Data has been collected on an extensive GR in northeast Germany located on a multi-storey car park at Berlin-Brandenburg Airport (BER; lat: 52.37°; lon: 13.51°, Fig. A1). The GR is located southeast of the city centre of Berlin on plain level terrain at a height of 61 m a.s.l. The roof has a slope of about 2 %, a size of 8600 m² and was built in 2012. The substrate composition is mainly expanded shale, expanded clay, lava, pumice and compost with a depth of 9 cm (Table 1). Initial substrate organic matter content, which was determined by loss-on-ignition, was 3.1 %, according to the manufacturer's specification. The roof is non-irrigated and has been initially planted mainly with Sedum and Phedimus species. Vegetation height varies between 0.1 and 0.3 m throughout the year. Plant surface coverage, as estimated from photographs taken at regular intervals (usually every 1–2 months) from up to 11 fixed spots distributed evenly throughout the GR (Heusinger and Weber, 2017a), increased from about 41 % in the summer of 2015 to ≈62 % in 2023 (Fig. A2a). Additional information about the site and substrate properties can be found in Table 1.

The study period lasted for 9 years from 1 January 2015 to 31 December 2023. In 2024, photovoltaic panels were installed across the entire BER roof, which prevents further analysis of the green roof–atmosphere exchange under the same conditions after the end of 2023. Carbon, water and sensible heat fluxes were measured using the eddy covariance technique with a temporal resolution of 10 Hz (Table 2). A Campbell Scientific CSAT3A sonic anemometer and an EC150 open-path infrared gas analyser were installed at a height of 1.15 m above roof level. Meteorological variables included air temperature, relative humidity and all four components of the radiation balance. The volumetric water content (VWC) of the substrate was measured at a substrate depth of 5 cm using Campbell Scientific VWC probes, applying the time-domain reflectometry method. The ground heat flux was measured with a Hukseflux heat flux plate located 5 cm deep in the substrate and corrected for heat storage above the heat flux plate using a soil temperature sensor and the calorimetric method (Liebethal et al., 2005; Weber, 2006). All data were stored with a 30 min time resolution.

The EC data were post-processed using the EddyPro software Version 7.0.9 (Fratini and Mauder, 2014; LI-COR Biosciences, 2022) according to established procedures in the EC community (Aubinet et al., 2012). The data processing workflow is based on procedures as documented in Konopka et al. (2021) but will be briefly summarised in the following: Spectral correction was applied according to Moncrieff et al. (1997), and despiking of the data was performed. Quality flags were calculated according to Foken et al. (2005), and flags ≥7 were rejected. Furthermore, data were rejected when CO₂ and H₂O signal strengths were <0.7 and data during and 30 min after precipitation were rejected. Data from the wind sector (36–54°), which is the sector where the measurement tripod affects data, was also excluded. The final data availability was 66 % for CO₂ and latent heat fluxes and 75 % for sensible heat fluxes. Missing data were gap-filled using the Look-Up-Tables (LUT) and Mean Diurnal Variation (MDV) approaches (Falge et al., 2001a, 2001b; Reichstein et al., 2005). The LUT algorithm uses data from similar meteorological conditions to replace missing values. If similar meteorological conditions are not found, the MDV approach is used, which is based on temporal-autocorrelation of the fluxes. We follow the standard micrometeorological sign convention: a negative flux is a flux directed towards the surface, whereas a positive flux is a flux directed away from the surface. Furthermore, daytime is defined for gap-filled data with a global radiation value of >10 W m⁻².

Heusinger and Weber (2017a) documented that the conditions found at BER, e.g. turbulence development, complied with the principles of EC. Measurement setup was the same during the 9-year study period; nonetheless, the data quality and turbulence development were checked for the present data set to investigate potential differences over time. As indicators, turbulence spectra as well as integral turbulence characteristics were calculated.

The integral turbulence characteristic, given by the ratio of the standard deviation of the vertical wind velocity (σW) to friction velocity (u*) under neutral stratification, was calculated for different wind directions to assess turbulence development above the GR. The influence of roughness elements on EC measurements was analysed using the aerodynamic roughness length (z0), calculated from the logarithmic wind profile.

Cospectra were calculated for each flux-averaging period using EddyPro. The program outputs “binned” cospectra by dividing the frequency range into 50 exponentially spaced frequency bins and averaging individual cospectral values that fall within each bin. This reduces noise that typically affects medium and high-frequency ranges. Subsequently, the program calculates ensemble-averaged cospectra for the whole study period that are quality controlled and sorted according to the atmospheric stability regime, defined by the value of the Obukhov length (L; unstable: -650 m <L<0 m; stable: 0 m <L<1000 m).

To check for potential contributions to the measured flux from areas outside the GR, we calculated the aggregated flux source area (FSA) using the FFP model (Kljun et al., 2015). The input parameter of the boundary layer height (h) was assumed as a fixed value of 1500 m. Kljun et al. (2015) reported minor influence on FSA dimensions when changing the value of h by ±20 % (up to ±3.7 % change in FSA peak location under stable conditions), indicating a low model sensitivity for h.

Multiple bulk substrate samples for laboratory carbon content analysis were taken every May or June at randomly selected locations on the GR starting in 2020. In the laboratory, the plant material (roots, leaves, etc.) was first removed from the substrate samples. A mortar and pestle were used for crushing, grinding, and mixing each sample. Subsequently, total carbon (C), nitrogen (N) and sulfur (S) in all solid samples were measured by means of an elemental analyser (EuroEA 3000, HEKAtech GmbH, Germany) that combusted 10–20 mg aliquots of each sample in a tin capsule calibrated with a sulfanilamide standard (CD41:750:17 %; ND16:260:22 %; SD18:640:18 %) and BBOT (2.5-Bis(5-tert-butyl-benzoxazol-2-yl)thiophene; CD72.52 %; ND6.51 %; SD7.44 %). The quality of the measurements was controlled by three certified reference materials – NIST 1515 apple leaves (ND2:250:19 %), NCG DC 73030 Chinese soil (CD0:6170:044 %; SD0:20:03 %), MOC soil standard (CD3:190:07 %; ND0:270:02 %; SD0:0430:005 %) – and sulfanilamide (1 for every 10 analyses).

Samples were not acidified prior to elemental analysis, so the measurements represent total carbon (TC). However, because the substrate consists exclusively of non-carbonate mineral aggregates and contains no carbonate-bearing components, inorganic carbon was assumed to be negligible, and the TC values will therefore be interpreted as total organic carbon (TC ≈ TOC).

In contrast to the laboratory measurements, the manufacturer's value of 3.1 % represents organic matter determined by loss-on-ignition, a method that quantifies total mass loss upon combustion and therefore includes true organic material as well as volatilised structural water and thermally unstable mineral fractions. Hence, the two metrics cannot be directly compared. However, the initial TOC range of 1.5 %–2.5 % was considered plausible based on the substrate composition (Table 1).

A mass-balance comparison between substrate TOC and measured EC carbon fluxes was conducted (Sect. 3.4.3, Fig. 11) with the following procedure: Substrate TOC for both the measured values (2020–2023) and the initial value (2012) were converted to g C m⁻². The values for the other years (2013–2019) have been linearly interpolated. In order to analyse the potential range of substrate TOC, two assumptions were made: (1) the initial TOC was set at 1.5 %, and bulk density was fixed at 0.78 g cm⁻³ (lower bound), and (2) the initial TOC was set at 2.5 %, and bulk density increased to 1.18 g cm⁻³ in 2020 (upper bound). For the mass-balance comparison, we added the annual carbon sequestration (2015–2023) from EC flux observations to the initial TOC value. For the years 2012–2014, where flux measurements were not yet operational, we assumed a value of 121 g C m⁻² a⁻¹, i.e. the average for the period 2015–2021.

The water balance is written as: 1P=Q+ET+ΔS with P = Precipitation, Q = Runoff, ET = Evapotranspi-ration and ΔS = Change in storage.

It is a common practice to neglect changes in the water storage (ΔS) over annual/multi-annual periods in natural ecosystems, because it is minor when compared to the fluxes of the other components (Xue et al., 2013; Greve et al., 2016). Some studies, however, found that ΔS cannot be neglected in the annual water balance (Han et al., 2020; Bruno et al., 2022). The water storage capacity of the BER green roof was found to be 35.8 mm, i.e. maximum possible ΔS (Markolf et al., 2024), indicating its minor contribution to the water balance. The water storage experiences fast fluctuations of storage levels due to the shallow substrate, indicating that water has a short residence time before it leaves the GR via evapotranspiration (ET) or runoff, i.e. extensive GRs are not comparable to natural ecosystems with groundwater access. Hence, we decided to neglect ΔS in the annual water balance. Due to the fact that P and ET are measured variables in this study, we attribute the residual to Q (Q=P-ET). Not measuring Q, but calculating it as a residual of the water balance, raises uncertainties in this term. However, by calculating the ratio of ET to P for annual and multi-annual periods, we are able to retrieve an indicator of the volumetric water retention by the GR.

To evaluate whether there is a statistically significant monotonic trend of a parameter over the study period, two non-parametric rank-based tests were applied. Spearman's rank correlation assesses the strength and direction of a monotonic relationship between a variable and time by converting the data values to ranks and calculating a correlation coefficient (ρ) between the ranked data and the ranked time sequence. In addition, the Mann-Kendall trend test was used, which evaluates all pairwise comparisons among data points and determines whether later values tend to be systematically larger or smaller, expressed by the Kendall's τ statistic. The p-value determines whether the observed trends are unlikely to have occurred by chance, and statistical significance was evaluated at a threshold of p=0.05. Both methods are robust to non-normality and outliers (Croux and Dehon, 2010), making them suitable for time-series data with limited sample sizes.

Dominating wind directions at BER are from the sectors W to SSW and E to ENE (Fig. 1). Wind directions 36–54° were excluded in the quality control process (cf. Sect. 2.2). Wind speed is 2.75±1.5 m s⁻¹ on average, but peaks in wind speed are distinctly higher (maximum of 30.3 m s⁻¹) due to the exposed location in the North German plain.

Mean air temperature was 11.1 °C over the study period, and mean annual precipitation sum was 494 mm (Fig. 2b). Mean air temperature was highest in 2019 (11.6 °C), and lowest in 2021 (10.3 °C), and precipitation was highest in 2023 (736 mm), 2.5 times as much as in the driest year, 2018 (295 mm). Precipitation was evenly distributed over the year, i.e. no rainy seasons (Fig. 2a), but more dry days occurred in summer due to convective rainfall, i.e. less frequent, but heavier events. Air temperature showed clear seasonality: highest temperatures were reached during the summer months (June, July, August), where daily maxima of ≥25 °C occurred (53 d yr⁻¹ on average). The coldest months are December, January and February, where temperatures often fell below the freezing point (56 d yr⁻¹ on average). Temperatures ranged between -13.4 and 37.9 °C during the study period.

Compared to data from the 30 years reference period (1991–2020) at the German Weather Service (DWD) station located near the GR (BER: Berlin Brandenburg, DWD-ID 427, 46 m a.s.l., distance 1.9 km) all the years during the study period were warmer than the average (Fig. 2b), with 4 years (2015–2017, 2021) showing lower deviations from the reference period (≤+0.7 °C) and 5 years showing higher deviations (≥+1.4 °C). Only 2017 and 2023 had higher precipitation sums than the average of the reference period (117 % and 141 %, respectively), but many years were close (86 %–99 %). The year 2022 and especially 2018 experienced precipitation sums well below the average of the reference period (77 % and 57 %, respectively). Yearly ET varies between 194 in 2018 and 315 mm in 2017 (Fig. 2c). The ratio of yearly ET to yearly precipitation, which is indicative of retention (cf. Sect. 2.5), is between 0.49 and 0.58, except for the drought year 2018 (0.66) and the wettest year 2023 (0.38). Considering the whole study period, the ratio was 0.51.

The integral turbulence characteristic σW/u* under neutral stratification is 1.33 on average with low variation across wind directions (Fig. A3a). This is very close to the value of 1.25 as proposed for well-developed homogeneous turbulence under neutral stratification (Foken and Mauder, 2024). Aerodynamic roughness length (z0) is consistent throughout wind directions (Fig. A3b), indicating homogeneous vegetation heights across the GR and minor influence of roughness elements on EC measurements (mean = 0.04 m).

The normalised cospectra for unstable stratification (-650 m <L<0 m; Fig. A4) exhibit the expected turbulent structure with clear peaks at the expected normalised frequencies, consistent scalar behaviour, and an inertial-subrange decay close to the theoretical f-4/3 slope. CO₂ and H₂O cospectra show an enhanced contribution to gas flux from lower frequencies, i.e. larger eddies, and typical high-frequency dampening when compared to the sensible heat cospectrum. The cospectra indicate that the data exhibit good quality, as the expected patterns of turbulence and flux are present with minimal noise or distortion.

The 80 % flux source area extends at a maximum of 39 and 36 m towards the dominating wind direction and over an area of 2550 and 1860 m² for night- and daytime respectively (Fig. 3). The peak contribution is situated in the sector 200 to 300° at a distance of approx. 2.5 m to the EC station. The 70 % FSA is located entirely within the GR, whereas the 80 % FSA extends beyond the GR, but only at the northern edge (277–13°), where fetch is the smallest. The nocturnal 80 % FSA exceeds the GR by a maximum of 4.8 m. The area exceeding the GR accounts for 4.7 % of the total FSA; however, it contributes <1 % to the total flux since wind from North is rare at BER (Fig. 1).

Peak flux density is reached around noon for the shortwave components of the radiation balance and in the early afternoon for the longwave components, corresponding to the time when substrate and atmosphere temperatures reach their daily maxima (Fig. 4a). The ratio of reflected to incoming shortwave radiation (albedo) is 19 % for the study period (Table 3). Available energy for turbulent heat fluxes is obtained as net radiation (Q*) minus ground heat flux (QG). Energy balance closure, derived from the regression between available energy and turbulent heat, is 79 %, indicating a 21 % closure gap (Fig. A5).

In the median diurnal cycle of the energy balance components, sensible heat flux (QH) is negative during night-time (Fig. 4b), indicating that heat is transported from the warmer atmosphere towards the colder GR surface. Also, heat is transported in the substrate towards the cooler surface. During daytime, the flux changes direction, and QH and QG become positive, warming the air above the GR and heat is transported into the substrate. Latent heat flux (QE) is positive throughout the diurnal cycle. However, there are individual events during the study period where QE becomes negative during night and water vapour is transported towards the surface, i.e. dew formation. QH is larger than QG and QE during the day, and the fluxes peak at noon, when the residual (i.e. the difference between Q* and the sum of the three other components) is largest.

The Bowen ratio (β), representing the ratio of sensible to latent heat fluxes, is a diagnostic metric that characterises the distribution of available energy between turbulent heat transfer to the atmosphere and the energy consumed by phase change through evapotranspiration. During winter (DJF) and during the night, β often becomes negative (Table 3) due to a negative QH, when heat is transported from the warmer atmosphere towards the cooler GR. During spring (MAM) and summer (JJA), there is more energy available, leading to larger substrate heat storage and stronger positive fluxes. During the day, β is at least 1 during all seasons except for winter and is highest during the summer months. Considering all data, β is >1 for spring and summer and 0.25 during autumn (SON). For the whole 9-year study period, β was close to 1.

There is a strong dependence of β on VWC (Fig. 5a), i.e. water availability for ET. If water availability is not limited, energy is primarily partitioned into latent heat flux (β<1), resulting in reduced sensible heat exchange and a lower ambient temperature (Fig. 5b). This effect diminishes once moisture levels are below VWC ≈0.05 m³ m⁻³. At that threshold, β increases strongly, and energy partitioning into QH is larger, thus reducing the cooling effect of the GR.

The annual NEE of the Berlin green roof based on direct turbulent flux measurements using eddy-covariance ranged between -154 and 8 with an average of -92±62 g C m⁻² for the period 2015–2023. For the same GR and measurement setup as in the present study, Konopka et al. (2021) documented a consistent annual net uptake of carbon with an average of -141 g C m⁻² over the course of 5 years (September 2014 until August 2019). The lower average annual NEE in the present study stems from the slight net carbon release during the last two years after consistent annual net uptake in the first 7 years of the study period. Kuronuma and Watanabe (2017) measured the plant and substrate carbon concentrations for Sedum mexicanum on a green roof in Chiba (Japan) over a one-year period. They reported annual carbon sequestration values of -276 under non-irrigated conditions and -336 g C m⁻² under regular irrigation. Yang et al. (2023) measured the carbon exchange of an intensive green roof in Nanjing (China) over 1 year using eddy covariance and found the annual NEE to be -61 g C m⁻². Getter et al. (2009) found that an extensive green roof system sequestered -375 g C m⁻² in above- and belowground biomass and substrate organic matter over two years. This results in an average of -187 g C m⁻² yr⁻¹. Hence, our estimates are at the lower end of the range reported for carbon sequestration in other green roof studies that employed different measurement approaches.

The same applies for the comparison to other types of urban green infrastructures (which are discussed as urban NBS) or natural ecosystems. Thölix et al. (2025) used different ecosystem models (JSBACH, LPJ-GUESS, SUEWS) combined with eddy-covariance measurements in Helsinki, Finland, to estimate the annual NEE of different urban vegetation types. The mean NEE for forested sites was approximately -92 and -143 for the period 2006–2021, and -76 and -123 g C m⁻² for an urban park, whereas the mean NEE for lawns was approximately -30 and -7 g C m⁻², for JSBACH and LPJ-GUESS, respectively. The entire vegetation sector, comprising different urban vegetation types, sequestered -40 (±34), -44 (±19), and -52 (±55) g C m⁻² as quantified from the JSBACH, LPJ-GUESS, and SUEWS model output, respectively. Nowak et al. (2013) quantified carbon sequestration by urban trees in the United States using urban tree field data from 28 cities and 6 states and applying biomass and growth equations. The annual net carbon sequestration varied between -81 and -401 g C m⁻² of tree cover and averaged to -205 g C m⁻² of tree cover.

In comparison with natural ecosystems, the GR shows a lower magnitude of annual carbon sequestration, given the structural constraints of extensive systems such as shallow substrates, limited rooting volume, and low water and nutrient availability, restricting vegetation development. Klosterhalfen et al. (2023), for instance, found the annual NEE of an unmanaged mixed beech forest in central Germany to range between -393 and -670 g C m⁻² over the period from 2000 to 2022. They reported a decline of carbon uptake by up to 54 % during the drought years 2018, 2019 and 2022. The annual NEE of 19 grassland sites throughout Europe was found to vary between -653 and +171, with an average of -150 g C m⁻² (Gilmanov et al., 2007). This indicates that natural ecosystems may exhibit large drought-induced variation in carbon uptake, whereas the GR's NEE fluctuates within a relatively narrow range, reflecting the stress tolerance and drought-adapted physiology of its vegetation. Additionally, the GR enables the sequestration of carbon in areas that would otherwise have no capacity for this, effectively converting impervious urban surfaces into carbon sinks.

The carbon flux measurements provide valuable insights into the temporal dynamics of the GR's carbon exchange. The GR shows seasonality in the carbon flux, with the length of the daily net uptake phase associated with sunrise and sunset timings. Peak net uptake is reached, on average, during May, followed by April, July, June and August. This suggests that during late spring, conditions are optimal for high carbon uptake due to high light availability accompanied by less extreme temperatures and higher water availability compared to the summer months. Higher incoming radiation and thus energy availability led to increased net uptake rates, while mean daily temperatures exceeding values of ≈20 °C during summer decreased net carbon uptake, potentially due to stomatal closure to reduce transpiration (Grossiord et al., 2020; McAdam and Brodribb, 2015). There is a clear trend towards lower carbon uptake by the GR when water availability drops below ≈0.05 m³ m⁻³ during spring and summer. At higher VWC values, there was no effect on carbon fluxes observed during summer, but reduced net uptake during spring, which might be due to the spring season incorporating colder phases (e.g. March) where VWC is usually high but light availability and therefore photosynthesis are low.

These results illustrate how carbon flux is affected by various environmental and biotic factors. During the first five summers, higher plant coverage corresponded to increased net carbon uptake. In subsequent years, net uptake was lower at comparable coverage, suggesting that other factors have exerted a stronger influence on carbon fluxes.

The carbon exchange of the GR–atmosphere system was strongly shaped by diurnal patterns, with respiration intensifying in the later years of the study (cf. Figs. 6, 10). Correspondingly, annual NEE became slightly positive in 2022 and 2023, after 7 years of consistent net uptake. Our findings indicate that increased respiration, rather than decreased photosynthetic uptake, caused the observed change in annual NEE. While sink season NEE showed no long-term trend, source season NEE increased distinctly. During the last two source seasons, temperatures did not exceed those of previous years, but they showed the highest mean VWC value. Studies show that microbial activity and soil respiration increase with soil temperature and peak at moderate levels of VWC, where they are not constrained by low water availability or by oxygen limitation under excessive soil moisture (Lloyd and Taylor, 1994; Yan et al., 2016). High soil moisture might have contributed to the observed higher net release rates, but cannot be seen as the sole explanatory factor. With a value of 2.5 % in 2021, substrate TOC remained close to its initial value of 1.5 %–2.5 %, possibly showing some substrate ageing (accumulation of organic matter from decomposing plants and animals), but then increased abruptly to 7.9 % in 2022 and remained elevated thereafter. Higher respiration during 2022–2023 is consistent with the increase in substrate TOC. Other studies indicate that respiration is substantially higher in high-organic GR substrates than low-organic substrates (Halim et al., 2022). Hence, carbon fluxes are sensitive to additional input of carbon into the system, pointing to the potential impact of management practices (e.g. fertiliser application) on carbon sequestration potentials of GR ecosystems.

A mass-balance comparison underpins that atmospheric carbon sequestration, as estimated from NEE observation, which is cumulatively added to the initial substrate TOC, agrees well with measured substrate TOC until 2021 (Fig. 11b). Besides the substrate, carbon from atmospheric sequestration is stored in above- and belowground biomass. In Fig. 11b, we show that even if we allocate the entire annual NEE to the substrate and do not consider export of carbon via runoff, the abrupt rise of TOC after 2022 cannot be explained by atmospheric sequestration alone.

This suggests external carbon inputs – such as unremoved plant litter after maintenance by gardeners (1–2 times per year, cf. Heusinger and Weber, 2017b) or application of carbon-rich material (e.g. organic fertilisers) – as the most plausible cause, although no specific maintenance events were reported by the operator of the GR. Hence, we argue that the observed shift from a robust carbon sink to carbon-neutral conditions was primarily driven by the sudden increase in substrate TOC, which originated from sources other than atmospheric CO₂ sequestration. Assuming that the threshold at which the GR is not a sink for carbon anymore (i.e. NEE ≈ 0 g C m⁻²) is defined by the substrate TOC of 7.9 % in 2022, it would take at least 32 years to reach that value when extrapolating the time series from 2022 onwards at an average rate of sequestration of -121 g C m⁻² a⁻¹ (the average rate of sequestration for the years 2015–2021). Further assessment is not possible after 2023 due to the installation of photovoltaics and the resulting alteration of the GR system.

During the 9-year study period, we found that 49 % of the precipitation resulted in GR runoff, meaning that 51 % was retained. Except for the years 2018 and 2023, annual retention ranged between 49 %–58 %. The retention capacity of the BER GR is within the range of values found in other studies for extensive GRs: Schultz et al. (2018) studied runoff patterns from two extensive GRs with substrate depths of 7.5 and 12.5 cm in Portland, Oregon, USA, for a full annual cycle. From a total of 808 mm precipitation, 23.2 (7.5 cm GR) and 32.9 % (12.5 cm GR) was retained. In 2023, precipitation sum for BER (9 cm substrate depth) was similar (736 mm), and retention was 38 % of the annual precipitation sum. Stovin et al. (2012) reported a value of 50 % retention from a UK extensive GR test bed and VanWoert et al. (2005) 61 % on a roof platform with full vegetation coverage. Our data suggests that high annual precipitation sums lead to lower annual retention. Studies report that meteorological conditions, e.g. precipitation sum, and other factors such as GR design may have a considerable influence on the retention capacity of GRs (Liu et al., 2019; Stovin, 2010; Schultz et al., 2018). Our results suggest that the water storage of the GR provides sufficient capacity to enhance ET and reduce runoff, which in turn has the potential to reduce the burden on wastewater structures as well as the risk of flooding in urban areas.

Since substrate water availability not only determines the hydrological performance but also directly affects the energy and carbon fluxes of the system, the interaction between water content, Bowen ratio, and carbon flux was further examined. When β reaches values ≥2 during summer and ≥3 during spring, net carbon uptake by the GR is reduced. This is the case when VWC is below a critical threshold of ≈0.05 m³ m⁻³. The three variables are coupled: Low VWC reduce ET, resulting in increased partitioning of available energy into sensible heat flux, i.e. higher β, as well as lower net carbon uptake. With regard to the ecosystem services of extensive GR, it would therefore be advisable to keep VWC values above that threshold to preserve the cooling effect and to increase carbon uptake by the GR. This could be achieved through a sustainable automated irrigation, e.g. using on-site harvested rainwater, which activates during dry conditions (e.g. Heusinger et al., 2018).