deposits) between the delta and the dam. The eroded coarse sediment was expected to redeposit on top of the bottomsets across the full width of the receding reservoir at progressively lower elevations as the delta prograded toward the remnants of the dam [62]. The deltas were expected to erode by both incision and lateral channel migration, though some parts were expected to remain as terraces along the valley walls [63]. Lake Mills was 5 to 10 times wider than the Elwha River, and Sawaske and Freyberg [34] and Morris and Fan [64] found that dam removals with large ratios of channel width to reservoir width have a reduced ability for the channel to access and erode all the reservoir sediment deposits.

An important element of ecosystem restoration involves recreating a fully connected channel–floodplain system [60,61] that is in a state of dynamic equilibrium (Table 2). Within the reservoirs, this requires the majority of the coarse delta sediment to be eroded and redistributed throughout the basin during dam removal and the first few post-removal floods. A conceptual understanding of how this might be achieved was boosted by outputs from the reservoir drawdown experiments (Section 3.2) and physical models built at the University of Nottingham (UoN) and the St. Anthony Falls Laboratory (SAFL) (Section 3.4). Due to base-level lowering during drawdown, channels on the delta surface in Lake Mills were expected to incise through their armor layer during flows as small as 28 m³/s, which is only 7% of the 2-year flood. Bank erosion driven by this incision was then expected to drive channel widening and lateral movement across the delta surface [63]. The SAFL model showed that excavating a pilot channel along the centerline of the delta prior to dam removal would substantially increase rates and magnitudes of lateral erosion compared with allowing channel erosion along the erosion-resistant margins of the reservoir.

According to the conceptual model [62,65], the rate of sediment release into the river downstream during dam removal was expected to vary with discharge and the rate and timing of dam lowering. Following dam removal, sediment discharge downstream was expected to vary with the timing and magnitude of floods, with progressively higher flood stages causing some degree of channel and floodplain widening within the remaining reservoir sediment [60,61,65].

The middle river reach between the two reservoirs features narrow bedrock canyons interspersed with wider, alluvial floodplain reaches [13]. Following dam closure, the alluvial reaches incised and armored, reducing lateral migration and allowing vegetation to colonize and stabilize the channel banks and bars. With a slope of roughly 0.7–0.8% [13], stream powers were initially expected to be sufficiently high that sediment would be predominantly transported through it after dam removal [62,65]. Restoration of coarse sediment supply to the lower reach downstream of Elwha Dam, which had a slope of roughly 0.4% [13], was expected to fill pools and increase river water surface elevations [60,62,65], restore the full range of bed material sizes (sand, gravel, cobble, and boulder), increase the active channel width, and rebuild the coastal delta. In addition, the wood load trapped within the reservoir sediment and the natural supply from upstream were expected to increase the presence of log jams. Lateral channel dynamism was expected to increase within the ~1 km wide corridor constrained between flood levees. Because of an increase in flood risk associated with these changes, cash payments were made to landowners between the National Park boundary and the upstream end of Lake Aldwell to pay for flood mitigation measures to be installed and for any future flood damage that might occur. Although sediment deposition was twice as much in the lower reach as in the middle reach, deposition along the middle reach was greater than predicted and triggered sufficient lateral channel movement to wash out roads and recruit trees to the channel.

Initially, fine sediment concentrations released downstream were expected to be small, as the small amounts eroded from the delta mixed with the receding reservoir pool and settled on the bed. As the incising delta surface channel cut into the original bottomset deposits, however, more fine sediment was expected to be recruited into an ever-shrinking reservoir pool, thus increasing the concentrations released downstream for a given reservoir inflow. Concentrations were expected to peak once the reservoir had completely disappeared and eroded bottomset sediments were released directly to the downstream river. Together with natural concentrations after dam removal, this was expected to increase water treatment needs for municipal and industrial water users [60,65], so the following water quality mitigation measures were implemented prior to dam removal:

• • Construction of the ESWI and EWTP to, respectively, divert and pre-treat river water for municipal and industrial users. (Numerical modeling was used to predict the concentrations to be dealt with.)
• • Construction of a new Water Treatment Plant for the City of Port Angeles.
• • Upgrades to the existing industrial water treatment plant.
• • New and deeper municipal wells were drilled for the Dry Creek Water Association and the Elwha Place Homeowners Association.

Turbidity and abrasion due to the high sediment concentrations were also expected to affect fish and macroinvertebrates, primarily during the water years (WYs) of dam removal. A new fish hatchery was thus constructed for the Lower Elwha Klallam Tribe to help native fish recover following dam removal.

The conceptual model concluded that phased dam removal could efficiently allow the river’s natural energy to erode and transport the reservoir sediments to the sea. When coupled with mitigation to improve flood protection and water quality treatment, the phased approach would provide the most cost-efficient approach to ecosystem restoration. The conceptual model also led to the conclusion that most of the related Lake Aldwell sediment transport would be over before the larger sediment volume in Lake Mills entered the river, and that the former Lake Aldwell reach would then experience sediment impacts just like other downstream alluvial river reaches.

Geomorphic analyses were undertaken to help quantify inferences based on the conceptual model and chart recovery towards dynamic equilibrium in both the reservoir and river channel downstream (Table 2). A state of dynamic equilibrium was expected to exist when rates of sediment input and output became balanced over time, even though sediment was continually being eroded, transported, and deposited and channel form was adjusting in response [66]. Recovery of the longitudinal pre-dam river profile through Lake Mills was expected to be achieved soon after dam removal was complete, while lateral erosion rates of reservoir sediment terraces were expected to be initially high and then decrease exponentially with each subsequent water year until a dynamic equilibrium was achieved.

Table 2. Use of geomorphic analyses to assess recovery of the Elwha River system towards dynamic equilibrium [22].

	Criteria	Findings	Information From
Reservoir Area
1	River incised to pre-dam valley bottom indicates low likelihood of further removal-related incision.	Exposure of pre-dam tree stumps, historical photos and pre-dam & contemporary topography indicate river through former reservoir reached pre-dam valley bottom by October 2017.	Repeat fluvial audit; contemporary ground-based photos; historical and contemporary surveys.
2	At least one flood peak greater than the 5-year flood has occurred since dam removal.	Seven peaks ranging from 2 to 10 yr recurrence intervals have occurred since removal.	Stream gauging and analysis.
3	Lateral erosion of terraces is localized (not along entire bank line), only occurs during floods, and a flood of a given magnitude erodes less than the previous flood of a similar magnitude.	Terrace erosion reduced annually since dam removal: larger, more frequent flooding in WY 2016 eroded less sediment than in WY 2015; a peak flow no larger than a 2 yr flood in WY 2017 eroded minimal sediment; a 2 to 5 yr flood in WY 2018 eroded 5–10% of the terraces eroded previously.	Repeat fluvial audit.
4	Annual erosion of remaining sediment is less than natural variability in year-to-year background sediment loads.	By the start of WY 2018, erosion volumes were less than the natural variability in background loads.	Stream gauging; sedimentological data; repeat topographical & bathymetrical surveys; calculation of natural background sediment loads; geomorphic change detection.
5	Net annual erosion from the reservoir relative to natural background loads exhibits an exponential decay over time.	Net erosion declined exponentially from 63 times background rate in WY 2013 to 0.4 and 0.8 times the background rates in WYs 2017 and 2018, respectively.	As for criterion 4.
6	Woody vegetation covering the majority of fine sediment deposits on reservoir hillslopes and terrace surfaces prevents erosion by rainfall runoff, snowmelt and wind.	By the end of 2017, reservoir hillslopes and sediment terraces were covered in dense vegetation, including woody species.	Repeat surveys; oblique time lapse photos.
Downstream river channel
1	Recovery of pool–riffle morphology indicates that most of the released coarse sediment has been transported to the coastal delta.	The pool-riffle morphology, with abundant gravel bars and wood deposits, re-established itself within a more complex, laterally migrating, naturally dynamic channel by 2014. Sediment erosion and deposition occur within the active channel and floodplain.	Repeat topographical survey & ground-based photography.
2	Most of the eroded reservoir sediment has been transported to the Elwha River mouth.	90% of the eroded reservoir sediment has been transported to the coast, greatly enlarging the delta [67]. Sediment remaining in the river has been sorted to restore fish habitat and natural river ecosystem function.	Historical survey; pre- and post-removal topographic and bathymetric survey; aerial LiDAR; sedimentological data; geomorphic change detection; gauging data; sediment transport data; calculation of natural background sediment loads; sediment budgeting.
3	Lateral channel migration and associated terrace and bank erosion is limited to periods during floods.	Lateral channel migration and bank erosion occurred during floods but reduced in longitudinal length and in surface area over time.	Repeat fluvial audit.

Table 2. Use of geomorphic analyses to assess recovery of the Elwha River system towards dynamic equilibrium [22].

	Criteria	Findings	Information From
Reservoir Area
1	River incised to pre-dam valley bottom indicates low likelihood of further removal-related incision.	Exposure of pre-dam tree stumps, historical photos and pre-dam & contemporary topography indicate river through former reservoir reached pre-dam valley bottom by October 2017.	Repeat fluvial audit; contemporary ground-based photos; historical and contemporary surveys.
2	At least one flood peak greater than the 5-year flood has occurred since dam removal.	Seven peaks ranging from 2 to 10 yr recurrence intervals have occurred since removal.	Stream gauging and analysis.
3	Lateral erosion of terraces is localized (not along entire bank line), only occurs during floods, and a flood of a given magnitude erodes less than the previous flood of a similar magnitude.	Terrace erosion reduced annually since dam removal: larger, more frequent flooding in WY 2016 eroded less sediment than in WY 2015; a peak flow no larger than a 2 yr flood in WY 2017 eroded minimal sediment; a 2 to 5 yr flood in WY 2018 eroded 5–10% of the terraces eroded previously.	Repeat fluvial audit.
4	Annual erosion of remaining sediment is less than natural variability in year-to-year background sediment loads.	By the start of WY 2018, erosion volumes were less than the natural variability in background loads.	Stream gauging; sedimentological data; repeat topographical & bathymetrical surveys; calculation of natural background sediment loads; geomorphic change detection.
5	Net annual erosion from the reservoir relative to natural background loads exhibits an exponential decay over time.	Net erosion declined exponentially from 63 times background rate in WY 2013 to 0.4 and 0.8 times the background rates in WYs 2017 and 2018, respectively.	As for criterion 4.
6	Woody vegetation covering the majority of fine sediment deposits on reservoir hillslopes and terrace surfaces prevents erosion by rainfall runoff, snowmelt and wind.	By the end of 2017, reservoir hillslopes and sediment terraces were covered in dense vegetation, including woody species.	Repeat surveys; oblique time lapse photos.
Downstream river channel
1	Recovery of pool–riffle morphology indicates that most of the released coarse sediment has been transported to the coastal delta.	The pool-riffle morphology, with abundant gravel bars and wood deposits, re-established itself within a more complex, laterally migrating, naturally dynamic channel by 2014. Sediment erosion and deposition occur within the active channel and floodplain.	Repeat topographical survey & ground-based photography.
2	Most of the eroded reservoir sediment has been transported to the Elwha River mouth.	90% of the eroded reservoir sediment has been transported to the coast, greatly enlarging the delta [67]. Sediment remaining in the river has been sorted to restore fish habitat and natural river ecosystem function.	Historical survey; pre- and post-removal topographic and bathymetric survey; aerial LiDAR; sedimentological data; geomorphic change detection; gauging data; sediment transport data; calculation of natural background sediment loads; sediment budgeting.
3	Lateral channel migration and associated terrace and bank erosion is limited to periods during floods.	Lateral channel migration and bank erosion occurred during floods but reduced in longitudinal length and in surface area over time.	Repeat fluvial audit.

3.2. Reservoir Drawdown Experiments

Reservoir drawdown experiments provided opportunities to collect sediment samples from the exposed delta and to observe and document the delta erosion processes prompted by base-level lowering. Observations and measurements made during the experiments greatly improved the accuracy of the conceptual model and dispelled a prior conceptual expectation that the river’s capacity to erode deltaic sediments would be limited by armoring and large wood.

In 1989, Lake Mills was lowered 6 m to allow sediment coring of the exposed delta, which included six auger holes, 13 piston core samples, and 32 thickness probe measurements [68]. Size gradations and stratigraphic data from the cores served as inputs to the mass balance (MB) models. Combined with 1994 field measurements, roughly equal proportions of fine and coarse sediment were found in the reservoir, with downstream fining and upwards coarsening of the deltaic sediments [69]. Daily channel evolution and the total sediment erosion volume of the exposed delta were not documented.

In 1994, Lake Mills was lowered 6 m over a one-week period, held at a constant level for a week and then allowed to refill. Daily monitoring of cross-sections, discharge, and sediment loads was conducted [63]. The total erosion volume was not determined, but erosion channels through the exposed sediment advanced the delta front 100 m farther into Lake Mills. Delta surface cross-sections were surveyed daily [63]. Discharge, suspended sediment concentration and bedload were measured daily, upstream, on, and downstream of the delta, to provide insights into the delta’s mechanisms of morphodynamic adjustment. As the reservoir was lowered, a tree-covered island at the head of the delta created two main channels. Channel incision dominated during reservoir drawdown and was greatest at the delta front, declining with distance upstream as the knickzone graded into the existing channel slope. Driven by incision and lack of sediment cohesion, braiding and topographic steering of the flow by terraces and the valley walls created extensive channel widening during the week with constant water level. Throughout the experiment, the right-hand channel remained attached to the bedrock valley wall in the upstream half of the delta. Large wood pieces were undermined by incision and did not inhibit either vertical or lateral erosion, though they did locally accelerate rates of lateral erosion, split flow, and generate braiding [63]. Eroded material was deposited as lobes at the downstream end of each channel, prograding first longitudinally, then laterally until a delta sweep (sediment deposited across the full reservoir width) was completed. This highlighted the importance of making hold periods during dam removal sufficiently long to meet the sediment management objective of maximizing lateral sediment erosion and redistribution within the reservoir. These findings informed design of the mass balance models and guided the development of research hypotheses investigated using the physical model. At this stage, the optimum increments of reservoir drawdown and hold period lengths for maximizing sediment redistribution remained unclear.

3.3. Mass Balance Modeling Phase 1

The drawdown experiments demonstrated that delta erosion involved interactive processes of incision, widening, lateral channel movement, and the release and mixing of fine sediment that the numerical models then available (circa 1996) were incapable of simulating. To fill this gap, cross-section-based mass balance models (MB1) were developed for Lakes Mills and Aldwell to simulate delta erosion driven by incremental lowering of the dam and to support production of Environmental Impact Statements [60,61]. MB1 used field data, empirical rules, regime equations and fine sediment settling equations to simulate vertical and lateral delta erosion, delta sweeping, sediment redeposition within the remaining reservoir pool and evacuation through the dam site, while conserving the mass of reservoir sedimentation, upstream sediment supply, and downstream transport [65]. The model relied on assumptions derived from the conceptual model and drawdown experiments [63] that varied depending on if a reservoir pool remained or had been completely drained. For example, while a reservoir pool remained, it was assumed that all the eroded coarse sediment and a portion of the fine sediment would deposit therein, that the width of the channel eroding the delta was a function of the discharge, and that channel width increased with distance downstream, reaching a maximum where it was sweeping across the receding reservoir pool.

These models proved useful in estimating the quantities of coarse and fine sediment eroded and redistributed during dam removal and by post-dam floods. Results also indicated when the receding reservoir pools would fill with sediment, triggering the first release of coarse sediment downstream. The models predicted:

• • That 15% to 35% of the coarse sediment and ~50% of the fine sediment would be eroded from the reservoirs;
• • The date on which coarse sediment would first be released downstream;
• • That maximum downstream fines concentrations would be in the range 10,000–50,000 mg/L;
• • That bed elevations in the lower river would rise appreciably;
• • That reservoir sediment erosion rates would decrease exponentially once dam removal was completed and a few floods had occurred.

Although MB1 predictions underestimated the percentages of coarse and fine sediment eroded from the reservoirs and overestimated peak concentrations of fine sediment relative to actual dam removal values, they helped guide the adaptive management monitoring design used to determine the rate of dam removal and set benchmarks against which to compare monitoring results.

Predictions concerning downstream sediment concentrations and aggradation helped guide decision making on securing the supply of drinking water for Port Angeles, water supply to the hatchery and industrial users, raising flood levees to maintain authorized levels of protection, and informed the pace of dam removal to achieve the desired ecosystem restoration outcomes and minimize impacts to fish during migration periods. Downstream results also informed environmental permitting on the expected duration and frequency of background sediment loads exceeding threshold values.

3.4. Physical Modeling

Physical modeling was performed to advance understanding of the sediment processes observed during the reservoir drawdown experiments at Lake Mills.

An un-scaled tabletop model built at the University of Nottingham (UoN) in 2000 simulated channel incision and widening by bank collapse following a drop in base level in two delta channels that qualitatively mimicked those that developed during the 1994 drawdown experiment (Figure 2a). Eroded material was deposited as two lobes, prograding first longitudinally and later increasingly laterally towards a delta sweep.

At SAFL, Bromley [70] built and ran a vertically distorted, scaled physical model of Glines Canyon Dam, Lake Mills and its delta (bottomset deposits not represented) (Figure 2b,c). The SAFL model was able to simulate varying-sized increments of dam removal and delta surface channel starting positions. This showed that lateral erosion and redistribution of delta sediment was more extensive when a pilot channel was excavated along the delta’s centerline (Figure 3a), eroding substantially more sediment than when the channel was allowed to align along either the left or right valley wall. These findings informed the decision to excavate a pilot channel in the real delta prior to dam removal (Figure 3b,c). The pilot channel allowed the incising channel to access virtually the entire delta surface through the operation of three interrelated mechanisms.

First, lateral erosion occurred most rapidly when one terrace topographically steered flow across the delta to impinge against the opposite terrace, creating the space and sediment load necessary for braiding. Second, once sediment loads decreased sufficiently, any sub-channel in the braided planform could capture the bulk of the flow to recreate a single-thread channel capable of rapidly eroding a different part of the delta surface. Third, avulsion of the single-thread channel could efficiently redirect erosion to a different part of the delta. Without a pilot channel, the stream flow contacted the reservoir basin boundary at an early stage and remained there for a greater proportion of the model run, which muted the operation of these mechanisms and constrained them to either the left or right half of the delta.

The SAFL model also established that the capacity of post-dam removal floods to further erode the delta depended strongly on the eroded delta’s topography. In runs with a pilot channel, floods eroded only about 5% more sediment because so much of the delta had already been eroded and redistributed. Post-dam removal delta erosion was also limited when the channel incised close to the left basin boundary, because the channel’s curvature directed the bulk of the floods’ erosive force away from the large terrace deposits to the right, causing the channel to incise quickly to the lowest point of the valley floor. In contrast, when the channel incised closer to the right basin boundary, the cross-valley slope from right to left drove flood flows against the left terrace, eroding large volumes of sediment before the channel could cut down to the valley floor.

Leung et al.’s [72] tabletop model, run after the actual dam removal, demonstrated the role played by large wood in creating a more locally complex, braided planform on the delta topset (the gently sloping delta surface). Individual wood pieces or jams were able to divide flow and help braid bars to emerge over two cycles of falling base level and a hold period, when compared to a control run without wood [72]. This concurred with observations made during the 1994 drawdown experiment and the actual dam removal. There was little difference in the rate of lateral channel migration and delta sweeping in the model runs with and without wood.

3.5. Mass Balance Modeling Phase 2

From the second drawdown experiment in 1994 to 2010, the volume of sediment stored in Lake Mills increased substantially from 10.6 million to 16.1 million m³ (52%) [73], which prompted the development of a second phase of mass balance modeling (MB2) [71].

MB2 retained many of the concepts and assumptions of MB1 but was updated with observations from the SAFL model and new field measurements of delta topography. Rather than using cross-sections, MB2 included a GIS module that reproduced the complex geometry of Lake Mills and allowed simulation of non-linear channel alignments, variable channel widths and variable erosion and deposition slopes, while supporting 2D and 3D visualization of the evolving delta. It partially corrected MB1′s under-predictions of future eroding channel width by using an additional function that increased erosion width as a function of the number of days above the average discharge during a given model time step [71].

Upstream boundary conditions included daily discharge and mean daily sediment loads computed using an equation derived from measured sediment transport data. The downstream boundary conditions replicated the planned reservoir drawdown schedule, policy limits on the incremental drawdown rate and the required hold periods after each 3–5 m drawdown increment and during 2-month-long, seasonally scheduled, fish migration windows, during which dam removal was halted. Digital elevation models derived from photogrammetry on a weekly to monthly basis during dam removal were also available, from which updated channel slopes and widths were extracted to provide additional boundary conditions [74]. The model ran fast enough to produce multiple simulations during each increment of dam removal, allowing its boundary conditions to be updated with the monitoring data ahead of each subsequent increment of dam removal. This informed the alignment of future delta erosion channels, resulting in accurate simulations of the current period and reduced errors in simulations of the subsequent period, despite channel meandering and migration not being simulated. This allowed adaptive management decisions to be made, notably the one-year-long hold period required to upgrade the EWTP and to allow downstream bed elevations to be reduced.

Depending on the incoming hydrographs used, 47% to 57% of the coarse sediment and 57% to 74% of the fine sediment were predicted to erode from Lake Mills, with maximum concentrations from 17,000 to 32,000 mg/L predicted for the ESWI downstream of Lake Aldwell. Outputs from MB2 were used to inform sediment management decision making, including the rate of reservoir drawdown, throughout the removal of Glines Canyon Dam.

3.6. One-Dimensional Numerical Modeling

Konrad [62] applied a 1D mobile-bed model as part of a larger effort to simulate historical sedimentation in Lake Mills and better understand how sediment would be eroded from the delta and transported and deposited along the Elwha River.

Konrad’s model simulated reservoir erosion solely through incision related to base-level lowering. A trapezoidal channel cross-section was specified, meaning that 37% of the sediment in Lake Mills was unavailable for erosion. Different sequences of wet and dry years were simulated, with annual maximum daily flows ranging from 125 m³/s to 600 m³/s. The model provided estimates of how much aggradation might occur along the downstream alluvial reaches as a result of sediment erosion and evacuation from Lakes Mills and Aldwell.

The 1D model predicted that most of the available sediment volume would erode from the reservoir after dam removal. After two water years of pre-dam removal, two water years of dam removal, and four water years of post-removal flows, the model predicted that between 38% and 52% of the 1994 sediment volume (28% and 35% of the 2010 volume) would have been evacuated from Lake Mills. The model predicted that, by the end of water year 7, 50% of the depositional reaches downstream from the two dams would aggrade by more than 1 m, while 10% would aggrade by more than 3.3 m [62]. Ritchie et al. [49] reported “1.0 to 1.5 m of widespread riverbed aggradation,” based on aerial surveys, water-level monitoring, and topographic surveys.

3.7. Two-Dimensional Numerical Modeling

Lai [75] used the SRH-2D model (version 2.2) and bank erosion module to simulate the lateral channel erosion and sediment deposition observed in the SAFL model run with a center pilot channel and three pieces of dam removed (run 3×C). Use of a moving mesh allowed accurate simulation of the spatial distributions of bank and bed erosion and deposition observed in the SAFL model. However, relative to the SAFL model, the 2D numerical model over-predicted bank erosion in the upstream half of the pre-dam removal delta and under-predicted it in the downstream third, which resulted in the size of the prograding delta being under-predicted. SRH-2D was able to reasonably simulate observed bank erosion and channel migration equal to about one-third of the channel width (~24 m) at a bend in the Rio Grande in New Mexico over a five-year period [76]. However, bank erosion of this magnitude and rate is modest compared to that observed at Lake Mills, where bank erosion and channel widening varied between 4.5 and 8 times the pre-dam channel width (~274 m to ~488 m) [22,49]. Lai [75] discovered that the moving mesh became unstable when simulating more than about one channel width’s worth of bank erosion and concluded that while simulation of lateral erosion using a moving mesh shows promise, in cases of extreme bank retreat a static mesh may be better suited.

3.8. Monitoring and Adaptive Management During Dam Removal

A program of monitoring and adaptive management was designed and implemented to verify and update model predictions and inform management decisions regarding when to proceed to the next phase of the staged dam removals [22]. This was necessary due to irreducible uncertainties in (a) discharge hydrographs, (b) the schedule of dam removal and reservoir drawdown, (c) the rates and amounts of reservoir sediment erosion, (d) how much of the eroded sediment would be retained within the shrinking pools and for how long, (e) the capacity of the river to transport sediment of different sizes downstream of the dams, and (f) the rates and long-stream distributions of deposition between the dams and the river mouth.

Flow did not significantly increase with distance downstream due to few tributary inputs, and flows were moderate during dam removal, with no floods exceeding the 2-year flood peak. Lake Mills was drawn down using the spillway and through dam removal in increments of 3–5 m, with ‘hold’ periods of constant base level lasting from two weeks to two months in between, to allow time for delta erosion and redeposition of coarse sediments in the remaining pool to sweep across the full reservoir width before the next increment of dam lowering was implemented. Hold periods were also specified during May and June, August through the first half of September, and November and December, to provide lower sediment concentrations during periods of fish migration [77,78,79]. Dam removal was halted from October 2012 to September 2013 because sediment was overwhelming the EWTP, a significant sediment mitigation facility constructed for the project [22].

Monitoring initially focused on vertical and lateral erosion of sedimentation within the reservoir, together with turbidity in the river downstream of the dam. The U.S. Geological Survey (USGS) additionally collected discharge, turbidity, and suspended sediment samples and created a relationship between sediment concentrations and turbidity [80]. The USGS measured Elwha River turbidity at 15 min intervals below Elwha Dam (gage number 12046260, Diversion near Port Angles, WA, USA) using two different methods:

• • Formazin nephelometric units (FNUs): water, unfiltered, monochrome near infra-red LED light, 780–900 nm, detection angle 90 ± 2.5 degrees for the period 1 August 2003, to 2 June 2019 (5917 measurements)
• • Formazin Backscatter Units (FBUs): water, unfiltered, monochrome near infra-red LED light source, 780–900 nm, detection angle 0–45 degrees to incident light (backscatter), for the period 18 September 2013, to 18 December 2019 (6417 measurements)

Reclamation measured bedload with traditional methods and impact plates in the lower river to contribute to the total sediment budget [80]. Monitoring of aggradation in the channel downstream began once coarse sediment was being transported past the dam [13,22,49].

The monitoring results validated predictions of reservoir morphodynamics made using the 1994 drawdown experiment and SAFL physical model. The pilot channel excavated on the Lake Mills delta kept the eroding channel away from the valley sides for most of the dam removal period. Incision dominated during incremental lowering of the dam and reservoir, while lateral erosion, channel widening, planform metamorphosis between meandering, wandering and braiding, and delta sweeps dominated during hold periods [22,48,49].

Aggradation occurred in unconfined reaches of river downstream from the dam sites [22] and reinstated the gravel- and cobble-bed over the boulder armor/bedrock bed that developed with the dams in place. Large wood released from the reservoir sediment formed wood jams consistent with the dynamically stable pre-dam channel [22]. Sediment deposition filled deep pools, temporarily raised riffle crests by ~0.6 m and raised flood stages by ~0.6 m, which allowed the channel to reconnect to the floodplain. In addition, restoration of the sediment supply and natural wood load resulted in the channel being wider and more laterally dynamic. Elwha River active channel widths increased 40 to 50% in the middle reach (between Lake Mills and Lake Aldwell) and by 10% downstream from Elwha Dam [22,49]. These increases were not anticipated pre-dam removal since the steepness of the middle reach suggested sediment transport would dominate, but the rate of sediment arrival was so great that it aggraded the bed and induced substantial lateral channel movement and erosion that resulted in part of a road and two campsites within Olympic National Park being washed out. While gravel and cobble deposition dominated in the main channels, sand and fine sediment tended to dominate deposition within side channels. The exception to this aggradational behavior was through the eroding Lake Aldwell delta, where both bed elevations and flood stages decreased because the channel incision following dam removal was greater than the subsequent deposition from sediment supplied from Lake Mills.

Monitoring after each increment of reservoir lowering informed management decisions as to whether dam removal should proceed as planned, be slowed, or even paused to minimize sediment-related risks to downstream communities [22,81,82].

Process-based projections of how delta sediments would erode and be transported downstream [65], which were validated through monitoring during the early stages of dam removal [22], supported accurate interpretation of monitoring data used to update the MB2 model when necessary [71]. That process facilitated adaptive management of the dam removal program. For example, dam removal was paused between October 2012 and September 2013 due to sediment-related concerns. As anticipated on the basis of modeling and monitoring performed up to October 2012, this slowed the rate of delta erosion and allowed sediment already in the river to be transported downstream past the ESWI, so that upgrades to ESWI could be completed prior to the resumption of dam removal [22].

3.9. Synoptic Forecasting/Earthcasting

Once dam removal was completed, a synoptic, or “earthcasting”, approach was adopted to forecast how much of the remaining delta and prograded deposits would be eroded during subsequent floods and how long it would take for downstream sediment loads to reach a new dynamic equilibrium. This approach was selected to leverage the insights, experience and expertise gained by the engineers and scientists leading the adaptive management program.

The modern term ‘synoptic’ comes from the Latin synopticus, (literally, “seeing everything together”), but embodies the Greek philosophical argot that wisdom comes from ‘a coherent understanding of everything together’ [83]. Synoptic forecasting thus involves a structured synthesis of quantitative and qualitative information on system behavior from a wide range of viewpoints and disciplines, which bridges gaps left by other modeling approaches. While synoptic forecasts rely on theory, precedence and experience, meaning they are, to a degree, subjective [84], they are not uninformed guesswork. More recently, Ferdowsi et al. [85] encapsulated the principles of synoptic forecasting into a structured framework they call earthcasting, which is aimed specifically at better predicting earth surface process. They suggest that earthcasts focus on temporal scales of ~1 to ~100 years and spatial scales of ~1 m to ~10 km of relevance to planning; that their design involves the stakeholders likely to be impacted by the processes being predicted; and that they generate clear and testable predictions that include quantitative uncertainties. Ferdowsi et al. [85] suggest that different modeling methods can be used for different components of the model. For example, Cui and Wilcox [86] used a first-principle model to develop a two-phase forecast for gravel and fine sediment transport associated with dam removal. Sensitivity testing was performed to characterize the potential model uncertainties.

The synoptic forecasting process described in this article involves engineers and scientists applying a blend of physically based models, empirical data, and their experience (expert judgment) to create detailed sediment forecasts, which were important for the adaptive management program.

In forecasting the post-dam terrace erosion and sediment transport, the first step was to build a fixed-bed, 2D hydraulic model (SRH-2D, version 2.2) with which to simulate velocity distributions for the 2-, 10- and 100-year recurrence interval floods. To forecast velocity vectors within the former reservoir during WY 2015, model results were generated for two possible flood scenarios: first, a moderate winter with multiple floods featuring “one large peak” (Figure 4a) and second for a moderate winter with a single moderate peak (Figure 4b) [22].

Erosion locations and extents were then inferred based on velocity vectors considered in the context of results obtained earlier using the SAFL physical model for post-dam flood erosion simulations, monitoring observations, experience gained during dam removal, the sediment team’s collective professional judgment, and a number of simplifying assumptions [22]. Specific assumptions included:

• • Post-erosion, the dynamically stable longitudinal channel slope would approximate the pre-dam channel slope;
• • Once the river incised to the pre-dam valley bottom, additional downcutting would be minimal;
• • Exposed bedrock, boulders and tributary fans would continue to constrain future channel alignments, such that the alignment observed at the end of dam removal would be maintained;
• • Erosion-resistant materials in the terraces (tree stumps, cohesive sediments) would continue to limit erosion locally;
• • Where velocity vectors were parallel to silt-clay terraces, there would be limited erosion;
• • Where velocity vectors were sub-parallel to silt-clay terraces, erosion would occur at a rate that increased with the angle of flow attack;
• • All else being equal, low terraces would erode faster than high terraces;
• • The width of the active floodplain would increase through time, with the rate of lateral erosion decreasing as width increased.

Using this synoptic approach, erosion volumes of 1,500,000 m³ and 30,600 m³ were forecast for WY 2015 under the large and moderate flood scenarios, respectively. The actual erosion that occurred in 2015 totaled 1,150,000 m³ and resulted from two floods with peak discharges that were, respectively, 11% and 2% greater than the 2-year flood peak. These were the largest flood peaks since the beginning of dam removal and were expected to generate an erosion volume closer to, but within, the high-end estimate of the synoptic forecast (1,500,000 m³). The forecast therefore overestimated the actual erosion by as much as 30%. The locations of terrace bank erosion were generally correct, but the erosion extent was overestimated for some areas.

In 2015, observed lateral erosion immediately upstream from Boulder Creek more closely matched the forecast for the scenario with multiple floods including one large peak, while downstream from Boulder Creek, lateral erosion more closely matched the forecast for the moderate winter scenario including one moderate flood peak. The synoptic forecast focused on the lateral erosion of sediment terraces, rather than the vertical channel incision. However, vertical incision occurred along the upper Lake Mills delta channel and to elevations lower than expected based on extrapolation of the 1926 pre-dam topographic map, whose contours did not extend all the way to the mouth of Rica Canyon.

During dam removal, the channel erosion slope through the delta was flatter than the pre-dam channel slope because the delta sediments were finer and more erodible than those of the pre-dam channel. By the end of WY 2015, pre-dam tree stumps had been exposed, indicating the channel had recovered to its pre-dam profile and with an equilibrium slope very similar to that of the pre-dam channel.

Recognizing that by the end of 2015, the channel had recovered its pre-dam longitudinal profile, the synoptic forecast was that there would not be significant incision in WY 2016. However, 610,000 m³ of lateral terrace erosion was predicted for WY 2016 for a scenario featuring one 10 yr flood. This forecast was close to the observed volume of erosion, which was 650,000 m³, resulting from one 10-year and three 2-year floods. Both the spatial distribution and extent of lateral erosion were forecast quite accurately (Figure 4c), including the fact that erosion was less extensive than it had been in WY 2015, as was expected given that the fluvial system was evolving towards a condition of dynamic equilibrium [22].

4. Strengths and Weaknesses of the Different Elwha Models

The strengths and weaknesses of the Elwha modeling tools are evaluated in the context of the reservoir sediment erosion volume, progression of long-stream reservoir profiles, channel planform evolution, and downstream sediment concentrations. A synthesis summary of modeling tools is then provided.

4.1. Comparison of Simulated and Observed Reservoir Sediment Erosion Volumes

Simulated and observed Lake Mills delta erosion volumes both show stepwise increases over time, with modest increases driven by incision during periods of drawdown interspersed with generally larger increases predominantly due to lateral erosion during hold periods (Figure 5a). Starting at 69% of the total observed reservoir drawdown, there was a one-year hold period during which upstream-migrating incision continued unchecked, because it was incising through reservoir sediments that were finer than those present in the pre-dam river bed. This resulted in substantially more incision and delta erosion than during the equivalent stage of all simulations (Figure 5 and Figure 6), none of which simulated a one-year hold. Other than at the very upstream end of the delta, the incising channel did not encounter the more heavily armored pre-dam channel bed until after 100% drawdown, although it did encounter armored floodplain and terrace surfaces that slowed erosion at times.

Despite the SAFL model’s vertical distortion of 3.7 and other scale effects, simulated erosion volumes across its three runs were within 10% of those observed in Lake Mills for zero to 40% of full drawdown and were still within 15% at 55% of full drawdown. Between 55% and full drawdown, and despite the scale effect described below, the SAFL model delta channel left and right runs suggest that much less delta sediment would have been eroded had the pilot channel not been dredged; only 50% and 58%, respectively, of the initial delta sediments were eroded during these runs as the incising channels became attached to the model basin boundary, compared to the 94% actually observed. In the pilot channel run, the SAFL model eroded 75% of the delta, an underestimate of 19% that probably stems from a scale effect. To avoid cohesion effects, SAFL model sediments had to be coarser than required by the scaling calculations. In the pilot channel run, only 2% of the delta was eroded between 68% and full drawdown due to armor layer formation in the upper delta. Armoring resulted in the upper delta channel slope at 69% and 98% drawdown being noticeably steeper than the observed delta slope at equivalent percentage drawdowns (Figure 6a,b).

MB2 better represents the actual sediment volume and drawdown schedule than MB1, so only its results are considered here. To illustrate the difference between running the model in field monitoring and predictive modes, a subset of results is presented showing simulation of the first 40% of reservoir drawdown using field-monitored boundary conditions and the remaining 60% using boundary conditions calculated by the previous model time step. The latter significantly underestimated the extent of lateral erosion, so MB2 predicted that only 30% of the delta would be eroded at 100% drawdown, rising to 53% after the first post-removal 2-year flood. This occurred because it was assumed that coarse sediment would offer more resistance than was actually the case, so lateral erosion coefficients and exponents were set too low [71]. If the model were run for a greater proportion of time using field-monitored boundary conditions, more appropriate erosion coefficients would be selected, the MB2 plots in Figure 5 would track the measured erosion plots for a greater proportion of reservoir drawdown, and the predicted and measured erosion volumes would more closely match at 100% drawdown. Such a scenario does not alter this study’s findings regarding the utility of mass balance modeling to adaptive management decision making.

Both simulated and observed results show that little sediment (<2%) was transported past the dam during the first two-thirds of drawdown, when flows did not exceed the 2-year flood and a reservoir pool remained, albeit with shrinking capacity (Figure 5b). From two-thirds to full dam removal, MB2 predicted that ~15% of reservoir sediment would be transported past the dam site, the SAFL model predicted from 1 to 27%, while the observed amount was actually 49%.

In the SAFL runs, simulated floods following dam removal flushed 11–31% of the sediment in Lake Mills past the dam. In MB2, simulated floods following dam removal flushed 35%, but monitoring showed that the true figure as a result of post-dam removal floods was 16%. The SAFL model’s vertical distortion and coarser bed material meant there was proportionately more sediment that was too large to entrain from the delta and re-entrain from the residual pool. This scale effect may explain both the under-simulation of initial delta erosion and transport past the dam (Figure 5). MB2 included fine-grained bottomsets, but the under-estimation of lateral erosion mentioned above also caused the model to underestimate sediment transport past the dam [17].

4.2. Comparison of Simulated and Observed Long-Stream Adjustments

Long-stream profiles generated using the SAFL model pilot channel run (Figure 6a), field observation (Figure 6b), and MB2 (Figure 6c) data all show delta erosion and deposition in the reservoir starting between 0% and ~30% of total reservoir drawdown. Sediment reached Glines Canyon Dam when Lake Mills was drawn down by ~70%, and this was replicated in MB2. Sediment did not reach the SAFL model dam, however, until close to complete reservoir drawdown. This was because the absence of bottomset deposits meant there was a greater volume in the initial reservoir to fill with eroded sediment than under field conditions.

At 94% of full drawdown, significant incision was observed throughout the reservoir due to the effect of the one-year hold period, and the observed and MB2 long profiles were similar. By 98% drawdown, the SAFL model had incised close to the pre-dam long profile in the downstream half of the reservoir. In the upstream half, however, the long profile remained 7–9% higher than observed in the field, probably for two reasons. First, the armoring scale effect helped lock the long profile into a steeper configuration at an earlier stage of dam removal. Second, within the area occupied by the initial delta, the incising channel encountered the model basin boundary at a different and higher elevation location than that of the pre-dam channel.

In MB2, the channel did not incise to the pre-dam profile until after complete dam removal and passage of a 2-year flood, as was observed in the field. At that time, the MB2 and observed profiles both closely matched the pre-dam profile, except for local differences that were probably related to differences between the pre- and post-dam planform positions of the channel. In contrast, the long profile in the SAFL model still did not match that of the pre-dam channel as closely as was observed in the field and simulated using MB2, probably for the two reasons stated above.

4.3. Comparison of Planform Evolution

Natural geologic controls from tributary deltas, exposed bedrock, and pre-dam terraces were represented in the models, which improved the accuracy of simulated planform adjustments. In particular, the Boulder Creek delta (Figure 1) pushed flow towards the right half of the reservoir basin and protected the high left bank terrace downstream over a distance of two kilometers. Also represented in the SAFL model and MB2 was the process of prograded delta sediment forming high terraces (>10 m) and slowing lateral planform erosion through basal endpoint control (the balance established between the erosion and removal of failed bank material from the base of the bank and the material being supplied from the bank and the upstream channel [87]) once the reservoir was gone.

The SAFL model pilot channel run, MB2, and observed conditions all showed complex planform evolution that involved switching between single-thread and multi-thread channels at different times and locations on the delta surface (Figure 7). Stream flows during August and September 2013 (80% drawdown) were low and the incising channel was observed to narrow in response. This behavior was replicated in MB2 (Figure 7). The SAFL model did not reproduce channel narrowing, because it was run with a steady discharge. In the SAFL model, channel width varied considerably as the planform switched between being single thread during periods of rapid incision and wandering or braided when inputs of sediment from eroding terraces were high.

4.4. Fine Sediment Concentrations

MB2 predicted daily fine sediment concentrations released from the reservoirs. For the removal of Glines Canyon Dam and Lake Mills, MB2 was calibrated by adjusting the representative grain size, and thus fall velocity, so that simulated suspended sediment concentration patterns matched measured turbidities downstream from Elwha Dam during the period over which Elwha Dam was removed. During that period, most of the eroded fine sediments were released from Lake Aldwell, which had been completely drained by March 2012.

An empirical power equation was used to convert simulated suspended sediment concentrations to turbidities. This power equation was calibrated to turbidities measured by USGS downstream of Elwha Dam during the period August 2011 through April 2012. The calibrated equation coefficient was 0.6 and exponent was 1.0, which reduced the power equation to a linear equation. For the calibration period, the comparison between the simulated and measured turbidities (Figure 8) demonstrated, based on visual inspection, the capability of MB2 to match the measured turbidity patterns and orders of magnitude. This same calibration was used in other model simulations.

In April 2012, Lake Mills had been drawn down 43% and the pool was still trapping a portion of the fine sediment. Daily predictions of turbidity made using the MB2 model assumed it was solely a function of fine sediment concentration although, in reality, suspended sand also contributed to turbidity [22,80].

In practice, MB2 predicted that turbidity would be less than 20 FNUs and greater than 1000 FNUs on twice as many days as was observed in the field and underestimated the number of days that turbidity would exceed 20 and 100 FNUs. MB2 assumed that knickzones would stop migrating upstream during hold periods, when in fact they took months to migrate the 3 km to Rica Canyon (Figure 1). This protracted head cutting continued to supply substantial amounts of fine sediment downstream during the hold periods. Because MB2 underestimated lateral erosion rates for coarse, non-cohesive sediments while overestimating lateral erosion rates for fine, somewhat cohesive sediment, MB2 overestimated fine sediment concentrations during floods. Prediction of suspended sediment levels was further complicated by the fact that dam removal and reservoir drawdown (base-level lowering) decoupled concentrations from flows, thus exhibiting a wide range of values for a given flow.

4.5. Synthesis Summary of Modeling Tools

The applicability and accuracy of the modeling tools to sediment management issues are summarized in

Table 3. Applicability and accuracy of models and tools to resource management questions.

Model or Tool	Lake Mills Sediment Erosion			Sediment Transport to the Sea	Sediment Concentration Patterns and Peaks	Aggradation and Flood Stage
	Channel Incision	Channel Widening	Erosion Volume
Conceptual model	QC	QC	QC	QC	QC	QC
Reservoir drawdown experiment	A1DP	A1DP	n/a	n/a	A1DP	n/a
Mass balance modeling 1	QC	QC	AF2	n/a	AF10	n/a
Physical modeling	AF2	AF2	QC	n/a	n/a	n/a
Mass balance modeling 2	QC	QC	AF2	n/a	AF10	n/a
1D numerical modeling	QC	QI	AF2	AF2	AF10	AF10
2D numerical modeling	QC	QC	QC	n/a	n/a	n/a
Synoptic forecasting	n/a	QC	AF2	n/a	n/a	n/a

Notes: QC = Qualitatively correct; QI = Qualitatively incorrect; AF10 = Accurate to within a factor of 10; AF2 = Accurate to within a factor of 2; A1DP = Accurate to within one digit of precision; n/a = not applicable.

. Some modeling tools (e.g., reservoir drawdown experiment, mass balance models) were only applicable to the simulation of reservoir sediment erosion while other modeling tools were applicable to the reservoirs and river channel. The physical model and 2D numerical model were only utilized to simulate sediment erosion in Lake Mills, but both models could have been applied to Lake Aldwell or Elwha River sediment transport. The accuracy of the modeling tools varied: qualitative, within an order of magnitude, within a factor of 2, and within one digit of precision. The mass balance models, 1D numerical model, and synoptic forecasting were accurate to within a factor of 2.

The strengths, limitations, and appropriate uses of the modeling tools for dam removal sediment simulations are summarized in

Table 4. Strengths, limitations, and appropriate uses of modeling tools.

Model or Tool	Strengths	Limitations	Appropriate Uses
Conceptual model	Describes the important physical processes and cause and effect relationships. Qualitative predictions guide subsequent numerical model choices and interpretation of results. Not limited by numerical model capabilities or physical model scale issues.	Quantitative predictions are not provided.	Applicable to reservoir sediment erosion and redeposition, downstream transport and deposition, and environmental impacts.
Reservoir drawdown experiment	Observations and measurements of reservoir sediment erosion and redeposition can be made during the first stage of simulated dam removal.	Only provides data for the first stage of reservoir drawdown. May not be allowed if downstream sediment release were to cause significant impacts.	For cases where the sediment erosion responses to reservoir drawdown are highly uncertain.
Mass balance modeling 1	Fast simulations of reservoir sediment erosion, redeposition, and downstream release	Inability of cross-sections to define complex reservoir geometry. Channel erosion slopes and lateral coefficients must be estimated and calibrated.	For wide reservoirs trapping decades or more worth of upstream sediment supply.
Physical modeling	Simulated all the important physical processes related to coarse sediment that occurred within the real reservoir.	Scale effects prohibited the simulation of fine sediment and cohesion and made armor forming grains too coarse.	For wide reservoirs trapping decades worth of upstream sediment supply.
Mass balance modeling 2	Relatively fast simulations of reservoir sediment erosion, redeposition and downstream release	Sensitive to the estimation of channel erosion slopes and lateral coefficients.	Wide reservoirs trapping decades or more worth of upstream sediment loads.
1D numerical modeling	Simulation of sediment erosion from relatively narrow reservoirs and downstream sediment transport and deposition.	Cross-section-based model cannot simulate lateral sediment erosion or deposition processes within the reservoir or downstream channel.	Relatively narrow reservoirs trapping several years’ worth of upstream sediment loads.
2D numerical modeling	Accurate hydraulic simulations for a given discharge, channel geometry, and roughness.	Lateral sediment erosion can only be accurately simulated over a distance less than one channel width.	Simulations of hydraulics with non-erodible boundaries, vertical sediment elevation changes, and modest lateral erosion.
Synoptic forecasting	Provides estimates of future localized sediment erosion during discrete events.	Requires the measurement of previous sediment erosion during discrete events, 2D hydraulic fixed-bed modeling, and knowledge of processes.	Lateral erosion of reservoir sediment terraces.

. The strengths and limitations were unique to each category of modeling tool.

5. Discussion

Experience gained during the project identified two challenges associated with the numerical modeling of sediment entrainment, transport and deposition within the reservoir footprint during a phased dam removal with a large, complex sediment deposit. First, it proved challenging to accurately replicate the wide range of geomorphic processes that occurred in the field. Second, long run times and an inability to simulate the magnitude of lateral adjustment with a 2D model would have made it difficult to obtain accurate results quickly enough to be of use in guiding real-time, adaptive management decision making. These problems prompted the development of the mass balance model.

5.1. Simulation of Geomorphic Processes

Accurate replication of geomorphic processes in these settings was challenging for several reasons. The range of fluvial processes encountered exceeds those observed during the ‘normal’ functioning of an alluvial river, particularly when large amounts of sediment are eroded quickly and when this erosion is decoupled from larger flows. In numerical models of river channels, degradation and the downstream transport of bed sediments may create an armor layer that limits bed material transport until it is breached, which usually only occurs during a major flood. Hence, once the armor layer forms, downstream sediment transport is governed by the competence of the discharge to mobilize a certain portion of the bed armor, with the bed only becoming fully mobile at about the 2-year flood [88]. In contrast, during dam removal, the rate of sediment transport can become decoupled from the discharge such that discharges substantially smaller than the 2-year flood are able to transport large volumes of sediment. This occurs because rapid base-level lowering coupled with abrupt changes in channel pattern and pathway across the delta surface as the stream follows the shortest route to the deeper parts of the reservoir, introduce migratory knickzones that undermine the armor layer, causing it to erode and preventing it from reforming for sustained periods of time [70]. As a result, channels are able to incise more rapidly and more deeply than would be possible in a river not experiencing base-level lowering. In turn, this drives lateral erosion through a wide array of processes including meandering, topographical steering of flow by one terrace into another, braiding, avulsion, mass wasting and basal endpoint clear-out. A similar decoupling of sediment transport from discharge is reported by East et al. [43] one to two years after dam removal under transport-limited sediment supply conditions, with major geomorphic adjustments taking place during low flows and with unit stream powers ≤60 W/m².

Each increment of base-level lowering triggers a characteristic cycle of channel incision, widening and planform movement and metamorphosis. When the next increment of lowering starts before the previous cycle of channel adjustment is complete, further morphodynamic complexity ensues. The experiences of Konrad [62] and Lai [75] in attempting to simulate sediment transport and morphodynamic responses to dam removal on the Elwha River illustrate the difficulties of applying both 1D and 2D models in such environments.

These complexities are most likely to arise during the phased removal of larger dams, where the vertical fall in base level is proportionately large, the delta is much wider than the incising channel, and the great distance between the dam and the upstream boundary of the sediment delta means that eroded sediments will prograde into the shrinking reservoir pool. When the degree of vertical change is restricted and lateral adjustments are more muted, simulation using 2D or 3D process-based models may be more feasible, e.g., [76], once the incising channel has reached the pre-dam profile and a smaller number of fluvial processes are occurring, or in narrower sections of valley where lateral adjustments are more constrained.

The wide range of fluvial processes operating when a large dam is removed emphasizes the critical importance of developing a good conceptual model. Such a model not only helps to guide decision making early in dam removal design but also supports the selection and calibration of appropriate mass balance, physical and/or 1D/2D numerical models. Once the quantitative model(s) have been optimized, a good conceptual model also helps the sediment management team interpret quantitative model outputs. If a multi-dimensional model is selected, the conceptual model provides the basis for identifying which of the complex fluvial processes must be simulated, as it is unlikely the model could successfully be used ‘straight off-the-shelf’.

Even with a good conceptual understanding of the key morphodynamic processes, finessing a multi-dimensional numerical model to adequately simulate them remains challenging. Even if this can be achieved, model run times are likely to be on the order of weeks. This may be acceptable for pre-removal planning purposes, but it rules out the use of deterministic, process-based numerical modeling to inform real-time, adaptive management during dam removal, which may require a decision to be made within days, or at best a few weeks. Recourse must therefore be made to a reduced complexity approach, with a well-calibrated, mass balance model proving to be an appropriate tool for this task during removal of the Elwha River dams.

When the Elwha River dam removals were planned and permitted in the 1990s, the mass balance model MB1 (and later MB2) represented an innovative and, to an extent, experimental approach, but going forward, similar models could be developed for any dam removal where the investment of time required is merited by the level of sediment-related risk. For the Elwha River dams, MB2’s predictive capabilities with respect to sediment erosion and export rates, morphodynamics and downstream turbidity all proved useful prior to, and during, dam removal. The ability to periodically update boundary conditions in MB2 (and possibly recalibrate it) as the dam removal progresses reduces the predictive uncertainty over time.

The wide availability of powerful desktop computers now makes probabilistic modeling using a mass balance model feasible, especially if models are set running simultaneously on multiple machines. The need for probabilistic modeling reflects irreducible uncertainties that cannot be resolved in a deterministic model. For example, it would allow the sediment management team to explore a range of realistic initial values for key variables and coefficients in the model that are difficult to define precisely, e.g., representative discharge, grain size distribution, terrace erodibility, the size of drawdown increments, the length of hold periods, and issues at downstream water abstraction points. Probability distributions could be developed for each variable and thousands of simulations run to identify the most likely system responses and thus to make quantitative calculations of future risks and benefits. This would not be possible without much more powerful, or even quantum, computers when using complex numerical models to simulate dam removal because of the long simulation times. Adaptive management is well suited for phased dam removals, as the intervals between drawdown increments provide the time required to modify multiple aspects of the adopted sediment management strategy, e.g., the duration of hold periods. The key requirement is to quickly and accurately predict the volume, caliber and rate at which sediment enters the downstream system and for three practical reasons a mass balance model may currently be the best tool for doing this. First, model simulations can be set up, run and interpreted within a single day, which matches the time scale needed for adaptive management reviews and decision making. Second, these short run times allow the model’s delta surface geometry to be updated in real time with field-monitored data defining current channel slopes, bank lines, and planform patterns. This significantly improves model accuracy and provides better understanding of what to expect during subsequent time steps, with or without an adaptive change to the removal plan or sediment management strategy. Third, once lateral erosion coefficients for coarse and fine sediments have been calibrated using monitoring data, erosion rates and volumes can be predicted more accurately, thus increasing confidence that the removal plan will perform as designed or that it requires modification. Identifying appropriate values for model coefficients again emphasizes the need for a good conceptual model that correctly identifies all the major process–form interactions likely to occur, coupled with field measurements of actual erosion made during the early phases of dam removal. Conceptual modeling and a field monitoring program that start prior to dam removal and continue during dam removal are thus vital for successfully utilizing a mass balance model to support adaptive management decision making. Significant monitoring elements that contributed to the success of the Elwha adaptive management modeling included near real-time structure-for-motion photogrammetry [74], discharge and turbidity data, web cameras, reservoir stage, and field inspections of reservoir landscape evolution [49].

Experience gained at the Elwha River demonstrates the value of physical modeling. For example, the decision to excavate the pilot channel would not have been made without the palpable evidence of its effectiveness that only a physical model like that operated at SAFL could provide. That said, caution must be exercised when interpreting quantitative system behavior based on the outcomes of physical models, due to unavoidable scale effects. In the SAFL model, disproportionately large, armor-forming clasts influenced the evolution of the incising channel’s long profile and, hence, the absolute amounts of vertical and lateral erosion. Inaccurate boundary conditions, such as the absence of fine-grained delta bottomsets, also led to differences in system behavior in the model when compared to what actually happened. Even if bottomsets had been included in the SAFL model, the impossibility of accurately scaling cohesion when model scale exceeds 1:1 would have precluded perfect replication of the influence of fine sediments on the morphodynamics observed in Lake Mills.

Accurately assessing the cohesive strength of fine sediment field samples is itself challenging. Tests of samples from Lakes Mills showed no significant cohesion, but they contained just enough clay and organic material to ensure that flows running parallel to fine sediment banks and terraces generally eroded them less than three meters, which resulted in the channel remaining narrower than expected. Conversely, when the flow struck the same type of sediments at an angle, more extensive erosion occurred. This suggests that the laboratory test for cohesion, which is designed to identify issues associated with the loadings imposed by buildings or roads, may be inappropriate for fluvial applications. These difficulties notwithstanding, the qualitative process similarities of the physical model and drawdown experiment compared to those of the full dam removal demonstrate the vital role these studies play in informing conceptual modeling and the formulation of numerical and mass balance models.

Incoming sediment from the upper Elwha River watershed was accounted for in the MB1 and MB2 models, but in the actual dam, removal was initially inconsequential because large floods did not occur until after dam removal was complete. One boundary condition not addressed in any of the Elwha models was sediment inputs from tributaries, which also incised during dam removal. This omission was inconsequential because the Elwha tributaries are much smaller than the mainstream river and their lateral evolution lags behind that of the mainstem. As a result, they contributed only about 1.5% of the sediment trapped in the reservoir. However, in reservoirs where tributaries have provided a significant proportion of accumulated sediment, they should be considered during sediment management planning and adaptation. Where tributary deltas have the potential to control patterns of erosion and redistribution of sediments during dam removal, they should also be included, as was done for Boulder Creek in this study. The downstream boundary condition for tributary models would be the receding reservoir water surface elevation, and fluvial processes and morphodynamics could be simulated using either a 1D or a 2D model as appropriate for the degree to which the tributary is laterally confined.

5.2. Remaining Knowledge Gaps and Future Directions

The purely numerical models applied to Lake Mills had difficulty quantitatively predicting the lateral erosion of exposed reservoir terraces. The 2D hydraulic and sediment model by Lai [75,76] attempts to simulate the physical processes, but more development of this, or similar models, is needed for lateral erosion occurring over multiple channel widths. A new method to measure the erosion resistance of these fine-grained sediment terraces is also required.

While the criteria for identifying recovery of the Elwha system towards a condition of dynamic equilibrium were confirmed by subsequent assessments to be accurate (Table 2), more recent studies continue to add detail to our understanding of post-system recovery. A common theme emerging from several studies is that different aspects of ecosystem process and form evolve towards equilibrium over different timescales. Ely et al. [45] found that the grain size distribution, braiding index and number of large wood jams (LWJs) rapidly increased following dam removal and then stabilized, while channel sinuosity and LWJ area were still evolving towards a dynamic equilibrium 6–10 years after dam removal was complete. Buscombe et al. [41] documented similar lags between the peak in LWJ deposition occurring in 2013 and the subsequent growth in size of associated bars reaching a maximum in 2017. McCaffery et al. [47] concluded that recovery can start early for physical processes such as flow, sediment, and temperature regimes and for short-lived animals like invertebrates, while riparian communities and fish populations can take longer to recover. Long-term monitoring programs are therefore needed to document channel and floodplain evolution after dam removal and biological processes such as riparian vegetation growth, macroinvertebrate development, and fish migration and spawning.

The studies by Buscombe et al. [41] and Ely et al. [45] are noteworthy in utilizing high resolution orthoimages, machine learning and terrestrial LiDAR to collect monitoring data. Such techniques offer considerable potential to rapidly and accurately track and quantitatively analyze system response both during and for long periods of time after dam removal. High-resolution LiDAR, for example, has been used to assess variations in riparian zone canopy height, number of trees and species composition [89], so could potentially be used to track the evolution of post restoration riparian woodlands towards maturity. Continued research is needed to improve the integration and speed of high-resolution remote-sensing tools so they can be more widely used.

6. Conclusions

Experienced gained over the three decades during which removal of the Elwha River dams was planned and executed, coupled with the results of monitoring restoration of the Elwha River and its sediment system, has provided useful insights into both the complex and protracted evolution of deltas in the reservoirs of large, incrementally removed dams, and into the nature, rate and duration of elevated sediment concentrations, bed profile adjustments, and planform changes in the river downstream. The work has highlighted the relative strengths and weaknesses of conceptual, numerical, physical, and reduced complexity mass balance models in simulating the complex morphodynamic behavior of river systems impacted by large dam removals, and of using multiple models in a complimentary fashion to provide the best possible sediment management advice. Also emerging from this work is an appreciation of the value of synoptic modeling, which provides a framework for synthesizing data and experience on system attributes and behaviors from a range of models and disciplinary viewpoints. It is hoped that the experiences reported in this paper will be useful to others tasked with managing the sediment-related uncertainty associated with other high-risk dam-removal projects.