The study site is located in Beijing Olympic Forest Park (40.02°N, 116.38°E), Megalopolis of Beijing, China. It is the largest urban forest park in Asia, with an area of 680 ha and vegetation cover of ~96%. The forest reserve is located in the middle of the northeastern section of the park, an area committed to ecological conservation and reclamation, with no open water or roads in the area. Tourists are restricted from entering our research plot in order to minimize human disturbance. The terrain is flat and uniform. The soil is mainly of the fluvo-aquic type, with soil porosity of 40.3%, pH of 7.8, and an estimated field capacity and permanent wilting point of 26 and 10%, respectively. The site is characterized by a temperate continental monsoon climate. The 50-year annual mean temperature and annual precipitation are 12.5°C and 592 mm, respectively. The seasonal distribution of precipitation is highly uneven, with 80% of annual rainfall falling from June to August. The plant species in the forest reserve are endemic to the North China Plain. Tree species present include Sophora japonica L., Fraxinus chinensis, Pinus tabuliformis, and Platycladus orientalis. The understory consists mainly of Iris tectorum and Dianthus chinensis. The tree stand density in 2015 is recorded to be ~210 stems ha⁻¹, with a mean tree height of 7.7 m and a mean diameter at breast height of 20 cm. The cover ratio of trees-to-shrubs is ~7:3. Shrubs include Prunus davidiana, Amygdalus triloba, Swida alba, and Syzygium aromaticum, with a mean height of 2.8 m. All trees in the reserve have shallow roots with a depth of ~0.08–0.4 m.

Water vapor fluxes were measured at 12 m above the ground surface with an eddy-covariance (EC) system, consisting of a three-dimensional sonic anemometer [CSAT3A; Campbell Scientific, Inc., Logan, UT, USA (CSI)] and a closed-path infrared gas analyzer (model EC155; CSI), which was calibrated once a month. The EC flux measurements began in April 2011. Continuous high-frequency data (at 10 Hz) were collected and processed in the calculation of 30 min fluxes. Fluxes were subsequently recorded on a datalogger (CR3000; CSI).

Hydrometeorological variables, including air and soil temperatures (T_a and T_s, °C), relative humidity (RH, %), net radiation (R_n, W m⁻²), precipitation (P, mm), and soil volumetric water content (VWC, %) were measured simultaneously. Air temperatures, relative humidity, and net radiation were measured at 12 m above the ground surface. Precipitation was measured within an opening (unvegetated) area near the EC tower. Soil volumetric water content and T_s were measured at a 10 cm depth at five locations around the tower. Data for all hydrometeorological variables were stored on a datalogger (CR3000; CSI). Further details on instrumentation and sampling procedures can be found in Xie et al. (2016) study.

To characterize land cover and temporal changes in the canopy structure in response to climate variations, we used time series processed images (i.e., image-pixel containing the EC tower) of normalized difference vegetation index (NDVI; MOD13Q1 products) at 250 m resolution downloaded from the Earth Observing System Data Gateway. The images were downloaded as 16-day composites covering the entire flux-measurement period from 2012 to 2017.

Post-processing was performed using the EddyPro 4.0.0 software (LI-COR Inc., USA), including spike removal, double coordinate rotation, time delay corrections, frequency response corrections, detrending (block averaging), and flux computation (Burba, 2013). Half-hourly turbulent fluxes and meteorological data were carefully controlled for quality and spurious data points. Outliers were removed following the approach described by Papale et al. (2006). Data gaps were filled using the standard method used by the Fluxnet Canada Research Network (Amiro et al., 2006). Three diagnostic variables, including the surface conductance (g_s), the Priestley–Taylor coefficient (α), and the decoupling coefficient (Ω), were used to interpret the relative importance of biophysical control on ET.

Variable g_s of a dry surface mostly reflect the stomatal control of plants on ET. The α-coefficient in the Priestley–Taylor model can be viewed as a normalization of ET (Blanken et al., 1997), which provides a comparison of measured ET to climatological expectations. Usually, α ≥ 1 occurs in mesic ecosystems, indicating sufficient water supply so that latent heat fluxes are limited by available energy. In contrast, α < 1 occurs in semiarid and arid ecosystems, where latent heat fluxes are limited by the availability of soil water (Jia et al., 2016). Ω indicates the degree of canopy decoupling from the bulk air, with Ω varying from 0 (completely coupling) to 1 (decoupling), as the biological control on ET becomes weaker as Ω approaches 1.

We estimated g_s (mm s⁻¹) by inverting the Penman–Monteith equation (Monteith and Unsworth, 2007), such that

where γ is the psychrometric constant (kPa K⁻¹), Δ is the rate of change of saturation water vapor pressure as a function of T_a (kPa K⁻¹), R_n is the net radiation (W m⁻²), λE is the latent heat flux (W m⁻²), G is the soil heat flux (W m⁻²), ρ is the air density (kg m⁻³), C_p is the specific heat of air (J kg⁻¹ K⁻¹), and VPD is the water vapor pressure deficit (kPa).

The aerodynamic conductance (g_a, in mm s⁻¹; Monteith and Unsworth, 2007) was calculated as

where u_* and u are the friction velocity and wind speed, respectively, both given in m s⁻¹. The Priestley–Taylor coefficient (α; Priestley and Taylor, 1972) was calculated from

Ω is calculated as (Jarvis and McNaughton, 1986)

α, Ω, and g_s were calculated based on non-gap-filled 30 min data from daytime hours (when PAR > 5 μmol m⁻² s⁻¹), during periods when the soils were unfrozen, i.e., when T_s > 0.5°C and for days with no rain. The α, Ω, and g_s values were aggregated to daily means and longer periods.

Growing season length (GSL) was determined from the daily mean gross ecosystem production time series, following procedures reported by Zha et al. (2009). Growing seasons spanned DOY (day of the year) 96–303, 104–298, 98–302, 108–302, 97–302, and 94–297 for 2012–2017, respectively. Normally, tree leaf expansion and coloration stages commence in April and October, respectively. According to the seasonal variation in NDVI, we defined March to June as the leaf expansion period, July to September as the leaf-expanded period, and October to November as the leaf coloration period. We defined the period from the previous November to March of the following year as the non-growing season. Following Xie et al. (2016), we defined drought as the occurrence of days when VWC was < 12.5% from March to November, March through May as spring, June through August as summer, and September through November as autumn.

Linear regression was used to examine the relationships between ET and the target biophysical variables (i.e., R_n, T_s, VWC, P, NDVI, and g_s) at various timescales. To evaluate the contribution of different phenological phases to the variability in annual ET, the difference between the standard deviation of cumulative ET from a given month or season and its preceding one (DstDev) was calculated to quantify the contribution of each month or season to the observed interannual variability of the ET. All statistical analyses were performed in Matlab (The MathWorks, Natick, MA, USA).

Variation in major environmental factors during the 6 years is shown in Figures 1–3. Compared to the annual analysis, some factors (i.e., monthly R_n, monthly VPD, monthly T_a, monthly T_s, and monthly NDVI) were much more variable during the years 2012–2017 (Figure 2; Tables 1, 2). Interannual variation of monthly mean R_nand monthly mean VPD were highest in June; monthly mean T_s and NDVI were non-negligible in April (Figure 2). By contrast, interannual variation of annual mean T_a, T_s, R_n, VPD, and NDVI were imperceptible (Table 1). Annual mean T_a above the canopy increased from 2012 to 2015 and then decreased slightly over the remaining 3 years (2015–2017). The years after 2014 were all warmer than the 50-year mean of 12.5°C (1961–2010; Table 1), showing a warming climate.

Precipitation regime and soil volumetric water content (VWC) varied significantly both seasonally and annually at the site (Figures 1C, D, 2B, F; Table 1). The VWC at 10 cm depth varied with soil thaw in early spring and P amount (Figures 1C, D), increasing abruptly in March or June with the start of the rainy season and peaking in the middle of the growing season. Interannual variation in monthly P and monthly mean VWC were highest in July (Figures 2B, F). The mean annual total P was 473.52 ± 168.54 mm, with a CV of 35.6% (Table 1), and the lowest and highest annual mean were 270.9 and 738.3 mm in 2014 and 2012, respectively. The wet year 2012 had many irregular and intensive precipitation events, with 24% of the annual total P falling during a single summer rainfall event (i.e., 176 mm on DOY 203, dashed vertical line in Figure 1C). The year 2016 had one intensive rainfall event with 172.3 mm on DOY 202 (dashed vertical line in Figure 1C), which accounted for approximately 31% of the annual total P. In contrast, annual P in other years was all below the 50-year mean, especially the dry years 2014 and 2015 with 311 and 268 mm (Table 1). The dry year 2014 had the lowest mean annual VWC of 10.4% among the 6 years (Table 1). High spring VWC in 2013 was mainly caused by high non-growing season P (~99 mm) during the preceding year (i.e., 2012; Figures 1C, D).

Soils in late spring to early summer were generally dry over the 6 years (Figure 1D). According to the number of consecutive low VWC days during the growing season, 2012 had one extremely dry soil period (DOY 125–176) in spring, with a mean VWC of 10.0%. The year 2013 had two extremely dry soil periods, one in spring and one during the leaf coloration phase (i.e., DOY 134–157 and 286–334), with a mean VWC of 10.9 and 10.8%, respectively. The dry year 2014 had three extremely dry soil periods, one during each of the leaf expansion, leaf-expanded, and leaf coloration phases (DOY 60–131, 190–210, and 270–330), with a mean VWC of 9.9, 10.5, and 10.1%, respectively. The dry year 2015 had three extremely dry soil periods, two during the leaf-expansion phase and one during the leaf-expanded phase (DOY 60–101, 137–167, and 187–199), with a mean VWC of 8.8, 10.2, and 11.5, respectively. The year 2017 had one dry soil period during the leaf expansion phase (DOY 105–172), with a mean VWC of 11.1%. There was no dry soil period in the year 2016.

Both monthly and annual ET varied significantly among years at the site (Figures 2A, 3A). Interannual variation of monthly ET was highest in June (Figure 2A). Over the 6 years, annual ET was highest in 2016 and lowest in 2015. The mean annual ET was 462.17 ± 83.58 mm, with a CV of 18% (Table 1).

Over the 6 years, monthly ET was positively correlated to P in July (Figure 4A), to T_s in August (Figure 4F), to R_n in September (Figure 4C), and to g_s in all months during the leaf-expanded period (with p's < 0.05; Figure 4H). Total ET over the spring and autumn were both positively correlated to mean VWC over the corresponding period (p's < 0.05; Figure 5), with a higher sensitivity in spring than autumn. Over the 6 years, annual ET values were positively correlated to annual total P, especially P in July (Figure 6) and the non-growing season (November through March; n = 5, p < 0.05, R² = 0.98; Figure 7A).

Cumulative ET showed a seasonal trend similar to cumulative P (Figure 8). Cumulative P exceeded cumulative ET on DOY 203 in 2012 and DOY 182 in 2013, remaining greater than the cumulative ET until the end of the year (Figures 8A, B), yielding annual cumulative difference between P and ET (i.e., P and ET) of 176 and 85 mm, respectively. In contrast, 2014 and 2015 provided cumulative ET greater than cumulative P after DOY 146, yielding annual cumulative P and ET of −135 and −72 mm, respectively (Figures 8C, D). During 2016 and 2017, annual cumulative ET approximately equaled that of P, with most rainfall refilling the soils in mid-summer (Figures 8E, F).

Daily ET decreased during dry periods (Figure 3). Daily α, Ω, and g_s displayed a trend similar to that demonstrated by ET. The extended dry spell in 2014–2015 led to an analogous reduction in annual ET, α, Ω, and g_s (Table 1), causing a substantial decline in the cumulative difference between P and ET (Figure 8). Mean annual ET for the dry year 2014 and 2015 (~292 mm yr⁻¹) was 30% lower than the non-drought year in 2016 (~565 mm yr⁻¹). Accordingly, the derived parameters α and g_s also showed large wet-dry contrasts, with dry vs. wet year reductions of 29% (α), 35% (Ω), and 35% (g_s). The shorter duration of spring drought led to annual ET being higher (Figure 9).

Our study demonstrated that the annual P, especially P in the non-growing season and July controlled the IAV of ET (Figures 6, 7A). This result supports the hypothesis that long-term, local-scale ET is generally sensitive to the IAV of P in water-limited environments (Fatichi and Ivanov, 2014). It was also consistent with previous findings that soil water replenishment through P during the non-growing period of the previous year and mid-growing season significantly affected current-year ET and was largely responsible for the interannual variation observed in ET in temperate forest ecosystems (Ma et al., 2018). Precipitation in different periods may affect annual ET through different mechanisms. Since the water availability in pre- and early-growing seasons was confirmed to be essential for the growth of forests in water-limited regions of temperate China (Wu et al., 2009, 2016), P in the non-growing season and leaf expansion period could primarily affect annual ET by its regulation on canopy development. At our site, we found that P in the non-growing season could provide higher spring soil moisture content that can benefit spring canopy formation (Supplementary Figure S1) and make a higher spring ET (Figure 5A). The result is comparable to the results obtained by previous research, which showed a lag effect of P on ET in temperate China (Xu et al., 2014). Notably, too much P during the non-growing season might lead to a decrease in spring ET. According to some reports, a wet winter can delay the growing season start and affect vegetation growth due to its effect on absorbed radiation energy and growing degree days of boreal forests at high latitudes (>40°N) (Piao et al., 2011; Fu et al., 2014; Yun et al., 2018), it is reasonable to assume that P (99 mm) during the non-growing season of 2012 caused the highest spring VWC but lower T_s, R_n (Table 2), leading to a delayed growing season start and restricted canopy development in early spring (Figures 1E, 5A), reducing days with abundant light and active vegetation, thus a lower spring ET (Figure 5A) of 2013.

After the canopy completed construction (July to September), our result showed that ET was positively related to P in July (ET_JulytoSep = 0.27P_Jul + 184.5, R² = 0.69, p < 0.05) but not to P during July to September (ET_JulytoSep = 0.21P_JultoSep+172.1, R² = 0.50, p = 0.12 > 0.05), indicating that the effect of P in July played a more important role than P in later months in regular ET during the leaf expanded period. This may have occurred for the following reasons. First, as the cumulative P–ET was always negative before July (Figure 8), P in July contains 45% of annual P in mid-summer non-drought years (Figure 5B; Table 2), and it is very important to supply soil water during the leaf expanded period. Our result also showed that P in July could explain 50% of the variation in VWC during the leaf expanded period (n = 6, R² = 0.5). Second, in instances of a more favorable water supply, ET becomes more constrained by atmospheric demand (Liu et al., 2018), and rainfall may offset the positive effects of available energy on increasing ET (Ryu et al., 2008; Kume et al., 2011). August and September had the highest mean monthly α, and Ω and relatively higher VWC during the 6 years (Figures 1D, 3, 4E; Table 2); more mild, longer duration P might lead to lower R_n and soil temperature, resulting in an uncertain effect of P amount on ET. That is the reason why the forest showed a strong interannual response of monthly ET to P for July but not for other summer months (Figures 4A, 6). Thus, the large variability of P at the onset of the monsoon may lead to IAV in the ET during the whole leaf expanded period in forests of Northern China. The results suggest that ET in our forest may be affected by possible changes in the rainfall regime. These observations together suggested the importance of the amount, intensity, and timing of precipitation control on IAV of ET.

The 6-year study (2012–2017) provided an opportunity to study the interannual responses of ET to drought. Drought events at the site occurred randomly at any phenophases and decreased the ET (Figure 1) whenever they happened. However, the mechanism and extent to which drought-affected annual ET differed with the timing and duration of drought occurrence.

In previous temperate forest studies, soil drought due to seasonal and interannual variations in rainfall could be a determining factor for interannual variations in ET due to stomatal closure or leaf fall (Wilson and Baldocchi, 2000; Thomas et al., 2009). Spring drought has been reported to suppress canopy development in a deciduous temperate forest (Noormets et al., 2008). At our study site, we found that the values of spring NDVI in years with spring drought were lower than the year 2016 (Figures 1E, 5A), while ET during spring had a positive relationship with spring NDVI (p = 0.06, R² = 0.65) and IAV in annual ET (p < 0.05, R² = 0.96; Figure 7B). This might indicate that the drought in spring can hinder canopy development and might result in an overall reduction in annual ET. As the photosynthetic structure in the leaf is determined by conditions at the time of leaf development (Kull and Tulva, 2002), the spring drought might have a detrimental effect on the canopy development of the growing season. That appears to be why the mean value of NDVI during the leaf expanding period was positively related to spring NDVI (y = 2.28 x – 0.97, n = 6, R² = 0.71, p < 0.05). As g_s depends on leaf area index (LAI) and leaf-level stomatal conductance, reduced NDVI can lead to suppressed g_s, and thus ET (Figure 7D). The drought in spring may also affect annual ET by regulating plant phenology. Some studies found that drought happening in spring or autumn can influence the lifespan of annual deciduous and grassland in temperate regions, thus affecting growing season length and regulating days with active vegetation and IAV in annual ET (Wilson and Baldocchi, 2000; Ryu et al., 2008; Thomas et al., 2009; Zha et al., 2010). Our study also showed that the GSL had a positive influence on annual ET (Figure 7C). The variability of the onset date exceeded that of the end (Figure 8; Table 1) and the variability in the NDVI during the onset of the growing season exceeded that during the end of the growing season (Figure 2D) due to soil water replenishment through P during the summer month and less soil moisture deficit during the autumn (Figures 1C, D). As a result, ET was shown to be more sensitive to the availability of soil moisture during the onset of the greenness period than the transition to the senescence period (Figures 2, 5). Consequently, due to the monsoon circulation features (soil water replenishment through P in summer), the variability in the intensity of drought in spring may play a more important role in regular GSL and annual ET than that in autumn in temperate forest ecosystems. The more intense spring drought may lead to lower annual ET. Our result is that the most severe early spring drought in 2015 came with the lowest VWC in March (Table 2), leading to the latest date of onset, the shortest GSL; adding days with inactive vegetation and the lowest annual ET over 6 years (Figures 1D, E, 7C) can further support this.

We found that low P in mid-summer coincided with extreme drought and low ET during July to September, which occupied 52% of annual ET on average, particularly in 2014 (Figure 1D; Table 2). During the mid-summer, T_a and R_n reached their maxima, and water supply was substantially limited (Figures 1D, 8). Drought during the mid-growing season has been reported to reduce both photosynthetic capacity and transpiration, resulting in inhibited plant growth, earlier canopy senescence, self-protection by stomatal regulation against water losses, and limited ET (Restrepo and Arain, 2005; Yang et al., 2016). At our site, the NDVI showed little variability (Figures 2D, 4G) and decreased slightly during the mid-growing season drought in 2014 and 2015 (Figure 1E). More importantly, we found that lower P coupled with higher T_a and VPD, which can induce higher stomatal closure (Barron-Gafford et al., 2007), lead to suppressed g_s and ET (Figures 4A, B, D; Table 2). This result is comparable to the results obtained by previous research, showing that the suppression of ET by summer drought was not related to early canopy senescence but resulted primarily from reductions in stomatal conductance in deciduous forests in temperate regions (Zha et al., 2010).

We conclude that the timing of drought can affect annual ET in different ways and that spring drought primarily affects ET by restricting leafing out and canopy development, while summer drought primarily reduces stomatal conductance in this region. However, to compare the effect of drought in different periods on annual ET, longer (decade scales) time series and a wider range of VWC and P conditions are needed. In fact, the data suggest that the variations of ET during spring appear to have had a smaller contribution to IAV in annual ET in this study (Supplementary Figure S2A) in part because of the smaller proportion of spring ET (< 25%) in the annual total. However, the result that IAV in ET during spring and summer were both the primary factors determining the IAV in annual total ET (Supplementary Figure S2A) and IAV of annual ET being more sensitive to IAV of spring ET (Supplementary Figure S2B) may indicate that the variability in the intensity of drought in spring may have played a more important role in regular annual ET than summer drought. Moreover, the extended duration of the spring drought resulted in reduced annual ET (Figure 9), indicating that, in addition to its immediate negative impact on ET during spring, the spring drought may also have a more prolonged carryover effect on annual ET compared to the summer drought. Some studies have reported that an earlier spring uses the soil water reservoir sooner, which leads to summer drought in the northern hemisphere (Angert et al., 2005). In the context of climate change that less precipitation occurs as winter snow in temperate regions, without enough water supply from winter snowfall and/or seasonally frozen soil (Wu et al., 2018), forests across temperate China might become more vulnerable to changing hydrological processes and intensified drought stress. Thus, irrigation during early spring is viewed as a practical and efficient management measure for sustaining tree growth and ecosystem services in urban forests of Northern China.