3.1. Rain Garden Network Response to Storm Events
The rain garden network buffered more than half of the inputs of stormwater volume from the combined sewer system (Table 2
). Our measured effectiveness compared well with that determined by other workers in different infiltration-type stormwater control measures (e.g., 52% in a monitored rain garden [17
] and 49% for a grass filter strip treating highway runoff [18
]). Our results also fall into the range of effectiveness determined by Hatt et al. [19
], which was 15–83% for three biofiltration installations with soil and mulch profiles similar to the present study; and the higher end of the range (13 to 62%) determined by Autixier et al. in a hydrologic simulation study [20
Precipitation in 2012 was the lowest study-wide at 500 mm, with greater average number dry days between events (5.2 days) longer than the four-year average (3.7 days). The frequency and volume of flows to the CSS were correspondingly lowest in 2012 with 95% effectiveness. During this time unmonitored bypass flow moved through the rain garden network subsurface drainage plumbing, which would have otherwise registered as high network effectiveness to (Table 2
). There was nearly twice the depth of rainfall in 2013 (940 mm) than 2012, but spread out over a much longer period (409 h). Flow volumes to the CSS was higher in 2013 due to bypass of upper rain garden flows to the CSS via the activation of the original parking lot drainage system, which remained unsealed through the 2013 warm season. By 2014, repairs and modifications to infrastructure were complete. The rainfall depth in 2014 was moderate at 680 mm, and distributed over the longest study-wide duration of 552 h (Table 2
), and we observed that 48% of the total volume inputs (inlet flow, direct rainfall, and runoff from adjacent turf landscapes into the gardens) did not enter the MSDGC CSS (Table 3
). Finally, the more intense rainfall pattern in 2015 had the same total rainfall depth as in 2013 (940 mm), though delivered in approximately half the time (238 compared to 409 h), with the shortest average number of days between storms (2.7 compared to an average of 3.7 days).
Analysis of storm hydrographs showed that any flow into the combined sewer systems was delayed by 4.5 h in the construction phase (2012–2013), and by 5.5 h in the operational phase (2014–2015). This rain garden function is a benefit to maintaining capacity in the combined sewer conveyance system. Based on the typology of events driving flows to the CSS, and accounting for 2012 and 2013 overestimation of retention, the frequency of events was similar across years (Table 3
). Large flows to the CSS were more variable in the construction phase, as much lower (2012), or higher (2013), compared to the overall 4-year average (Table 3
). As the system stabilized in the 2014–2015 operational phase, the volume retention of the rain garden network averaged slightly greater than 50%, with nearly 90% of all warm-season events fully retained. Approximately 5% (averaged over the four-year monitoring period) of all events produced either threshold or large flows to the combined sewer system (Table 3
The hydrologic effectiveness of a rain garden is dependent on the amounts and timing of runoff volume delivered, transmitted, and stored in the different layers of each rain garden. We expect that stormwater volumes conveyed to the sewer system are most strongly correlated with the primary precipitation and rain garden hydrologic processes regulating them. Since we did not control for variables experimentally, we cannot make firm claims of a causal relationship between the correlated variables and stormwater regulation. We used multivariate analysis (principal components analysis, PCA) to qualify the potentially interactive role of measured fluxes to assess the hydrology of this network in the construction and operational phases. The PCA approach equally weights individual observations, and simultaneously places and scales the different hydrologic fluxes along independent, explanatory principal components, delineating fluxes that operate either in concert or independently of each other. The first two principal components explain 77 percent (Figure 3
A; construction phase 2012–2013) and 73 percent (Figure 3
B; operational phase 2014–2015) of total variance in the dataset, with re-scaled data evenly distributed around the space formed by the components. PCA Factor 1 (x-axis; Figure 3
A) shows that for both construction and operational phases, total event rainfall depth was coincident with flows into, through, and out of the network. PCA Factor 2 shows the expected, inverse relationship between average event rainfall intensity (Rint2) and duration (Rdur2), with these metrics ordinated independent of Factor 1 hydrologic processes. Evapotranspiration losses (Etloss2, Factor 2) are indicated to play a more predominant role in the operational than construction phase. With the exception of areas immediately around the inlets, the initial plantings quickly established in the construction phase from 2011–2012, with canopy coverage in 2016 (4 years in to the operational phase) estimated at 97%. Although the thick surface mulch layer likely restricted evaporative loss, the amount of transpiration may have increased due to increased vegetative cover and presumably increased removal of soil moisture through likewise expanded root systems. Our analysis suggests that total event rainfall depth (an input) and evapotranspiration (a loss) are primary factors regulating flows through the rain garden network.
We found that events with the largest flows to the combined sewer system had high total rainfall depth delivered over longer durations (i.e., 24 h). This suggested that average event intensity was not as important as total event rainfall depth. We note here that the network retained higher-intensity events (~0.7 cm h−1
), which for a different study, was a threshold intensity for driving a rain garden into overflow [17
]. Overall, the frequency of the largest outflow events was minimized and retention efficiency maximized by ongoing improvements to the network conveyances. These served to push the threshold for outflow higher; for example, flow into the CSS would occur only after the water surface exceeded the lower garden standpipe invert.
We next analyzed how these different factors synthesize to create conditions for overflow to the CSS. In Figure 4
, factor scores for events are plotted pairwise (Factor 1, Factor 2) against event outflow volume. For events when system capacity is never exceeded and stormwater volume inputs are stored, factor scores for both Factor 1 and Factor 2 are small. This is related to overall low values of total event rainfall depth, low inflow volume, and an absence of flow moving through the network (Figure 4
). As Factor 1 scores (as rain event depth) increase, flow through the network is initiated, which activates threshold or larger categories of outflow volume to the CSS (Figure 4
), and evapotranspiration losses are proportionally less important in the regulation of these outflows. For the largest rainfall events (some of which may also have high intensity), outflow volumes are considerable, ranging from 10,400 to 338,300 L (Figure 4
), and differences between Factors 1 and 2 are maximized. The only significant loss from the network under the largest stormflows is outflow to the CSS. Factor 2 scores are highly variable, which we attribute to event-wise differences in evapotranspiration losses, which remove soil moisture from the rooting zone over the course of days, whereas an increase Factor 1 values indicates rapid influx of stormwater volume occurring at the scale of minutes or hours.
An anomaly in this otherwise consistent set of relationships summarized by the PCA factors is that both factor scores are both highly positive for two extreme storm events, namely, 1 September 2012 (0.7 cm in 0.2 h, average intensity 4.1 cm h−1, 5 days between storms, flow into upper garden only, no flow out to the CSS); then 5 May 2013 (6.3 cm in 25.2 h, average intensity 0.2 cm h−1, 10 days between storms, flow into upper garden and lower garden, ~190, 600 L outflow to the CSS). Monitoring data for each event showed that inlet volume (2012 event) and then outflow volumes (2013 event) exceeded measured volumes in the rain garden network. Both events occurred prior to any modification or repair of the network drainage system. Although rare in the study record (2 out of 233 events over a four-year period), the impacts of each event are considerable in different ways, and tested the operational capacities of the network under drier antecedent conditions. Additional flow sources may have been initiated by early-term infiltration-excess runoff production due to high rainfall intensity in the 2012 event; and by possible expansion of source areas and overall complete saturation and inundation of the network for the longer-duration 2013 event.
3.2. Soil Hydrology
Compared with high runoff potential in the surrounding turf areas, the rain garden system offered four-times greater infiltration rate, and 100 times greater internal drainage rate. The infiltration rate of the surrounding turf areas (clay loam soils) is 0.6 ± 0.1 cm h−1
, which falls into the lower end of the range determined on residential lawns in central Pennsylvania [21
]. The design infiltration rate of engineered soil fill material for this network is 5 cm h−1
. Infiltration rates measured at the mulch layer surface were consistently (2012–2015), which is about half of the assumed design value at 2.2 ± 0.4 cm h−1
, and 2.0 ± 0.5 cm h−1
for upper and lower gardens, respectively (Figure 5
). In the period 2014–2016, the upper garden loamy sand exhibited a mean infiltration rate that was overall highest and most variable (in 2016 at 12.9 ± 1.7 cm h−1
); whereas the lower garden sandy loam had a mean infiltration rate of approximately 2 cm h−1
that did not vary considerably over the study period (Figure 5
). Although measured infiltration was less than design rates, it compares well to the lower-permeability biofilter soils modeled by Le Coustumer et al. [22
], which found that when biofilter infiltration rates were initially low (approx. 2 cm h−1
), these rates tended to not change over time. In addition, our near-saturated measurement condition eliminated the flow contribution of macropores exceeding about 0.5 mm in diameter (e.g., space between mulch particles), and so would likely register conductivity less than measurements made under saturated conditions.
Once water saturates the Oi and Bw layers, drainage rate regulates drawdown and movement of water into the gravel layer. At the inception of the study in 2012 the average drainage rate at saturation (Ksat
) was unrestricted in the upper rain garden (due to a highly-permeable unstructured loamy sand), and 20.3 ± 1.3 cm h−1
for the lower rain garden. By 2015, drainage rate was both higher and more variable in the upper rain garden (75.8 ± 28.3 cm h−1
) than the lower garden, which had declined to 4.2 ± 0.3 cm h−1
. Clay loam soils under turf areas were slightly more compact at an average of 1.35 ± 0.05 g cm−3
compared to rain garden soils. For both the upper and lower rain gardens, the organic and mineral layers were, on average, less compact in 2016 than 2015 (Table 4
). Bulk density values are in agreement with those for samples with correspondent soil textural classes (loamy sand, sandy loam), as measured by Asleson et al. [7
] in their survey of 12 Minnesota (USA) rain gardens. There was no significant regression relationship between bulk density and sampling distance from the inlets for either garden in the 2015–16 period (p
Soil profiles developed over time, and the stratification of the surface mulch layer was similar for both gardens. Serial, bi-annual mulching (2012, 2014) led to the development of a pronounced organic horizon in both gardens, which we attributed to the hardwood-chip mulch composting in place (Figure 2
). Over the ensuing six years since construction, the surface horizons in both gardens stratified into the coarse, newer mulch layer that comprises the Oi horizon, which transitioned to the finer, older layer of organic matter that define an Oa horizon (Figure 2
). By 2016, the total organic layer thickness ranged from 4 to 13 cm, and 7 to 25 cm in the upper, and lower gardens, respectively. Due to inlet stormflow, mulch was redistributed in each garden, and especially in the upper garden so that combined depth of the Oi, Oa horizons varied widely; ranging from about 8 cm (closer to the inlets) to 25 cm (farther from the inlets), indicating that mulch was pushed outwards from points nearest the flow outlet that communicated flow into the rain garden. Further profile development occurred in the mineral horizons, but only for the lower rain garden. By 2014, a thin A horizon had developed in the lower rain garden (Figure 2
), which may have contributed to keeping infiltration rates consistent in the lower garden mineral soil layer. The development of new soil layers—and the accompanying flow discontinuities—may have further regulated the infiltration process in the lower rain garden, which may explain observed ponding nearest the inlets, where soil formation is most active and layering is most pronounced. The coarser sands (larger particle size) in the upper rain garden had high infiltration and throughflow rates due to its overall weak structure, and a high degree of structural macroporosity. Although limited to observations on bulk density soil cores, soil macroinvertebrate presence was observed in the upper, but not the lower rain garden. Macroinvertebrate activity (e.g., ant and earthworm burrowing) creates macropores that—when soils are saturated—can rapidly channel flow to the gravel storage sub-layer. Therefore, a combination of both soil development and biogenic processes may have furthermore maintained high infiltration rates in the upper garden, and regulated the same in the lower rain garden [22
Coincident sediment deposition and post-event ponding was noted in the upper garden starting August 2012. We hypothesized that sediment suspended in inflow would settle, clog surface-connected pores, and thereby degrade infiltration rates [7
]. Although we expected that soils closest to the inlets to have decreased infiltration rate and become denser with time, neither of these relationships were significant for either garden. We were particularly surprised that there was no evidence of degradation in upper garden infiltration rates, where the mass of sediment delivered ranged between 0.1 to 56 kg with a median of 8 kg per event [30
Overall, the upper rain garden acted as a fine sediment filter, protecting the lower garden from sedimentation, such that the study-wide, event-wise maximum suspended sediment load into the lower garden was only 2 kg [30
]. This 75% decrease in fine sediment loading is in agreement with other field studies which reported 68 [29
] to 90% [19
] reductions in suspended sediment loads in networked rain gardens. Jenkins et al. [31
] observed that although the texture of rain garden surface soils was changed by settling of fine sediments over an eight-year study period, infiltration rates did not change. Taken in the context of the present study, the specific composition and thickness of the surface mulch layer may regulate the impact of sediment load on rain garden hydrology. Based on our data, we speculate that sediments were well-dispersed in the vicinity of the inlet, and ultimately incorporated into the thick organic surface soil, where their impact on infiltration rate was minimized.
The absence of clogging effects over a four-year monitoring period in this study may also suggest that the amount of fines in the engineered soil mixes were initially low, and that excess capacity due to oversized garden areas may have provided more surface area over which fine sediment was distributed [8
]. Dumouchelle and Darner [13
] observed nearly vertical infiltration in soils near the lower garden inlet wherein water percolated though this finer soil to the gravel layer, stopped at the underlying restrictive clayey subsoil, at which point flow was lateral, and the gravel layer filled from the bottom upwards. We assume that this same pattern holds for the upper garden where soils are coarser. From a practical standpoint, the event peak depth (via crest stage gages) was always lower than the maximum freeboard depth in either rain garden; total inflow volume for any event was insufficient to fill either rain garden. This suggests that a smaller proportion of each rain garden was active in infiltration and drainage processes. Given the amount of unused surface area (and hence retention capacity) in both rain gardens, future outflow events in this network may be better mitigated by increasing the usable area. Some practical approaches that may be generalizable to other rain gardens include: engaging the unused network storage volume via flow-spreaders; facilitate movement of water to the perimeter by re-grading the gardens to create a slight slope toward the outer perimeter of each garden; and limiting the drainage area of underdrains to a close proximity near the inlet, forcing lateral water flow (fully leveraging subsurface storage) once the maximum vertical flow rate is attained.