The DFNP east of Austria's capital Vienna (Figure 1) is, with 96 km², the largest natural floodplain landscape of its kind in central Europe, where the Danube River still has the dynamic character of a mountain river. To ensure navigation and reduce flood hazards, extensive channelization, regulation through rip-raps and groins, as well as active disconnection of the floodplain through levees, have been conducted since the nineteenth century. Additionally, heavy damming since the 1950s has further impacted natural dynamics, with the DFNP being the longest remaining free flowing stretch of the Austrian Danube (36 km) (Habersack et al., 2016).

Starting in 1996, the first side arm in Haslau-Regelsbrunn (3) was partially reconnected, followed by two floodplain systems in Orth (2) and Schönau (1) 5 years later, connected at lower water levels but still in a technical manner (Figure 1, Supplementary Figure 1, Supplementary Table 3). To increase natural dynamics while ensuring navigation, a pilot project by the Austrian waterway administration (viadonau, http://www.viadonau.org) (2012–2014) successfully tested a combination of river engineering measures including the complete reconnection of a side arm (Johler Arm−5), optimization of groins, removal of rip-raps, and river bed improvement to counteract riverbed incision. The remaining side arms (Petronell−4, Spittelauer Arm−6, Roethelstein−7) are solely groundwater fed until the Danube discharge reaches >2,600 m³ s⁻¹, well above mean discharge (MQ).

We consider these seven side arm systems in the DFNP with different degrees of connectivity to the Danube flow regime (Table 1) and different sizes (here as length of branch and maximum flooded area) to model retention of nitrate and total phosphorus for wet, dry, and average hydraulic years. The considered river stretch belongs to the Upper Danube Catchment with a nival flow regime influenced by snow melt from the Alps. Hainburg is the representative gauge for the study site. Here, the catchment has a size of 104,727 km² and the MQ is 1,930 m³ s⁻¹. At the study site, the Danube River transports between about 119,015 t (dry year 2003) and 199,000 t (wet year 2002) of dissolved inorganic nitrogen (DIN), with NO₃-N being the main component. The annual load of TP ranges between 2,699 t (dry year 2003) and 6,700 t (wet year 2002) (ICPDR, 2002, 2003).

To detect the effect of river restoration, nutrient retention was modeled in the current state (CUR) and in a scenario of complete implementation of all proposed restoration measures (ALL), according to the management and restoration plan of the Austrian waterway authority. This scenario includes the complete reconnection of seven side arms and the maintenance of stable riverine water levels due to sediment addition to the river channel. CUR and ALL were calculated for a hydrograph of an extremely wet (2002) and an extremely dry year (2003) as defined by Weigelhofer et al. (2015).

Hydrological variables of the Danube were obtained by the hydrological portal of Austria (https://ehyd.gv.at/, last download 15th of July 2019) and water quality data on nutrients from the federal ministry of sustainability and tourism database (https://wasser.umweltbundesamt.at/h2odb/, last download 15th of July 2019). Water quality data in the side arms was gathered from monitoring and research projects of past reconnection measures in the systems of Regelsbrunn (2) (Heiler et al., 1995; Hein et al., 1999, 2002), Orth (2) (Hein et al., 2003), and Johler Arm (5) (Viadonau, 2015). Water quality was measured at discharges of the Danube River up to 5,130 m³ s⁻¹ in the vicinity of the respective inlets and outlets. To ensure comparability, we included only data sets where sampling and nutrient analysis were conducted with standardized methods. The degree and duration of surface water connection of the side arms was approximated in all scenarios by linearly interpolating the flow between the topographical point of inflow and a discharge of 5,130 m³ s⁻¹. The latter value was obtained from the simulation output of a one-dimensional hydraulic model (HEC-RAS).

Information on DFNP borders was obtained from EEA (2018, https://www.eea.europa.eu/data-and-maps/data/natura-10/natura-2000-spatial-data). Wetted area and water depth at three discharges were obtained by intersecting surface and groundwater tables with the DEM (Hohensinner et al., 2008). For higher discharges, a flood hazard map of a frequent HQ30 flood (9,290 m³ s⁻¹) for the DFNP (https://maps.wisa.bmnt.gv.at/hochwasser, last download 15th of July 2019) was georeferenced and intersected with DFNP borders, showing that the entire floodplain is inundated at this discharge. To estimate the extent of each side arm system, we delineated watersheds on the basis of a digital elevation 1 m grid (provided by viadonau, tiles 78, 89, and 81) using the ESRI ArcGIS “Hydrology” tool (Figure 1). Shortwave radiation data on a monthly mean basis was downloaded from EUMETSAT (http://www.cmsaf.eu/pum) to consider additional nitrate uptake by macrophytes following the method described in Venohr et al. (2011).

Mass balances for the statistical model were deduced from concentration measurements and the hydraulic model in investigated side arms. During surface water connection, longitudinal concentration differences (dc, in g m⁻³ m⁻¹) were calculated (Equation 2), and during isolation, changes in concentration over time (Δc, in g m⁻³ d⁻¹) (Equation 3). Therefore, the available dataset was divided accordingly, resulting in pairwise observations in upstream and downstream parts of the side arm in the connected state (n = 54) or uninterrupted periods between measurements in the disconnected state (n = 125).

Statistical models were developed for TP and NO₃-N, each following independent routines on the basis of concentration measurements in all three side arms. For TP, identified significant relationships (Figure 2) were used to calculate nutrient retention by applying a multiple adaptive spline regression model (Millborow, 2015). However, because dc and Δc differed considerably between the three side arm systems (Figure 2, Supplementary Table 1), three different connectivity classes were assigned (Supplementary Table 2). The connectivity classes, indicating three different TP retention potentials, were used to extrapolate nutrient retention to side arms without concentration measurements based on the average surface water connection (d a⁻¹) and the ratio of the inlet opening to the side arm width (Table 1). Assuming that the maximum retention equals the load entering the side arm, we modeled the longitudinal TP concentration decline as a logistic growth function (Equation 1). TP input from the Danube River was computed according to an adapted exponential discharge-dependent TP concentration model from Zessner and Kroiss (1999). During periods of isolation, concentration differences were not significantly different from zero and therefore not considered in further calculations for TP (one-sample Wilcox-test, p > 0.1).

For nitrates, no correlation with any measure-specific parameters could be detected, but retention values significantly exceeded zero (one-sample Wilcox-test, p < 0.01, n = 179). Therefore, mean retention values in connected (9 × 10⁻⁷ g m⁻³ m⁻¹, n = 54) and isolated (718 × 10⁻⁵ g m⁻³ d⁻¹, n = 125) conditions were used to calculate retention. By multiplying the resulting TP and NO₃-N retention rates (g m⁻² d⁻¹) by the factor 10 and the days under consideration, commonly used rates in kg ha⁻¹ month⁻¹ or kg ha⁻¹ yr⁻¹ were calculated.

dc = concentration difference of TP between inlet and outlet [g m⁻³]

dc/m = system-specific retention rate per m [g m⁻³ m⁻¹]

cin = inlet concentration [g m⁻³]

a = concentration-dependent retention rate (connectivity class) [–]

L = length of the side arm [m]

During connection:

During isolation:

Ret = retention [g m⁻² d⁻¹]

Q = discharge in side arm [m³ d⁻¹]

dc = concentration difference between inlet and outlet [g m⁻³]

Δc = concentration change over time [g m⁻³d⁻¹]

A = area [m²]

d = water depth [m].

This semi-empirical mesoscale model is based on two submodels. The first submodel calculates incoming load and inundated floodplain area considering the statistical relationships between the given discharge ratios and the incoming discharge and inundated area on a daily basis for each side arm system (Natho et al., 2013; Natho and Venohr, 2014). With this data given, the nutrient retention was calculated for nitrates and total phosphorus on a monthly basis for each side arm system following the approach of Venohr et al. (2011).

The incoming nitrate and total phosphorus load was calculated on a monthly resolution according to ICPDR (2017). On the basis of the discharge-flooded area relationship, three parameter sigmoidal functions were derived (SIGMAPLOT version 13.0 Notebook, Systat Software Inc.) to describe the proceeding inundation of floodplains of each side arm system with increasing discharge (Supplementary Figure 2). This key parameter represents the degree of lateral hydrological connectivity of each side arm system and was developed successfully for rivers in Germany (Natho and Venohr, 2012, 2014; Natho et al., 2013).

As flood waves normally occur over a short period, the incoming nutrient load was calculated on a daily basis with daily average discharge. In line with seasonal dependencies of NO₃-N concentrations found by Zweimüller et al. (2008), Venohr (2006), and Venohr et al. (2011) and complex patterns of TP fluxes during high discharges (Zessner et al., 2005; Bowes et al., 2008; Chen et al., 2013), concentration proxies for NO₃-N and TP were applied (Supplementary Tables 4, 5). The incoming daily load was calculated with an average nitrate concentration for each month and with an average TP concentration considering winter and summer seasons and discharge levels below or above MQ. Based on the daily loads and the known inundated area, the retention model (Equations 4, 5) developed on the mesoscale was applied at a monthly resolution (Venohr et al., 2011):

Nret% = retention in % of incoming nutrient load [%]

R = short wave radiation [W m⁻²]

T = temperature [°C]

HL = hydraulic load [m yr⁻¹]

a = 5.7 [–], a₁ = 0.025 [–], and b = 0.067 [–]

Pret% = retention in % of incoming nutrient load [%]

HL = hydraulic load [m yr⁻¹]

c = 15.91 [–]

Both Equations (4) and (5) consider hydraulic load (HL) as an equivalent for retention time (see Venohr, 2006 for details), meaning lower areal retention in case of high HL. For nitrogen, temperature and shortwave radiation have also influenced the metabolic rate of denitrification and plant uptake (Venohr et al., 2011).

Due to the dependence on connectivity in both models (Figure 3), reconnection of all side arms (ALL scenario) was predicted to considerably improve nutrient retention capacity in the DFNP. According to both models, the calculated absolute NO₃-N retention in the DFNP doubled in the wet year and even rose by a factor of 3.75–5 in the dry year (Figure 5). For TP retention, the trend was similar to NO₃-N according to the semi-empirical model. For the statistical model, TP retention after restoration was higher than the NO₃-N retention in absolute terms. Generally, the reconnection effect was most visible for large side arms with low connectivity under current conditions [Regelsbrunn (3), Petronell (4), and Spittelauer Arm (6); Figure 6]. In the current state, NO₃-N and TP load reduction in the floodplains accounted for between a marginal 0.01 and 0.03% for both models and hydrographs. Considering the complete implementation of all restoration measures, load reduction may increase to 0.05–0.07% according to the semi-empirical model. The statistical model resulted in 0.04–0.06% reduction of NO₃-N load and an unlikely high 1.8–3% reduction of TP load in the dry and wet hydrographs, respectively.

In the wet year 2002, the hydrograph exceeded the highest navigable discharge (HSQ) at 14 days (March, August and November), of which 12 days exceeded an annual flood. The semi-empirical model was found to adequately illustrate these events (Table 3). By including these events, the annual retention (t a⁻¹) increased by a factor of 1.6 for TP (0.04% river load reduction) and even 1.9 for NO₃-N (0.06% reduction) compared to discharges <5,130 m³ s⁻¹. However, if all side arms were completely reconnected (ALL scenario), the annual retention <5,130 m³ s⁻¹ would still exceed the current state including flooding events by a factor of 1.14 for both TP and NO₃-N. If the 14 flooding events were then included in the ALL scenario, the factor would even rise to 1.9 (0.07% river load reduction) and 2.1 (0.1% river load reduction) for TP and NO₃-N, respectively, compared to the current state. This finding supports once more the importance of frequent inundations for the efficient removal of nutrients from floodplains.

A detailed model validation in terms of observable nutrient retention was not possible because there was no data available suitable for this purpose. Instead, we compared our model results in selected systems and with the literature values presented in Supplementary Tables 6, 7. Whereas, for TP our model results (CUR: 1.4–5.1 kg P ha⁻¹ yr⁻¹) lie rather in the lower to mean range of values, our calculated NO₃-N retention (CUR: 65.5–114.5 kg NO₃-N ha⁻¹ yr⁻¹) is in the mid- to upper range of values found. The variation in published areal retention rates is great, even though only floodplain systems in temperate areas were selected. Reasons for this may be attributed to differences in floodplain size, hydrological conditions, nutrient loading, vegetation type, sampling design, or retention mechanisms (Fisher and Acreman, 2004). In the case of TP, retention is partly a temporary process of sedimentation (Fisher and Acreman, 2004; Venterink et al., 2006) and a highly variable transport with increasing discharge. Noe et al. (2019) found positive net retention when sedimentation inputs of P were high and leaching rates were low, e.g., processes described by Hoffmann et al. (2012) and (Schönbrunner et al., 2012).

The results of both models clearly show that the effect of restoration in the study region is the improvement of lateral hydrological connectivity and thus more frequent inundation, so-called “hot moments” (McClain et al., 2003). In our models, more frequent and thus higher nutrient input into floodplains leads to higher absolute nutrient retention (Newcomer Johnson et al., 2016; Noe et al., 2019) because of higher accumulation rates of N and P (Noe and Hupp, 2005) and greater removal of reactive nitrogen with increasing load (Jordan et al., 2011). Frequent inundations with less destructive power are important, because N removal efficiency (%) is reported to decrease with higher N load (Fisher and Acreman, 2004; Bernot and Dodds, 2005; Venohr, 2006; Jordan et al., 2011). Similar findings, though less pronounced, are described for TP retention (%) in wetlands (Reddy et al., 1999; Fisher and Acreman, 2004). Because of its ability to be released, not only with increasing incoming load, but with increasing duration of input, phosphorus retention efficiency in floodplains may decline (Fisher and Acreman, 2004).

In combination with connectivity and nutrient load inputs, Newcomer Johnson et al. (2016) showed in an extensive review that nutrient retention in floodplains could be enhanced through restoration, leading to increased water surface area and increased hydraulic residence time. In both models, discharge is a determinant for inundated area, days of inundation and incoming nutrient load. The highest retention in the statistical model is found in frequently connected areas with high nutrient load inputs, following rising patterns of the hydrograph. On the other hand, the highest retention rates found by the semi-empirical model occur in flat topographies with high input loads [e.g., Regelsbrunn (3)], where area increases disproportionately faster compared to the discharge, hence reducing HL and flow velocity. Maaß and Schüttrumpf (2019) also model more effective sedimentation in restored, lowered floodplains than in restored, elevated rivers. Saunders and Kalff (2001) and Craig et al. (2008) confirm that low flow velocities increase retention through longer water-sediment contact time, which is responsible for denitrification and sedimentation. Under current conditions (CUR) inundation happens less frequently, but if it does happen, the area is larger on average. In contrast, under higher connectivity levels, inundation happens more frequently, leading to reduced average areas but longer inundation periods of far more days. This indicates the importance of small inundation events with low HL for effective nutrient retention and further underlines the importance of improved lateral hydrological connectivity of the whole floodplain, not only selected side arms, as the main goal of floodplain restoration. Furthermore, prioritizing reconnection of larger side arm systems creates synergies of high retention potential with the provision of habitat for protected and endangered rheophilic communities (Funk et al., 2013, 2019).

In summary, increasing connectivity and its consequences for retention is estimated differently by the two models, but generally leads to higher retention, in answer to our second research question.

We considered a total of 3,200 ha of floodplains that can be inundated with a probability of once every 30 years, making up one-third of the DFNP. This is the lowest probability found in Flood Hazard Maps showing inundation of all the floodplains already considered. With an average inundation area of 481 ha (wet) and 282 (dry), which we would call a very well-connected floodplain, the semi-empirical model calculated for the ALL scenario a river load reduction of 0.07–0.1% for NO₃-N and 0.05–0.07% for TP. This seems to be very little and would not be detectable in the river. However, we analyzed a stretch of only 30 km and argue that the floodplain area is very small compared to the river size. Thus, we ask the question: How much well-connected floodplain area would be needed to achieve a nutrient retention of 1% of the yearly transported river load? A one-percent decrease in river load would already be detectable and within an achievable range in the light of our modeled retention efficiencies.

From our model results, we can conclude an areal retention rate of up to 334 (dry)−400 (wet) kg NO₃-N ha⁻¹ yr⁻¹ and 7 (dry)−9 (wet) kg TP ha⁻¹ yr⁻¹. A river load of 199,000 t dissolved inorganic nitrogen and 6,700 t TP is reported by ICPDR (2002). To achieve a retention of 1% of the river load in a wet year, we would need 4,975 ha for NO₃-N and 7,444 ha for TP. As we stated above, only 15% of the DFNP is flooded on a regularly basis, which would then result in a floodplain area 10.4 and 15.5 times larger than the DFNP. Alternatively, on a 100 km river stretch we would need a floodplain that is 0.5 km wide on each side for NO₃-N and 0.75 km for TP. In the dry year, the Danube transported 119,015 t DIN and 2,699 t TP (ICPDR, 2003). The area required to remove 1% of this load is 3,563 ha for NO₃-N and 3,856 ha for TP. Considering an average inundated area of only 9%, an area of about 12.6 and 13.7 times the DFNP is required for a 1% removal of NO₃-N and TP.

We suggest increasing reconnection not only through side arm connection but also by allowing overbank flow, which inundates the entire floodplain complex during frequent floods with lower water levels (annual floods). This increases river-floodplain interactions and might lead to higher sedimentation and phosphorus and N-enriched sediment deposition (Maaß and Schüttrumpf, 2019; Noe et al., 2019). Despite the clear evidence that the re-establishment of hydrological pulsing increases denitrification and decreases emissions of greenhouse gases (Mitsch et al., 2008; Welti et al., 2012), we are aware that nutrient retention in restored floodplain areas is a complex topic and that the potential for phosphate release from soils of floodplains with restored wetland hydrology is a large concern. Although some PO₄-P release processes might occur after floodplain reconnection (Figure 2; Aldous et al., 2007; Schönbrunner et al., 2012), over the long run positive net retention after restoration will appear (Owens and Walling, 2002; Noe and Hupp, 2005; Venterink et al., 2006; Wassen and Olde Venterink, 2006; Anderson and Mitsch, 2007; Noe et al., 2019).

Based on these considerations, we can answer our third research question and confirm that the measures planned in the ALL scenario will not be enough to retain a significant amount of nutrients (1% threshold). Only when reconnection at low flow conditions leads to very frequent inundation of the floodplain area as a whole could a significant amount of river nutrient load be retained in the DNFP. Finally, knowing that floodplain restoration may increase the multi-functionality of these ecosystems worthy of protection, it should still be our major goal to manage water quality primarily through reduction of punctual and diffuse nutrient emissions into river systems rather than filtering out pollutants afterwards.