Water quality impacts of young green roofs in a tropical city: a case study from Singapore

As cities grow, vegetation is replaced by impervious surfaces. This landcover change creates a host of environmental problems that include increased flood risk, poor water quality, biodiversity loss and urban heating that may be partly addressed by green roofs (Berndtsson 2010). Studies have shown that green roofs can retain approximately 50–89% of incoming rainfall, lowering stormwater runoff volumes and peak flows at both the building and catchment scales (Mentens et al. 2006). The vegetation/soil complex also removes contaminants found in urban rainfall and atmospheric deposition (trace metals and aerosols). As a result, green roofs have been widely implemented in many regions, since the 1960s, for example, in Europe, the United States, China, Hong Kong and Singapore (Nguyen et al. 2019).

A global review of green roof studies shows that they generally neutralise acid rainfall found in urban environments and remove trace metals (Berndtsson et al. 2009; Bliss et al. 2009; Van Seters et al. 2009; Vijayaraghavan et al. 2012; refer to Supplementary Material, Appendix A). However, their presence contributes eutrophication-causing nutrients such as nitrogen (N) and/or phosphorus (P) to urban runoff, especially when fertilisers are used (Supplementary Material, Appendix A). Only a few studies have reported the removal of N and P by green roofs (Kohler et al. 2002; Beck et al. 2011; Speak et al. 2014). These results point to the potential negative impact of green roofs on the urban runoff quality from a nutrient perspective.

Clearly, differences in green roof designs (intensive versus extensive), substrate and vegetation determine the amount of nutrients present in their outflow water. Both controlled laboratory experiments and field trials have identified the important role substrates play in elevating the levels of eutrophication-causing nutrients in green roof runoff (Vijayaraghavan & Joshi 2014; Chen et al. 2018; Qiu et al. 2020; Akther et al. 2021). Phosphorus concentrations tend to be controlled by the substrate, while nitrogen concentrations are controlled by both the substrate and vegetation (Chen et al. 2018).

Furthermore, the temporal variability in green roof runoff quality has important implications for urban water quality. Newly established green roof systems that are less than a year old tend to have the poorest runoff quality (e.g., Vijayaraghavan et al. 2012; Harper et al. 2015), sometimes with nutrient concentrations comparable to the wastewater or agricultural systems (Kuoppamäki & Lehvävirta 2016; Mitchell et al. 2017). The initial flushing (first flush effect) of poor green roof runoff quality, for nutrients and trace metals, was observed for the first rainfall event in a series of events that occurred after a dry period (e.g., Vijayaraghavan & Joshi 2014; Harper et al. 2015; Buffam et al. 2016). Over time, pollutant concentrations may decline after they are flushed out of the substrate and/or removed by vegetation (Berndtsson et al. 2009; Buffam et al. 2016).

In Singapore, green roof implementation has increased since 2017 to combat the urban heat island effect, improve the stormwater quality and increase liveability. Under the Landscaping for Urban Spaces and High-Rises programme (LUSH programme, the high-rise greenery is expected to reach 200 ha by 2030 (The Straits Times, November 10, 2017). Currently, there are already 133 ha of high-rise greenery installed. These applications include green roofs, rooftop farms and vertical greenery. Although the standard guidelines exist for green roof systems (https://www.nparks.gov.sg/skyrisegreenery/news-and-resources/guidelines), only limited research has been conducted to understand their water quality dynamics. Green roofs in Singapore have positive hydrological impacts, reducing runoff volumes (∼45%), dampening peak flows (80–85%) and sustaining downstream flow during dry periods (Vergroesen & Joshi 2010). The water quality impacts are variable. In an earlier study conducted approximately 0.56 km north of this study, Vijayaraghavan et al. (2012) reported that a newly established green roof system (2-month-old) using local garden soil and Sedum mexicanum vegetation removed trace metals, but nitrate (0.34–0.86 mg/L), phosphate (19.8–40 mg/L) and sulphate (76.7–109.2 mg/L) concentrations were higher than the rainfall samples. The phosphate concentrations exceeded the US Environmental Protection Agency (USEPA) recommended concentrations for freshwater (0.05 mg/L) and may be an eutrophication risk for downstream waterways.

Apart from the green roof studies in Singapore and Malaysia (Vijayaraghavan et al. 2012; Kok et al. 2016), there has been limited research on their impacts on urban runoff quality in the Southeast Asian region where green roofs are starting to be implemented widely. It is within this context of effectiveness uncertainty around water quality impacts of green roofs in the tropics that we conduct a study comparing a commercial substrate with a soil substrate recommended for blue-green infrastructure in Singapore and to consider the implications of widespread implementation on urban water quality in the region. The main objective of our study was to investigate the effect of growing medium and vegetation on the green roof outflow quality for a newly established green roof using a commercial green roof setup that is widely used in Singapore. We also examine the changes in the green roof outflow quality over the 9-month monitoring period to characterise the water quality impacts associated with these green infrastructures.

The green roof experimental configuration units were located on the second level of a small roof linking two buildings at the Department of Geography, National University of Singapore (1°17′42.36″, 103°46′15.91″). Singapore experiences uniform temperature (24.9–29.9 °C), high humidity (60–88%) and abundant rainfall (annual rainfall is approximately 2,400 mm) throughout the year due to its location just north of the Equator (Joshi & Balasubramanian 2010). Rainfall variability is associated with the Northeast Monsoon (December–March) and the Southwest Monsoon (June–September). The site is in the western part of Singapore where there are considerable industrial activities, including port facilities, oil-fired power plants and offshore refineries (Jurong Island).

The following two types of substrates were tested in this study: a local garden soil (ASM) and a commercially available green roof mix resembling sawdust that was supplied by a local company. We refer to these media as SOIL and COMMERCIAL, respectively, in this paper. The SOIL medium is commonly used in Singapore for roadside greenery and blue-green infrastructure. It is composed of three parts of local soil, two parts of woodchips and one part of sand, resulting in a texture of clay (5–30%), silt (5–60%) and sand (20–75%) (National Parks Board Singapore, cited in Lim & Lu 2016). The COMMERCIAL medium is a proprietary material. We know it contains coconut husk and pumice. The COMMERCIAL substrate has a higher organic content (loss on ignition (LOI) = 98.9%) and higher specific electrical conductivity (SEC) 247.7 μS/cm than the SOIL substrate (Table 1). The vegetation used for the green roof is couch grass (Cynodon dactylon), commonly used in water-sensitive urban design infrastructure in Singapore. This grass is suitable as green roof vegetation because it is drought-resistant, fast-growing and requires little maintenance.

The configuration units used in this study are commercially designed small green roof units that can be connected to produce an extensive green roof. Each unit (46 × 46 cm plots) contains an upper layer holding the growing medium (75 mm deep), a geotextile material in the middle layer and a 7-L storage reservoir at the bottom. Each configuration unit is connected by a drainage pipe that discharges the green roof runoff through a single outlet into a sampling bottle (Figure 1).

This study had 10 configuration units and considered the effect of two growing media and the presence or absence of vegetation (bare versus grass). This resulted in four configurations that are summarised below:

1. Bare SOIL (SOILbare, n = 2 units)

2. Vegetated SOIL (SOILgrass, n = 2)

3. Bare COMMERCIAL (COMbare, n = 3)

4. Vegetated COMMERCIAL (COMgrass, n = 3)

There were only two units for the SOIL configurations as one unit from each of the bare and grass configurations was used to monitor the green roof runoff depth and hence could not be sampled for water quality analysis. We placed the 10 green roof configuration units randomly across the site on which they were located to minimise any effect of spatial variation in rainfall on the experimental setup (Figure 1).

The green roofs were installed at the end of February 2014. The grass mats were grown in a local nursery and placed on the substrate during installation. The grass mats were watered with tap water for a month to ensure that it was established on the green roof substrate before the monitoring started in April 2014. None of the 10 configuration units were fertilised or irrigated during the monitoring period. Bulk deposition (including both dry deposition and wet deposition when rainfall occurs) was collected using a plastic container (0.33 × 0.46 m) placed next to the configuration units as a comparison to the runoff collected from the green roof configuration units (Figure 1). Bulk deposition is termed as the RAINFALL in this paper.

A controlled batch leaching study in the laboratory was conducted to examine the leachate quality of both substrates and to compare this to the green roof runoff quality. A 20:1 liquid:solid ratio was used for the batch study. Samples were placed in Erlenmeyer flasks filled with deionised water acidified to a pH value of 4.2 to simulate the mean acidity of Singapore's rainfall (Balasubramanian et al. 2001; Hu et al. 2003). Triplicates of each substrate were tested. The flasks were agitated continuously for 24 h before samples were filtered and preserved for analysis. We chose 24 h because it was the hypothesised mean residence time of water in the green roof system.

Ten rainfall events were sampled between April and December 2014 (Table 2, Figure 2). Field monitoring started in April which corresponds to the inter-monsoon season that is generally the drier period in the year. Coincidentally, drought conditions prevailed during the first 3 months of 2014, exacerbating the already dry conditions experienced during the inter-monsoon season.

We divided the experimental period into three periods (Figure 2). Period 1 corresponds to the first half of the year, which was relatively dry. Only 31% of the total annual rainfall had fallen when the third event was sampled in end-June 2014. Period 2 corresponds to the period just as the wet season was starting, representing wetting up conditions for the green roofs. The events sampled in Period 3 represented wet season conditions during the Northeast Monsoon season (Figure 2, Table 2). The rainfall properties of Periods 2 and 3 are quite similar. The greatest distinction in rainfall properties between the three periods is seen in the event rainfall amount, which is highest for Period 3 (Table 2).

The total volume of green roof runoff and RAINFALL samples were collected immediately after each rainfall event. An aliquot was taken from the green roof runoff sample for water quality analysis and represents the event mean concentration for the storm event. Both the green roof runoff and RAINFALL samples were stored in acid-washed containers. Samples were tested for pH, SEC, total organic carbon (TOC), anions (Cl⁻, SO₄²⁻, NO₃⁻-N and PO₄³⁻-P) and cations (Ca²⁺, Mg²⁺, K⁺, Na⁺ and NH₄⁺-N). Metals (Fe, Cu, Zn, Cd and Pb) were determined for the final seven events when the scope of the initial study expanded to include more variables. Water quality parameters such as pH, SEC and TOC were tested on unfiltered samples. A 0.45-μm Millipore PTFE syringe filter was used to filter the samples. TOC was determined using an Elementar Vario TOC cube using catalytic high temperature combustion. Cations and anions were tested using the Dionex Ion Chromatography ICS 5000 System with a capillary column each for cations and anions, respectively. Trace metal extraction was conducted using the USEPA3051A method, and concentrations were determined using an ICP-MS (Agilent Technologies, 7700 Series). Five calibration standards were used for the analysis together with two certified reference materials (Environment Canada Certified Reference Material TMDA 64.2 and TM23.4).

Statistical analyses were conducted using the SPSS Statistics 22 software. The data were screened for normality using the Shapiro–Wilk test. Levene's test of homogeneous variances was performed on the data. The paired sample t-test was conducted to test for statistically significant differences between the outflow water quality from the four green roof configurations. The Spearman correlation coefficient was used to test for correlation between the RAINFALL and green roof outflow water quality and environmental variables such as event rainfall amount and antecedent rainfall amount. The 95% significance level (α = 0.05) was used to test for the significance.

The leaching experiments provide insights into the substrate characteristics and leachability of the various water quality parameters examined in this study. Leachate concentrations from the COMMERCIAL substrate had higher concentrations of SEC, TOC (239.7±7.3 ppm), NH₄⁺-N (42.6±0.41 ppm), Cl⁻ (113.8±1.21 ppm) and Zn (39.1±5.3 ppm) (Table 3). This may be related to its higher organic content (LOI = 98.9%) and SEC (247.7 μS/cm) than the SOIL substrate, highlighting the greater risk of poor water quality when this substrate is used for green roof installations (Table 1).

All the sampled rain events were acidic (4.1±0.18) and were comparable to the previous studies in Singapore (rainfall pH 3.8, Vijayaraghavan et al. 2012; Table 4). The configurations neutralised the acidic rainfall, such that the runoff pH was close to neutral (6.7–6.9), very similar to an earlier study conducted nearby (pH 7–7.5, Vijayaraghavan et al. 2012) and with the other studies elsewhere (e.g., Teemusk & Mander 2011; Whittinghill et al. 2016; Wang et al. 2017a). The least neutralising effect was observed for the COMgrass configuration (pH 5.8±1.0).

The RAINFALL samples contained NH₄⁺-N (0.30±0.12 ppm) and NO₃⁻-N (1.0±0.52 ppm), but PO₄³⁻-P was not detectable (Table 4). Concentrations were similar to an earlier study conducted in Singapore (He et al. 2011). The green roof configurations added PO₄³⁻-P with runoff concentrations higher than the RAINFALL samples with enrichment ratios quite similar for both the substrates (SOIL: 1.3–4.4, COM: 2.2–4.2, Table 5). However, phosphorus pollution is not an issue with the green roof configurations in this study. The PO₄³⁻-P concentrations were lower than the Singapore's guidelines for the trade effluent into a controlled waterway and a study conducted at a nearby location (6.46–13.0 ppm, Vijayaraghavan et al. 2012), other green roofs (0.05–0.72 ppm, Razzaghmanesh et al. 2014) and a rooftop farm (1.38 mg/L, Whittinghill et al. 2016) (Table 4).

For nitrogen, the SOIL configuration units were a sink for NO₃⁻-N. This positive nitrate performance is reflected in the high RE for NO₃⁻-N (RE = 70.7%, SOILbare) and NH₄⁺-N (approximately 97% for both SOIL configurations). The COMMERCIAL configurations were a source of N-nutrients, with NO₃⁻-N and NH₄⁺-N concentrations that were 76 and 177 times higher than the RAINFALL (Table 4), close to or above the USEPA (2009) standards for drinking water (10 ppm NO₃⁻-N) and Singapore's guidelines for the trade effluent into a controlled waterway (Table 4). In particular, the NO₃⁻-N concentrations (75.6±53.9 ppm) from the COMgrass configuration were above the limit for surface freshwater eutrophication (11.3 mg/L NO₃⁻-N, Speak et al. 2014), at least an order of magnitude higher than the other published studies (0.29–8.9 ppm, Aitkenhead-Peterson et al. 2011; Teemusk & Mander 2011; Razzaghmanesh et al. 2014), higher than the runoff from a rooftop farm (NO₃⁻ = 4.95 mg/L, Whittinghill et al. 2016) and comparable to agricultural runoff (0–75 mg/L, up to 376.4 mg/L in nursery farms, cited in Whittinghill et al. 2016), highlighting a significant eutrophication risk for downstream waterways/waterbodies from nitrogen pollution.

The previous studies have shown that the green roof vegetation aided in nutrient removal (e.g., Kohler et al. 2002; Berndtsson et al. 2006, 2009; Aitkenhead-Peterson et al. 2011; Vijayaraghavan & Joshi 2014). The results from this study were mixed for N and P nutrients. For phosphate (PO₄³⁻-P), grass configurations removed phosphate for both the SOIL and COMMERCIAL substrates, with lower runoff concentrations, compared to the bare configurations. For N-nutrients, the SOILgrass removed both N-nutrients from the RAINFALL with removal efficiencies of 41.7 and 97.9% for NO₃⁻-N and NH₄⁺-N, respectively. However, the COMgrass configuration units resulted in NO₃⁻-N and NH₄⁺-N concentrations that were 10 times and 2.4 times higher than the bare configuration (Table 5). The reason for the poor runoff quality may be due to the high organic content in the substrate which decomposed, converting organic N into more labile forms, NH₄⁺-N and NO₃⁻-N as well as the nitrification of existing NH₄⁺-N in the substrate into NO₃⁻-N. Kuoppamäki & Lehvävirta (2016) also found that their vegetated configuration units had higher concentrations of total nitrogen (approximately 2–7 times higher) and total phosphorus (3–5 times higher) compared to the bare configurations due to the decomposition of dead vegetation and nitrogen-fixing leguminous plants (Trifolium repens and Lotus corniculatus). Plant growth was particularly luxurious for the COMgrass units, possibly attracting birds, whose faeces contribute NH₄⁺-N which quickly nitrifies to NO₃⁻-N (Berndtsson 2010; Buffam & Mitchell 2015). Others have also acknowledged the role of birds on green roof nutrient runoff quality (e.g., Teemusk & Mander 2011).

The SOIL substrate removed the potentially toxic metals such as Pb and Cd, with high RE (>80%) when compared against the RAINFALL samples. The COMMERCIAL configurations, by comparison, released these metals; Pb concentrations were 13.9 (COMbare) and 9.7 (COMgrass) times higher than they were in the RAINFALL samples. The Cd concentrations in COMgrass runoff were 7.2 times higher than they were in the RAINFALL samples (Table 5). Despite their proximity to an industrial area, the green roof runoff concentrations were below the USEPA (2009) guideline values for drinking water and freshwater and Singapore's guidelines for the trade effluent into a controlled waterway (Table 4). The higher trace metal concentrations from the green roof configuration units, in comparison to the leaching samples, highlighted the effect of field exposure which subjects the green roofs to additional inputs of trace metals from metal-enriched roadside sediments via wind resuspension (Yuen et al. 2012). The Fe (1.7–13.7 ppb) and Cu (1.3–2.3 ppb) concentrations in this study were also approximately an order of magnitude lower than the concentrations obtained from a nearby study (Fe: 43–113 ppb, Cu: 37–56 ppb, Vijayaraghavan et al. 2012). These differences in runoff concentrations between the two studies can be attributed to the roof exposure, substrate and vegetation.

Grass vegetation generally had a positive effect on trace metal concentrations in green roof runoff. The removal performance was better for the SOIL substrate, especially the SOILgrass configuration which removed toxic metals, such as Zn and Cd, effectively (RE > 80%). The COMgrass configurations also removed Fe, Cu and Zn, with lower removal efficiencies of 21.3, 33.7 and 4.7%, respectively. Cu and Zn retention by grass configurations reflects the vegetation uptake as these trace metals are micronutrients for plant growth (Whittinghill et al. 2015).

Carbon concentrations in green roof runoff were higher than the RAINFALL (Table 4). The SOIL configurations resulted in TOC concentrations 2.1–2.7 times higher than the RAINFALL. This difference increased to 16.8–22.4 times higher for the COMMERCIAL configurations (Table 5) The differences in TOC concentrations in runoff from the two substrates reflect the higher organic content in the COMMERCIAL substrate (Table 1).

Vegetation had a minimal impact on TOC concentrations in green roof runoff for both substrates as did Harper et al. (2015) who suggested that the carbon in green roof runoff originates from the substrates.

The green roof configurations elevated cation and anion concentrations, particularly Ca²⁺ (COMgrass = 17.8±6.0 ppm) and K⁺ (SOILbare = 9.3±8.2 ppm) when compared against the RAINFALL inputs (Table 4). Cation and anion concentrations in this study were lower than the concentrations recorded by Vijayaraghavan et al. (2012) as well as other green roofs elsewhere (Ca: 0.07–32.5 ppm, Mg: 2.8–12.1 ppm; Berndtsson et al. 2009; Berghage et al. 2009; van Seters et al. 2009; Teemusk & Mander 2011).

The grass configuration had minimal impact on runoff quality for the SOIL substrate, given the similar runoff concentration levels as the bare configuration. For the COMMERCIAL substrate, the COMgrass configuration resulted in poorer runoff quality, with higher outflow concentrations of Ca²⁺, Mg²⁺ and K⁺ that were at least four times higher than the bare configuration units (Table 5). Vijayaraghavan et al. (2012) found that S. mexicanum had a positive effect by reducing the concentrations of these ions in the runoff when compared against their bare configurations.

Environment conditions such as rainfall properties and antecedent dry weather affect the runoff quality from green roofs. Our results showed that as rainfall amount increases, the green roof runoff concentrations decrease for almost all water quality parameters, especially for events sampled during Periods 2 and 3 (p < 0.05). This trend was also observed in other studies (e.g., Zhang et al. 2014; Gong et al. 2019). However, other studies have reported higher runoff concentrations for larger rainfall events, especially for soluble nutrients (NO₃⁻-N, NH₄⁺-N and PO₄³⁻-P), which are readily flushed out of green roofs (e.g., Kuoppamäki & Lehvävirta 2016; Wang et al. 2017b; Gong et al. 2019). Bigger and more intense rainfall events cause rapid water flow through green roof systems, reducing the residence time needed for pollutant adsorption and increasing the likelihood of flushing (Wang et al. 2017b). The negative relationship between the rainfall amount and the green roof runoff concentration observed in this study reflects the dominant effect of frequent and heavy wet season rainfall events on green roof runoff quality just before and during the Northeast Monsoon wet season. This seasonal influence was also observed in Chongqing where a negative relationship between rainfall amount and green runoff concentrations was observed for the wet season (Zhang et al. 2015). Buffam et al. (2016) found a weak relationship between the rainfall amount and the runoff concentration due to the over-riding effect of temperature and antecedent dry weather on green roof runoff quality probably through their influence on plant nutrient cycling. Although this study focused only on concentrations, it is important to incorporate load estimates of water quality variables in future work on green roof runoff quality. The greatest water quality impact on downstream systems from a mass load perspective will be from high rainfall events with high concentrations from green roof configurations using the COMMERCIAL substrate, conditions likely to occur at the beginning of the wet season.

Antecedent dry weather is also an important factor affecting the green roof runoff quality in this study. Our results showed that both the RAINFALL (e.g., NH₄⁺, Na⁺ and SO₄³⁻) and the green roof runoff quality were poorer when the 7-day antecedent precipitation rainfall (ADP7) was lower. For green roof configurations, significant positive correlations were observed for almost all configurations, mainly for cation and anion concentrations for events from Periods 2 and 3 (Ca²⁺, Na⁺, K⁺, Cl⁻ and SO₄³⁻, p < 0.05). Other studies also found a positive relationship between antecedent dry weather and runoff concentrations (e.g., Zhang et al. 2014; Buffam et al. 2016). Longer dry periods increase the accumulation of pollutants from sources such as dry deposition and weathering of building material onto green roofs (Berndtsson 2010).

The RAINFALL quality changed over the monitoring period. The highest concentrations for most water quality parameters examined were observed for Period 1 samples (Period 3/Period 1 ratio < 1). Nitrate-N showed the greatest change over the monitoring period (Period 3/Period 1 = 0.13), whereas Mg²⁺ and K⁺ showed minimal changes (Period 3/Period 1: Mg²⁺ = 1.0, K⁺ = 1.07; Table 6, Figures 3 and 4).

The general trend for green roof runoff quality was a reduction in median concentrations over the monitoring period, often >50%, for many water quality variables measured, especially seen for both the SOIL configurations (Table 6). Visible changes in vegetation and the substrate were observed over the 9-month monitoring period. The grass configurations established well without additional watering, fertilisation or maintenance (e.g., plant trimming, Figure 5). For the SOIL configurations, important exceptions to the general trend included trace metals, including large increases in Fe concentrations (both SOIL configurations), while Cu (SOILgrass) and Cd (SOILbare) concentrations were largely unchanged over the wet seasons (Table 6). All other important differences were associated with the COMgrass configuration, including large increases in NO₃⁻-N (Monsoon/Dry = 5.44), NH₄⁺-N (Period 3/Period 1 = 2.46) and SEC (Period 3/Period 1 = 2.1) and small increases in Ca²⁺ (Period 3/Period 1 = 1.16) and PO₄³⁻-P (Period 3/Period 1 = 1.07) over time.

Many water quality variables (e.g., SEC, NO₃⁻-N, NH₄⁺-N, PO₄³⁻-P, Na⁺, K⁺ and Cl⁻) exhibited high concentrations for events sampled in Period 2, just before the start of the Northeast Monsoon wet season, especially for the COMgrass configuration units. Highest concentrations were recorded for most water quality parameters in the second or third event of a series of four events, especially for NO₃⁻-N and NH₄⁺-N (Figure 3). This was most likely due to the accumulation of organic matter and their subsequent decomposition and mineralisation when the substrate wetted up with rainfall events from Period 2 as well as from rainfall inputs that had relatively higher NH₄⁺-N concentrations (RAINFALL: Period 2/Period 1 = 1.31) (Buffam et al. 2016). The first event of Period 2 just wetted up the green roofs given the low ADP7 (2.3 mm), which increased to 33.8 mm after the second event (Table 2). Between the two substrates tested, the COMMERCIAL configurations recorded highest N and P concentrations during this period: NO₃⁻-N (144±61.1 ppm) and NH₄⁺-N (172±58.1 ppm) due to its high organic content.

By the end of the monitoring period, runoff concentrations from the SOIL configurations were lower than the samples collected at the start of the study (Period 3/Period 1 < 0.7) and were also close to or below RAINFALL concentrations. This was not the case for the COMMERCIAL configurations, especially the COMgrass configuration, which recorded high concentrations of NO₃⁻-N (67.7±17.1 ppm) and NH₄⁺-N (53±12.0 ppm) for Period 3 (Supplementary Material, Appendix B). The NO₃⁻-N concentration was above the USEPA guidelines for drinking water, Singapore's trade effluent guideline concentration for discharge into a controlled watercourse (4.4 mg/L NO₃-N) and more than an order of magnitude higher than stormwater runoff concentrations for different urban catchments in Singapore (0.038–2.77 ppm, Lim 2003; Wang et al. 2017a).

The temporal variability in runoff quality observed for the first 9 months of our configuration units agrees with other studies, i.e., green roof runoff quality is the most variable in its first year (e.g., Hathaway et al. 2008; Harper et al. 2015; Kuoppamäki & Lehvävirta 2016) due to rapid changes in the substrate and vegetation (Monterusso et al. 2004; Schrader & Böning 2006). Others have reported delayed increases in N and P nutrient levels due to seasonal effects related to freeze–thaw cycles and additions from the decomposition of organic matter from dead vegetation originating from well-established vegetation or from systems where vegetation is struggling to survive (e.g., Teemusk & Mander 2011; Harper et al. 2015; Kuoppamäki & Lehvävirta 2016). Here, the SOIL configurations stabilised and reached low runoff concentrations that were comparable or lower than the RAINFALL within the first 9 months of deployment. The time required for the COMMERCIAL substrate to achieve the same result is unknown and requires longer term monitoring. Other green roof studies have found that it can take up to 5 years (Berghage et al. 2009) or even 9 years after installation before the green runoff nutrient concentrations are comparable or below rainfall inputs (Berndtsson et al. 2006). For a roof in Germany, phosphorus retention increased from 26% in the first year to 80% in its fourth year of installation due to increased P-uptake and substrate retention after the vegetation established (Köhler & Schmidt 2003). Some factors that affect this stabilisation period include plant growth dynamics and changes to the substrate due to environmental exposure (Alsup et al. 2010; Buffam et al. 2016). Repeated wetting and drying have been found to affect organic matter decomposition and N-mineralisation rates in green roof substrates, affecting their runoff nutrient levels over time (Nagase & Dunnett 2011). The luxuriant grass on the COMgrass configuration suggests that this configuration may be a continual source of nitrogen-N nutrients if such growth is maintained.

The different water quality results reported by Vijayaraghavan et al. (2012) and this study are due to dissimilar substrate and vegetation composition. Comparison of the substrates used in both studies revealed that pH is most similar, while SEC and organic matter content show differences that may account for the variability in green roof runoff quality observed for the two studies (Table 1). Green roof runoff from the study of Vijayaraghavan et al. (2012) reported higher concentrations of almost all water quality parameters tested, especially for phosphorus due to the addition of fertiliser in the substrate, whereas nitrogen (NO₃⁻-N and NH₄⁺-N) was a problem identified in this study.

As the SOIL substrate and grass configurations achieved low concentrations of important water quality parameters (N-nutrients, TOC and trace metals) within a relatively short period of 9 months, we recommend this configuration combination for green roof implementation in Singapore and elsewhere. The low organic matter in this substrate (LOI = 16.9%) and lack of fertiliser additions are probably the reasons for the good water quality outcomes for both SOIL configurations. For comparison, the COMMERCIAL substrate contained excess organic content compared to the recommended values between 4 and 12% (FLL guidelines, Whittinghill et al. 2016). It is likely that commercial substrates are still preferred because they are designed to provide the nutrient needs of green roof vegetation, water retention requirements as well as the lightweight properties required for deployment on roofs. Of the 44 papers on green roof runoff quality compiled in Supplementary Material, Appendix A, about half of them used commercial substrates (45%) and 60% of those that used commercial substrates, including this study, reported poor green roof runoff quality when compared against rainfall and/or bare roof runoff quality. For widespread green roof implementation, we recommend thorough testing of substrate properties before field deployment for properties such as their pH, SEC, carbon, N and P content. This should be done for commercial substrates as well as for locally designed substrates, which may be the case in Southeast Asia, where commercial substrates may be too expensive. Lab results may be different from field results due to additional pollutant contributions from the atmosphere, urban activities and wildlife, especially for Southeast Asian cities with dense traffic and heavy industrial activities. Regular substrate testing should continue even after green roofs are installed to check if they still contain the necessary nutrient stocks for plant growth as this informs fertiliser application strategies (Rowe et al. 2006; Aitkenhead-Peterson et al. 2011; Buffam et al. 2016).

Vegetation had a positive influence on green roof quality, whether it is grass or sedum varieties. Because green roof vegetation also provides habitats for urban wildlife, cooling benefits and improves liveability, maintenance activities to ensure healthy vegetation should be a priority in green roof management especially in the tropics where warm and wet conditions encourage rapid plant growth. The healthy grass growth achieved in the 9 months of this study will need trimming maintenance to avoid problems of increased roof loading and potential poor water quality from overgrown green roof vegetation (Rowe et al. 2006).

By the end of the experiment, the overgrown grass vegetation appeared to be slightly brown which suggested it needed irrigation and fertiliser to maintain healthy growth in the long term. Green roof vegetation tends to experience a boom-and-bust phenomenon due to their exposure to extreme heat and drought (Monterusso et al. 2005; Bates et al. 2015). The sedum plants on some configurations in Vijayaraghavan et al. (2012)’s study also failed to survive. Even though phosphorus was not a problem in this study, P pollution is a common issue due to fertiliser or compost additions to boost plant growth (e.g., Hathaway et al. 2008; Vijayaraghavan et al. 2012). Rowe et al. (2006) found that plants needed a minimum amount of fertilisers (50 g/m²/year) to sustain healthy vegetation growth. The danger of luxuriant growth may be that vegetation is at risk from greater damage during drought periods (Rowe et al. 2006; Nagase & Dunnett 2011). Green roof vegetation in Singapore and other tropical cities will likely require different fertiliser concentrations due to the high plant nutrient demands to sustain their rapid growth in the hot climate. High rainfall in the tropics also encourages frequent flushing of fertilisers out of the root zone before plant uptake occurs (e.g., Wang et al. 2017b). In such situations, the solution may lie in the use of slow-release fertilisers at concentrations sufficient just to maintain healthy plant growth. Because of the lack of information relating to plant nutrient needs for green roofs in the tropics, further research on fertiliser applications in terms of their amount, timing and frequency relative to plant needs are urgently required with the widespread implementation of green roofs in this region (e.g., Emilsson et al. 2007).

The non-point source pollution from green roofs, especially young installations, can be managed, so that the overall benefits, particularly in urban cooling, is realised for tropical cities (Sangkakool et al. 2018). One solution is to connect green roofs to other blue-green infrastructure located downstream for additional water quality configuration or to other urban green spaces (Kuoppamäki & Lehvävirta 2016; Vijayaraghavan 2016). The green roofs in our study could be connected to other green spaces on the university campus to irrigate them, reducing the need for fertiliser applications for these systems. This may be a more practical approach to manage the poor nutrient quality from green roofs for cities in the Southeast Asian region which may not have access to improved green roof substrates that have been modified with materials such as biochar and zeolite (Vijayaraghavan 2016). Furthermore, tropical cities will increasingly face more severe urban water stress due to climate change and population increase, especially for cities located in dry environments as well as megacities (Razzaghmanesh et al. 2014; He et al. 2021). The capture and reuse of green roof runoff as greywater provides a sustainable approach to augmenting the urban water supply. A holistic approach that includes appropriate green roof design coupled with substrate testing, well-devised fertilisation strategy and long-term green roof monitoring, incorporating pollutant mass load estimates, is required with a widespread implementation of these green infrastructure in order to maximise their benefits in the urban environment.

The green roof runoff quality in its first 9 months is highly dynamic and controlled largely by the substrate. The green roof configurations using the SOIL substrate displayed a gradual decline in concentration levels for most of the water quality parameters. The COMMERCIAL substrate displayed more variable patterns, with highest concentrations for the N and P nutrients observed just before the start of the Northeast Monsoon wet season. Frequent rainfall activity resulted in a gradual decline in concentrations over the course of the wet season. All configuration units neutralised acid rainfall and removed trace metals for a site located close to an industrial area. Runoff from the COMgrass configuration resulted in high concentrations of carbon and N-nutrients that make it suitable as a source of irrigation water for other urban green spaces in Singapore. We recommend the SOIL as a suitable green roof substrate for Singapore and other cities in the region. This substrate supported healthy grass growth during the monitoring period. Runoff concentrations from the SOIL-based configuration units were also similar or lower than the rainfall inputs. We also recommend a thorough testing of substrates before they are used for widespread deployment of green roofs in cities to minimise their water quality impacts while enjoying the benefits they provide for a more sustainable urban environment.

We thank Loshini Prabhai and Tian Yuan for their efforts in sample collection and laboratory analysis. We gratefully acknowledge the rainfall data from A/P Winston Chow and comments from Professor Chris Margules and two anonymous reviewers.

All relevant data are included in the paper or its Supplementary Information.