Introduction

The ongoing climate change increases the frequency and severity of extreme weather events (IPCC 2013). One of the key ecological parameters affected by changing climate is primary production. Extreme drought events have been shown to reduce primary production in Europe (Ciais et al. 2005) and the increase in drought frequency negatively affected grassland production worldwide (Zhao and Running 2010). At plot scale, climate manipulation experiments are particularly effective way to study ecological consequences of climate change (Wu et al. 2011), especially long-term multi-site field experiments (Kröel-Dulay et al. 2015).

In climate manipulation experiments, effects of precipitation treatments on vegetation performance are often estimated once per year, during or right after seasonal treatments (Köchy and Wilson 2004; Brancaleoni et al. 2007; Mänd et al. 2010; Byrne et al. 2013; Tielbörger et al. 2014). However, the effects emerging in the treatment period may change during the rest of the growing season. Thus, additional within-year measurements may provide further information about the changes in plant biomass, either targeting legacy effects in long-term experiments before the yearly treatments, or aiming to follow the relaxation period after. Aboveground plant biomass may recover after the drought period in semiarid grassland and shrubland communities adapted to high variability in precipitation (Miranda et al. 2011). Even if drought strongly reduces aboveground biomass, it can recover quickly during the late summer, because belowground parts are less affected by the drought (Shinoda et al. 2010). Therefore, treatment effects should be checked multiple times during the growing season.

Conducting multiple within-year measurements on aboveground plant biomass is one of the major challenges in long-term field experiments. For this purpose, application of non-destructive sampling methods is suggested (Gamon et al. 1995). When effects on primary production are in focus, field spectroscopy is one of the feasible solutions for estimating aboveground biomass, or leaf area index (Goodin and Henebry 1997; Pontailler et al. 2003; Mulla 2013; Nestola et al. 2016; Ónodi et al. 2017a). The normalized differential vegetation index (NDVI) obtained by field spectroscopy is an accurate proxy for aboveground biomass estimations (Gamon et al. 1995; Ónodi et al. 2017b). However, it is rarely applied in multi-seasonal measurements in long-term ecological experiments, but see (Goodin and Henebry 1997; Filella et al. 2004; Boelman et al. 2005; Wang et al. 2016; Nestola et al. 2016).

Our goal was to study the effect of 2-month drought treatments (i.e., rain exclusions) on the aboveground biomass, and primary production via proxies (NDVI, and the sum of positive NDVI increments accordingly) both during the treatments as well as during the whole growing seasons. We applied field spectroscopy to observe within-year changes in aboveground green biomass of vascular plants in the semiarid Kiskunság forest-steppe vegetation (Lellei-Kovács et al. 2008a; Kröel-Dulay et al. 2015). According to climate change scenarios for Hungary, the frequency of extreme dry and wet years is expected to increase in the study region (Bartholy et al. 2003; Bartholy and Pongrácz 2007).

Our specific questions were as follows: What is the effect of drought treatment on the aboveground biomass estimate (NDVI) in different seasons in a long-term climate manipulation experiment? What is the effect of drought treatment on annual primary production and production of different seasons, i.e., treatment and post-treatment changes of biomass calculated as the sum of positive NDVI increments?

Methods

Study site and experimental design

Our study site is part of the EU FP5 VULCAN and the EU FP7 INCREASE projects (Beier et al. 2004; Peñuelas et al. 2007; Kröel-Dulay et al. 2015) representing the continental semiarid forest-steppe vegetation of Central Europe in the multi-site surveys. The site is in the Kiskunság National Park (N46°52′, E19°25′), in a Pannonian sand forest-steppe vegetation mosaic (Lellei-Kovács et al. 2008b) of high plant diversity and nature protection value (Fekete et al. 2002; Molnár et al. 2012). In our study plots, we sampled open grassland patches where also shrubby root suckers of white poplar (Populus alba) occurred. The soil is calcaric arenosol which enhances the semidesert character of the vegetation. Climate of the study area is temperate continental. The vegetation period starts in April and finishes in October. Based on regional 30 years average values (1961–1990), mean annual temperature is 10.4 °C, mean monthly temperature ranges from − 1.9 °C in January to 21.1 °C in July, while mean annual precipitation is 505 mm with a peak in June (Kovács-Láng et al. 2000).

The climate manipulation experiment described by Lellei-Kovács et al. (2008a,b ) was conducted in three replications of controls and drought treatments. The vegetation of the replicates differed from each other in the abundance of poplar shoots, but within each control–drought treatment pair, the plots were similar in this respect. Plot size was 4 m × 5 m, and the experiment started in 2002. Automatically controlled rain exclusion during May and June was applied as drought treatment.

Sampling design and data collection

In our study, we estimated aboveground green biomass of vascular plants (referred as plant biomass hereinafter) by means of NDVI in the control and drought treatment plots in 2011 and 2013. The planned 2012 measurements had to be cancelled for technical reasons. We applied a multi-seasonal non-destructive plant biomass sampling method (Fig. 1). As the first step, we measured a baseline of the plant biomass at the beginning of the vegetation period of the first year (M0 in Fig. 1), when plant activity is still very low as the soil surface is covered by litter. Afterwards, we estimated the plant biomass three times a year: at the turn of April and May (before-treatment measurement, M1), at the turn of June and July (after treatment measurement, M2), and after a relaxation period at the turn of August and September (end-of-summer measurement, M3). Precipitation was monitored in all plots separately. Rain exclusion data were calculated as differences between average values collected in the three control and the three drought treatment plots. Annual and monthly precipitation data were calculated as average values of the control plots (Online Resource 1).

Fig. 1
figure 1

The timing of measurement events (M0–M3). Horizontal bars stand for time intervals of drought treatments. Vertical bars are proportional to the monthly precipitation, while the unfilled parts of the vertical bars show the amounts of excluded precipitation during the drought treatments. The heights of the bars range from 0 mm (November 2011) to 126.3 mm (March 2013). See dates and more values in the text, as well as in Online Resource 1

Full size image

In 2011, drought treatment started at 30 April and ended at 07 July. During this period, we excluded 88.8 mm out of the 112.8 mm precipitation (78.7%). Annual precipitation in 2011 was 408.0 mm. Dates of the biomass estimation measurements were: (M0) 01 April, (M1) 02 May, (M2) 28 June, and (M3) 30 August.

In 2013, drought treatment started; later, it was conducted between 15 May and 30 June. During this period, we excluded 111.7 mm out of the 118.4 mm precipitation (94.4%). However, 30.6 mm rain was not excluded during the first 2 weeks of May. Annual precipitation in 2013 was 597.8 mm. Dates of the biomass estimation measurements were: (M1) 29 April, (M2) 10 July, and (M3) 04 September. Thus, 111.7 mm precipitation out of 149.0 mm (75.0%) was excluded between the M1 and M2 measurements.

We estimated the amount of plant biomass by non-destructive field spectroscopy techniques in each measurement event (Online Resource 2). We applied a portable Cropscan MSR87 multispectral radiometer (Cropscan, Inc., Rochester, MN) for measuring incoming and reflected light intensity. We used an aluminium frame for moving the sensor above the plots at a height of 1.5 m. In each of the six plots (three control and three drought-treated plots), we sampled twelve subplots arranged in a 3 × 4 grid. The area of the circular subplots was 0.44 m2 (diameter: 0.75 m), and the distance between centre points of the neighbouring subplots was 1 m. The frame allowed us to repeat the sampling of each subplot at the same position during the different measurement events. We calculated NDVI (Rouse et al. 1974) values based on the measured light intensity data at red (660 nm) and near infrared (810 nm) wavelengths. According to our previous investigation, NDVI provides an accurate proxy for plant aboveground green biomass estimation in the studied vegetation complex (Ónodi et al. 2017b). Thus, differences in NDVI values are interpreted as differences in aboveground green biomass henceforth. Baseline NDVI data collected at the first (M0) measurement event (NDVIAVG±SE = 0.205 ± 0.003) provide an empirical zero point for calculation of increments of yearly plant biomass. The 0.205 average is in agreement with our long-term experience (Ónodi et al. 2017a) and the low standard error value which we got shows that the baseline is not sensitive to the differences in litter cover and composition. We consider the increase in NDVI as proxy for aboveground primary production. Thus, we count the sum of the positive increments as proxy for the annual aboveground net primary production (ANPP), according to Sala and Austin (2000).

Statistical analyses

In the first analysis, dependence of the measured NDVI values on treatments, years, and measurement events and their interactions (including three-way interaction) were analysed by fitting linear mixed models (Zuur et al. 2009). In this analysis, subplots nested in plots were random factors in the model, since simplification of the random part would result in higher AIC values. To avoid this, while not losing the inside-plot variation information, we applied the nested design, in line with Colegrave and Ruxton (2018). Significance of fixed factors was tested by maximum-likelihood ratio tests (Zuur et al. 2009).

The following null-hypotheses were tested using contrasts. The hypothesis 1 refers to the measured NDVI values in each sampling date. The hypotheses 2–4 refer to the changes of the NDVI values in time, and they are arranged into pairs where (a) probe whether there is a significant increase or decrease in the given time span at the level of a certain treatment, and (b) compare the changes between control and drought.

  1. 1.

    NDVI values in control and drought treatments do not differ (tested in each measurement event);

  2. 2 (a)

    changes in NDVI between M1 and M2 (hereafter called treatment change) do not differ from zero;

  3. 2 (b)

    treatment changes do not differ between control and treatment plots;

  4. 3 (a)

    changes in NDVI between M2 and M3 (hereafter called post-treatment change) do not differ from zero;

  5. 3 (b)

    post-treatment changes do not differ between control and treatment plots;

  6. 4 (a)

    changes in NDVI between M1 and M3 (hereafter called whole-season change) do not differ from zero;

  7. 4 (b)

    whole-season changes do not differ between control and treatment plots.

P values were corrected by single-step procedure (Hothorn et al. 2008) to avoid their inflation due to multiple testing.

In the second analysis, the sum of positive NDVI increments for each subplot, as a proxy for ANPP was the dependent variable, while year and treatment were fixed factors in the model. The random part was the same as in the previous analysis. Significance of fixed factors was tested by series of maximum-likelihood ratio tests (Zuur et al. 2009).

All calculations were done in R statistical environment (R Core Team 2017) using nlme (Pinheiro et al. 2017), multcomp (Hothorn et al. 2008), and lsmeans (Lenth 2016) add-on packages for fitting models, doing post hoc tests, and drawing figures, respectively.

Results

We found significant three-way (treatment × year × measurement event) interaction (likelihood ratio = 12.875, df = 2, p = 0.002) on the NDVI data, and thus, treatment effects had to be tested in each sampling time by post hoc test using contrasts. Post hoc tests showed that drought treatments significantly affected NDVI values after treatment measurements (M2), in both years and also at the end-of-summer measurement (M3) in 2013 (Table 1, upper six rows, NDVI with C vs. D comparisons). However, the differences are not significant in the other sampling times, even if NDVI values were higher in control plots in all six measurement events (Fig. 2, positive estimates in Table 1).

Table 1 Comparisons tested using contrasts in the mixed-effect linear model fitted to NDVI
Full size table
Fig. 2
figure 2

NDVI values (least-square means and 95% confidence intervals estimated by the fitted mixed-effect model) for the measurement events in 2011 and 2013 in the control (C) and drought treatment (D) plots (see also Online Resource 2); before treatment: M1, after treatment: M2, end-of-summer: M3. Asterisk denotes significant drought effect

Full size image

Regarding the increase or decrease of plant biomass between the measurement events, we found a significant increase of NDVI values during the treatment change (M2–M1 in Fig. 3; M1 vs. M2 in Table 1) and its significant decrease during the post-treatment change (M3–M2 in Fig. 3; M2 vs. M3 in Table 1) except the drought treatment in 2011. There were no significant changes in the NDVI values in the whole-seasons (M3–M1 in Fig. 3; M1 vs. M3 in Table 1).

Fig. 3
figure 3

Estimated changes of NDVI values between measurement events (least-square means and 95% confidence intervals): treatment change, i.e., M2–M1; post-treatment change, i.e., M3–M2; whole-season change, i.e., M3–M1; in 2011 and 2013 in the control (C) and drought (D) treatments; asterisk denotes significant drought effect

Full size image

Regarding the treatment effects on the changes of plant biomass during the treatment periods, we found significantly higher biomass increase in the control than in the drought treatment in both years (M2–M1 in Fig. 3; C vs. D comparisons of ∆NDVI (M2–M1) in Table 1). In the post-treatment periods, the plant biomass decrease was significantly greater in the control than in the drought treatment in 2011 (M3–M2 in Fig. 3; C vs. D comparisons of ∆NDVI (M3–M2) in Table 1).

In the analysis of the sums of the positive increments as proxy variables for ANPP (Fig. 4), the two-way interaction between treatment and year proved to be significant (likelihood ratio = 8.809, df = 1, p = 0.003). Effect of treatment (likelihood ratio = 16.046, df = 1, p < 0.001) was significant in both years (2011: z = 2.224, p = 0.040; 2013: Z = 3.823, p < 0.001); however, it was stronger in 2013 (t = 3.018, df = 70, p = 0.004).

Fig. 4
figure 4

Sum of positive NDVI increments from the 0.205 start-of-season baseline (M0) to the end-of-summer measurements as a proxy for annual aboveground net primary production, in 2011 and 2013 in the control (C) and drought treatment (D) plots (least-squares means and 95% confidence intervals). The difference between control and drought treatment is significant in both years and its value is significantly greater in 2013

Full size image

Discussion

Based on multiple NDVI measurements, we found consistent negative effects of drought treatment both on yearly peak plant biomass and on the ANPP, in line with Estiarte et al. (2016) and Reinsch et al. (2017). Drought treatment decreased the biomass in both years in June (M2 in Fig. 2), and at the end of summer in 2013 (M3 in Fig. 2). However, NDVI values showed no overall significant treatment effect. Treatment and measurement event had interactive effects on biomass, similar to Hoover et al. (2014) who also found both significant year effect and significant year × drought treatment interactions in their 2-year extreme drought and heatwave experiment in central U.S. grassland. In our study, we showed that besides the significant treatment effect in June, the plant biomass did not differ in the treated and the control plots at the beginning of the studied vegetation periods (M1 in Fig. 2) and at the end of summer in 2011 (M3 in Fig. 2). The treatment and post-treatment changes of NDVI values (Fig. 3) show also strong effects of drought. While the biomass increased markedly in the control plots, we did not find increment in drought plots in both years (M2–M1 in Fig. 3), only in 2013, when it was significantly less than in the control. Furthermore, the post-treatment biomass decrease was also less in 2011 compared to the control (M3-M2 in Fig. 3). Consequently, the estimated ANPP decreased in the case of our drought treatment (Fig. 4), similar to the findings of most studies in arid or semiarid ecosystems (Beier et al. 2012).

The NDVI values responded sensitively to the treatments. The detected treatment effects depended on the relative timing of treatments and measurement events. Delay of starting the treatment resulted in detection of significant biomass increase during the treatment period also in the drought treatment plots in 2013 (M2–M1 in Fig. 3), even if this increase was significantly smaller compared to that in the control plots. We assume that the reason for the biomass increase is that the study site had 30.6 mm precipitation during the 2 weeks long delay period, which promoted a significant vegetation growth also in the drought-treated plots.

We applied multiple sampling of biomass in a year to gain deeper knowledge on the pattern of plant biomass changes in grasslands. First of all, multiple biomass estimates are required for monitoring the amount of biomass in the course of the vegetation season, revealing which periods of the growth season were affected by the treatment. We found that drought eliminated peak biomass in June (M2 in Fig. 2, as well as M2–M1 in Fig. 3), characteristic for the open sand grasslands (Kovács-Láng 1974), while it has slight or no effect at the early and late season stages. On the other hand, multiple estimates are required for assessing the primary production following the method of the sum of positive increments in plant biomass (Sala and Austin 2000). This allows a more reliable comparison of ANPP than estimation only using a measurement of peak biomass (Scurlock et al. 2002). The method which we applied is based on the calculation of the positive increments between repeated measurement events and it needs an estimate for the baseline. For this purpose, we executed a sampling plan which covers the starting point (M0) and three measurement events (M1, M2, and M3) during the vegetation period. Application of the sum of positive increment method allowed us to take the biomass at the time of all the three measurement events into account. Besides the already mentioned drought effect (Fig. 4), we detected that difference between ANPPs in control and drought plots was higher in 2013. This is in accordance with the fact that spring precipitation in 2013 was much higher than in 2011 (Fig. 1), which resulted in higher peak biomass in the control. Furthermore, the late summer drought in 2013 (Fig. 1) prevented the regeneration in the drought-treated plots. However, our ANPP estimate, being mostly governed by M2–M1 difference, is not sensitive to the regeneration of the plant biomass by the time of the next treatment.

We emphasize the importance of biomass measurements multiple times in the growing season in an experiment where yearly drought treatments are applied, in contrast to most of the studies from which only annual data are published. Our results supplement the findings of Estiarte et al. (2016) and Reinsch et al. (2017) who got consistent drought effect applying one annual biomass estimation by point-intercept method right after the treatment period. Our study reveals that in late-successional grassland-shrubland ecosystems, like ours (Kröel-Dulay et al. 2015), compensation may occur before the next drought. With one measurement per year, we could only detect the effect of drought treatment on the peak biomass. Although our investigation started in the 10th year of the climate change field experiment, we could not observe general treatment effect on the biomass taking three annual measurements into consideration. While both summer drought treatments caused significant differences in NDVI by the end of the treatment periods, among four before-treatment and end-of-season measurements, only one showed a significant treatment effect. Furthermore, we found no whole-season (M3–M1) differences in NDVI between the control and treated plots. Thus, the studied ecosystem proved drought resistant both in terms of Vicente-Serrano et al. (2013), reacting to the drought only at a short time scale, and according Hoover et al. (2014), recovering by the end of the season. This resistance is in agreement with the findings of Tielbörger et al. (2014) in long-term experiments in Mediterranean shrublands. We suppose that the main reason for rapid recovery of biomass in the studied vegetation mosaic is that the drought treatment did not lead to regime shift which occurs after strong disturbance events (Kröel-Dulay et al. 2015). The presence of poplar shoots might contribute to the late season recovery of the grass layer after drought through shading, in line with the findings of Erdős et al. (2014). In contrast with our results, in the post-fire successional vegetation of the Catalonian VULCAN site, Filella et al. (2004) found long-term around-the-year divergence in biomass (also estimated by NDVI) due to drought treatment from the first year after the start of the experiment, which is in accordance to the findings of Kröel-Dulay et al. (2015) in early successional ecosystems.

According to Ónodi et al. (2017a), drought can temporarily change the NDVI–biomass relationship. Several structural and physiological changes may result in lower NDVI readings, such as decreased specific leaf area, light absorptance, and green biomass to standing biomass ratio because of drought treatment (Cornic and Massacci 1996). However, Filella et al. (2004) and Mänd et al. (2010) found NDVI a reliable proxy for biomass estimation across treatments, seasons, and sites in the same experimental design. As there were not remarkable long-term compositional changes in the vegetation due to the drought treatment at our site (Kröel-Dulay et al. 2015), we conclude that the lower NDVI value after drought treatments indicated less aboveground green biomass because of increased drying and reduced sprouting.

The loss of biomass peak in consecutive years due to drought could lead to severe changes in the carbon budget of the ecosystem. Nagy et al. (2007) found that net ecosystem exchange (NEE) in semiarid grasslands of the same ecosystem can turn to positive (i.e., carbon releasing) in dry years. However, according to Pintér et al. (2008), the NEE in the same vegetation type is negative (i.e., carbon accumulating) in years of normal or above normal precipitation. Our finding that plant biomass recovers by the next drought treatment show the resilience of this drought-adapted vegetation. Considering the long-term climate prediction of increasing frequency of both extreme dry and extreme wet years (Bartholy and Pongrácz 2007), there is no direct danger of desertification in the studied community, as the carbon loss in dry years can be compensated by carbon accumulation in wet years.

In conclusion, we want to underline two of our findings. First, by means of application of field remote sensing, we demonstrated the negative effect of drought treatment on the aboveground plant biomass and the ANPP in a diverse semiarid shrubland–grassland community. At the same time, we showed that only one yearly measurement right after the treatment may overestimate the effect of drought, disregarding the compensation processes of late-successional ecosystems, which can be detected using multiple within-year measurements.