Immediate and carry‐over effects of late‐spring frost and growing season drought on forest gross primary productivity capacity in the Northern Hemisphere

Forests are increasingly exposed to extreme global warming‐induced climatic events. However, the immediate and carry‐over effects of extreme events on forests are still poorly understood. Gross primary productivity (GPP) capacity is regarded as a good proxy of the ecosystem's functional stability, reflecting its physiological response to its surroundings. Using eddy covariance data from 34 forest sites in the Northern Hemisphere, we analyzed the immediate and carry‐over effects of late‐spring frost (LSF) and growing season drought on needle‐leaf and broadleaf forests. Path analysis was applied to reveal the plausible reasons behind the varied responses of forests to extreme events. The results show that LSF had clear immediate effects on the GPP capacity of both needle‐leaf and broadleaf forests. However, GPP capacity in needle‐leaf forests was more sensitive to drought than in broadleaf forests. There was no interaction between LSF and drought in either needle‐leaf or broadleaf forests. Drought effects were still visible when LSF and drought coexisted in needle‐leaf forests. Path analysis further showed that the response of GPP capacity to drought differed between needle‐leaf and broadleaf forests, mainly due to the difference in the sensitivity of canopy conductance. Moreover, LSF had a more severe and long‐lasting carry‐over effect on forests than drought. These results enrich our understanding of the mechanisms of forest response to extreme events across forest types.


| INTRODUC TI ON
Terrestrial ecosystems fix about 30% of anthropogenic carbon dioxide (CO 2 ) emissions and are key players in the global carbon (C) cycle (Ballantyne et al., 2012;Le Quéré et al., 2009). Even minor changes in terrestrial C pools could trigger large climate feedback (Arneth et al., 2006;Xing et al., 2022). For example, global warming can prolong the length of the growing season for forest growth, enhancing forest C sequestration, and increasing C sink, that is, negative C-climate feedback (Xu et al., 2016). On the other hand, the "greening" effect brought on by the extension of the growing season may be countered by warming-induced climatic extremes, which may reduce the C sink and result in positive C-climate feedback (Ciais et al., 2005). Climate change increases the intensity and frequency of climatic extreme events globally by increasing air temperature and changing precipitation patterns (Kharin et al., 2018). Forests, as long-lived terrestrial C pools, are inevitably influenced by climate extremes (Frank et al., 2015;Reichstein et al., 2013).
Significant inter-annual variability of the global C cycle has been detected due to changes in C fluxes in terrestrial ecosystems (Ballantyne et al., 2012). Gross primary productivity (GPP) is the main component of global C fluxes. It has been reported to be more than 12 times larger than anthropogenic C emissions (Bi et al., 2022;IPCC, 2022). Quantitative and conceptual models describe GPP as a function of vegetation properties and environmental factors.
Vegetation properties encompass physiological variables, such as the maximum rate of GPP at the leaf or canopy level, whereas key environmental factors include photosynthetically active radiation, temperature, and soil water content (Running et al., 2004). Early research considered GPP to have a linear relationship with photosynthetically active radiation and be restrained by other factors, which is inconsistent with reality (Landsberg & Waring, 1997). Later lightuse efficiency-based approach has been used to correct for nonlinearity between GPP and photosynthetically active radiation (Mäkelä et al., 2008). The model, however, still introduces bias in the estimation of GPP because soil moisture dynamics (Leuning et al., 2005) and foliar nitrogen concentrations (Cook et al., 2008) are not captured. Thus, the prediction of GPP's inter-annual variability using models is still a great challenge (Hu et al., 2018;Keenan et al., 2012), suggesting that more studies on the inter-annual variability of C fluxes are still needed.
GPP capacity is the maximum GPP under light saturation. A recent study (Musavi et al., 2017) proposed that this maximum rate of GPP is a good proxy for explaining inter-annual variability of terrestrial C fluxes since it can act as an indicator of the ecosystem's functional stability that reflects the ecosystem's physiological response to its surroundings. Indeed, under natural conditions, canopy photosynthesis seldom reaches light saturation. However, the light response curve can be used to estimate light-saturated rate of photosynthetic production (Reichstein et al., 2012), helping to reveal deeper relationships between forest physiological changes, ecosystem functional stability, and forest resource usage (Musavi et al., 2017;Xu, Xiao, Zhang, Ollinger, et al., 2020). Despite the potential importance of environmental factors for inter-annual variability in GPP capacity, only a few studies have addressed this relationship and its response to environmental factors (Bao et al., 2022).
Extreme events, including late-spring frost (LSF) and drought, have been found to affect forest productivity and stability and have received widespread attention (Frank et al., 2015;Marquis et al., 2020). The influence of LSF and drought on forest growth have been reported to be of similar importance (Vitasse et al., 2019) but through different mechanisms. Damage due to LSF is caused by a mismatch between the phenological rhythms and temperature changes (Hänninen, 1991), which has been called a "false spring." LSF can damage photosynthetic organs and depress forest productivity for an extended period (Benomar et al., 2022;Dittmar et al., 2006).
Drought stress arises from an imbalance between water supply and demand. It triggers tree dieback if hydrological stress is not promptly relieved (Di Francescantonio et al., 2020;Guarín & Taylor, 2005).
Similar to LSF, severe drought events also reduce tree growth and may lead to a prolonged reduction in forest productivity, although various species employ diverse acclimation and adaptation strategies against drought (Di Francescantonio et al., 2020). In general, LSF and drought affect peak summer GPP and, thus, annual GPP and GPP capacity by limiting photosynthesis (Musavi et al., 2017).
Numerous approaches, such as tree ring-based studies (D'Andrea et al., 2020;Gazol et al., 2019), eddy covariance data-driven studies Xu, Xiao, Zhang, Ollinger, et al., 2020;Zhang et al., 2016), remote sensing studies (Li et al., 2019;Schwalm et al., 2017), and modeling studies (Anav et al., 2015;Wolf et al., 2016;, have been extensively applied to the topic of extreme events. Whereas these studies have made important contributions to our understanding of forest resilience and resistance, the results have nevertheless remained ambiguous. This can be attributed to two reasons. First, the responses of forest ecosystems to climate extremes vary and are highly dependent on forest type or species (Renne et al., 2019;Teuling et al., 2010). Lee et al. (2021) and Sturm et al. (2022) claimed needle-leaf forests were more sensitive to drought events than broadleaf forests in Switzerland and Korea. By comparing ecosystem resistance and resilience at biome and global scales, Huang and Xia (2019) found that broadleaf forests had higher drought resistance and stability than needle-leaf forests under a global increase in drought occurrences. Second, previous studies have used various climate settings and assumptions for the modeling framework of GPP.  and  identified a slight fluctuation in the inter-annual variability of GPP globally using eddy covariance data-driven model tree ensemble, whereas Anav et al. (2015) showed large inter-annual variability of GPP by process-based models. These notably divergent results about the inter-annual variability of GPP from modeling studies also highlight the impacts of extreme events  and . Together, these findings demonstrate the urgency of model improvements, especially to capture impacts of extreme events on ecosystem productivity. Besides, those studies focused on the influence of single extreme event on ecosystem productivity, few studies have investigated cases in which both LSF and drought occur during the same year (Di Francescantonio et al., 2020;Gazol et al., 2019). The interaction between drought and late-spring frost can affect water balance and C metabolism of plants, further influencing forest vulnerability (Charrier et al., 2021). Thus, understanding these interactions can help us better predict populations and distributions of forests (Charrier et al., 2021;Gazol et al., 2019).
Forest growth states or meteorological conditions during the current growing season may influence subsequent growth, which are called carry-over effects (O'Connor et al., 2014;Ogle et al., 2015).
For example, Jin et al. (2020) found that vegetation growth has a significant carry-over effect on later growth via the changes in soil moisture through transpiration. Similarly, snow melting can promote to recharge water in deep soil, which benefits plant growth in the subsequent growing seasons (Lian et al., 2020). In forest ecosystems, carry-over effects caused by extreme events have been observed in above-ground primary productivity (Aubinet et al., 2018;Liu, Schwalm, et al., 2019;Liu, Yu, & Jia, 2019), stomatal activity , and forest mortality (Anderegg et al., 2013). These detected carry-over effects may be related to a variety of mechanisms, including canopy damage, root function loss, and water transportation damage, but ultimately point to a change in the climate sensitivity of the forest (Peltier & Ogle, 2019). The duration of carry-over effects has been reported to last several years (Anderegg et al., 2015). However, the current C cycle models simply assume that the physiological functioning of trees recovers almost immediately to its full state after stress has ceased, which does not seem to be realistic (Anderegg et al., 2015). Insufficient consideration of carry-over effects further increases uncertainty in the prediction of the global C cycle (Anderegg et al., 2015).
Here, we used a long-term time series of eddy covariance from 34 measurement stations in boreal and temperate regions to study different responses of annual GPP capacity to extreme events in needle-leaf and broadleaf forests. As extreme events, we considered late-spring frost (LSF) and growing season drought (hereafter drought). The specific aims of this study were (1) to quantify the immediate and carry-over effects of LSF and drought on GPP capacity in needle-leaf and broadleaf forests and (2) to investigate possible mechanisms behind the different responses of needle-leaf and broadleaf forests to drought and LSF using structural equation models.

| Carbon fluxes and meteorological datasets
GPP, ecosystem respiration, and meteorological datasets with half-hourly and daily time resolutions were obtained from the FLUXNET2015 dataset (https://fluxn et.fluxd ata.org/data/fluxn et201 5datas et/). The FLUXNET2015 dataset was aggregated into multiple time scales and subjected to standardized quality control and gap-filling procedures (Papale et al., 2006;Pastorello et al., 2020;Vuichard & Papale, 2015). To detect the drought and frost records, we initially selected sites that had at least 7 years of observations in the northern hemisphere. We then restricted our analysis to sites where at least 75% half-hourly GPP data were observed or gap-filled with high-quality data according to the quality flag in each year (e.g., flag 1 means good quality gap-filling). Forty sites fulfilled the above criteria. All sites were in North America or Europe, including two evergreen broadleaf forests, 11 deciduous broadleaf forests, five mixed forests, and 21 evergreen needle-leaf forests. A prior study pointed out that the difference both in anatomical and physiological structures between needle-leaf and broadleaf forests may relate to different drought responses (Isaac-Renton et al., 2018). Water transport in the xylem by tracheid is common in conifer species. Broadleaf species transport water by using xylem vessels. Besides, broadleaf species are reported to have more efficient water transport systems.
They have wider xylem conduits, which transport water from the roots to the leaves more rapidly, ensuring that the plant receives enough water for photosynthesis (Feild & Brodribb, 2000). The different anatomical and physiological structures for transporting water may regulate canopy stomatal conductance differently when drought stress appears, which may imply different water use and survival strategies (Liu, Schwalm, et al., 2019;Liu, Yu, & Jia, 2019;Tyree & Ewers, 1991). Thus, to analyze the different responses of forest GPP capacity to climate extremes between forest types, the sites were further divided into two main forest types-needle-leaf and broadleaf forests. Mixed forests were excluded. Needle-leaf forests (21 sites) only included evergreen needle-leaf forests. The broadleaf forests (13 sites) included both evergreen and deciduous broadleaf forests. All data analyses ( Figure S1) were based on these 34 sites ( Figure 1).

| Yearly GPP capacity calculation
Site-level annual GPP capacity was estimated from a non-rectangular hyperbolic light response curve function using half-hourly GPP and incoming photosynthetic photon flux density (PPFD) (Saito et al., 2009) as follows: where (mol CO 2 ⋅ mol ⋅ photon −1 ) is the initial slope of the light response curve of photosynthesis; it is also an indicator of the quantum yield of CO 2 assimilation at canopy scale; P max ( mol ⋅ CO 2 ⋅ m −2 ⋅ s −1 ) is the maximum photosynthetic capacity at light saturation; is the convexity of the curve; R d ( mol ⋅ CO 2 ⋅ m −2 ⋅ s −1 ) is daytime ecosystem respiration and PPFD ( mol ⋅ photon ⋅ m −2 ⋅ s −1 ) is the incoming photosynthetic photon flux density. (1) To fit the parameters of Equation (1), we used only the daytime dataset. The daytime dataset was filtered by PPFD (PPFD >4 mol ⋅ photon ⋅ m −2 ⋅ s −1 ) instead of the time interval (e.g., 08:00 am-18:00 pm), because the "midnight sun" phenomenon exists at the site in the north of the arctic circle (e.g., FI-Sod site with latitude of 67.362° N), which means that photosynthesis can still be active even at midnight. Friction velocity (U * ) with a threshold of 0.4 m ⋅ s −1 was then used to filter the low turbulence data (Knohl et al., 2003). Data were chosen using a 7-day moving window method, and Equation (1) was fitted using this data. According to Reichstein et al. (2012), a challenge for light response curve estimation is that the ideal light level under field conditions is realistically unreachable.
Hence, P max may not represent the real maximum photosynthesis capacity. We computed GPP at a PPFD of 2000 mol photon ⋅ m −2 ⋅ s −1 by Equation (1), which represents the light-saturated photosynthesis capacity (GPP2000) (Musavi et al., 2017;Xu, Xiao, Zhang, Ollinger, et al., 2020). We assigned the estimated parameters and GPP2000 to the middle date of the 7-day moving window. We only retained parameters ( , P max , ) with R 2 between raw and estimated GPP greater than 0.6 ( Figure S2). Thus, a robustly calculated daily GPP2000 was produced. We further transformed the daily GPP2000 to yearly by calculating 95 percentiles for given years across all sites.
Definitions of the photosynthetically active period GPP-based time series were used to detect the start and the end of growing season, since phenological measurements are unavailable for most sites. Prior studies have successfully linked the seasonal inflection points in the time series of GPP to determine the transition dates associated with the phenology of ecosystem processes (Richardson et al., 2010;Wu et al., 2012). We applied the exponential smoothing function in the "statsmodels" module (Seabold & Perktold, 2010) in

| Definition of late-spring frost and growing season drought
LSF was defined as a daily minimum temperature below −2.2°C between the start of the growing season and mid-July (Schwartz et al., 2006). A threshold of −2.2°C is widely used for calculations of the False Spring Index (Schwartz, 1993), which is used to signify the likelihood of damage to occur from a late-spring frost (Chamberlain et al., 2019). Although the threshold to initiate cold damage varies by species and phenostage, this criterion represents a hard freezing threshold known to cause potential mortality or severe tissue damage (Peterson & Abatzoglou, 2014). Besides, it was reported that actual temperature in plant tissues was even several degrees F I G U R E 2 Schematic for defining the start and end of the photosynthetically active period in one-year daily GPP data. Doy means day of year. F I G U R E 1 Site distributions in this study. Map lines delineate study areas and do not necessarily depict accepted national boundaries.

Longitude
Latitude Needle-leaf forest Broadleaf forest lower than air temperature due to radiative cooling (Leuning & Cremer, 1988). Plant tissues can be damaged by a day during the growth season with a daily minimum temperature below −2.2°C (even only several hours) (Dy & Payette, 2007).
We first sorted the half-hourly air temperature for each day and selected the daily minimum during the active growth period and thereafter detected the frost event. The absolute value of LSF, which represents the severity of frost during this period, is referred to as the late-spring frost severity (LSFS). We used the highest value of LSFS to represent extreme LSF for a given year since it indicates the strongest freezing stress that the forest was exposed to during the active growth (Hänninen, 1991), Overall, a total of 267 extreme LSF events were detected from 34 sites.
Unlike late-spring frost, which is a sudden and unpredictable event, drought is a gradual process. Typically, drought events that affect the ecosystem are initiated by meteorological events and water deficit develops slowly in the soil (Sturm et al., 2022). Thus, studies related to drought usually use drought index with monthly or multi-month scale. We used an unstandardized drought index-climatic water balance (CWB) to detect growing season drought. CWB was calculated by subtracting the monthly potential evapotranspiration using the "Thornthwaite" function (Thornthwaite, 1948) from the monthly precipitation in R (R Core Team, 2020). Finally, CWB was aggregated during the growing season across all sites as a proxy of water stress during growing season for a given year. As an alternative approach we also tested the Standardized Precipitation and Evapotranspiration Index (SPEI) to support the results from CWB. The time scale for SPEI was tested by Pearson correlation between SPEI and GPP2000 by using 4-10 months span ( Figure S3). Finally, The SPEI time scales were selected with 10 months and 8 months for needle-leaf and broadleaf forests, respectively ( Figure S3). We used the minimum SPEI value to represent the drought event for a growing season for a given year, as it represents the maximum cumulative water deficit during these growing periods. SPEI was calculated via the "SPEI" package (Vicente-Serrano et al., 2010) in R (R Core Team, 2020). Overall, the late-spring frost severity and drought index were calculated on different time scales to capture different processes between late-spring frost and drought.

| Canopy conductance calculation
Canopy conductance (G c , m ⋅ s −1 ) was calculated using the Penman-Monteith equation (Monteith, 1965) by half-hourly data: is the net radiation, G (W ⋅ m 2 ) is soil heat flux, (P a ⋅ K −1 ) is the psychrometric constant, ET (W ⋅ m −2 ) is the latent heat flux, Δ (P a ) is the slope between saturation vapor pressure and air temperature, and G a (s ⋅ m −1 ) is the aerodynamic resistance, which is calculated as follows (Launiainen, 2010): where u (m ⋅ s −1 ) is wind speed, u * (m ⋅ s −1 ) is friction velocity, and KB −1 is the dimensionless sub-layer Stanton number that we set as 2 (Kasurinen et al., 2014).

| Statistical analysis
2.5.1 | Analysis of immediate and carry-over effects of late-spring frost and growing season drought We used linear mixed models with random slopes and intercepts to analyze the immediate and carry-over effects of CWB and LSFS on GPP2000 in needle-leaf and broadleaf forests separately. For the immediate effects analysis, the initial linear mixed effect model was expressed as follows: where y ij is the GPP2000 for site i and year j, x ij is LSFS or CWB in year j at site i, 0 and 1 are the intercept and slope across all sites, respectively, 0 and 1 are the random intercept and random slope in site i, respectively, and ij is the random error.
For carry-over effects, only the variable that was significant in the immediate effects analysis was considered. The initial linear mixed model was formed as follows: where y ij is the GPP2000 for site i and year j, x i,j , x i,j−1 , and x i,j−2 are LSFS or CWB in year j, j − 1 and j − 2 at site i, respectively, 0 , 1 , and 2 are fixed effects, 0 , 1 , and 2 are random effects, and is the random error.
We then applied stepwise analysis to determine the random structure of the above full models, based on the Akaike information criterion. This led to some models in which all random effects were removed from the model. Linear mixed models and stepwise analysis were done using the "LmerTest" package in R (Kuznetsova et al., 2017;Lefcheck, 2016; R Core Team, 2020)). Multiple linear regression was conducted using the "lm" function in R (R Core Team, 2020) when no random effect was significant.

| Path analysis
Path analysis was used to evaluate the direct and indirect effects of key factors (Table 1) on GPP2000 in both needle-leaf and broadleaf forests. Path analysis has been widely utilized to quantify the complex causal links of interactions of biophysical factors on ecosystem functions (Astigarraga et al., 2020;Xu et al., 2018). Here, the paths were built based on hypotheses provided by published studies. For frost (Figure 3a), we assumed that (1) LSFS influenced canopy conductance (G c ) and the maximum quantum yield of CO 2 assimilation ( ) of canopy (Göbel et al., 2019;King & Ball, 1998) (Waring et al., 1995), and (3) LSFS, G c , and directly affected the GPP2000 (Landsberg & Waring, 1997;Waring et al., 1995). For drought (Figure 3b), we assumed that (1) air temperature (T a ), CWB and vapor pressure deficit (VPD) indirectly influenced GPP2000 by regulating G c (Berkelhammer et al., 2020;Steiner & Chameides, 2005), (2) (T a ) influenced the CWB and VPD (Sun et al., 2021), (3) CWB is influenced by VPD (Sun et al., 2021), and (4) all factors mentioned above except VPD affected the GPP2000 directly (Zhang et al., 2016). Piecewise structural equation model was used to evaluate our conceptual hypotheses because it permits the inclusions of random effects of the observation site (Shipley, 2013). Piecewise structural equation model provides marginal (R 2 m ) and conditional contribution (R 2 c ) of key factors and random effects in forest GPP2000 by the linear mixed model structure. R 2 m represents the variance explained by fixed effects only, whereas R 2 c reflects the variance explained by both fixed and random effects (Nakagawa & Schielzeth, 2013).
Based on the conceptual hypotheses, piecewise structural equation model was constructed using the "piecewiseSEM" package (Lefcheck, 2016) in R (R Core Team, 2020). Akaike and Bayesian information criterion were both used to test whether a random site effect was necessary. Lastly, the path coefficients were estimated using the 1000 times bootstrapping method by "semEFF" package (Murphy, 2021

F I G U R E 3
Conceptual diagrams showing (a) key paths of late-spring frost severity (LSFS) on canopy conductance (G c ), maximum quantum yield of CO 2 assimilation ( ), and forest GPP capacity at 2000 (GPP2000), and (b) the key paths between air temperature (T a ), vapor pressure deficit (VPD), climatic water balance (CWB), canopy conductance (G c ), maximum quantum yield of CO 2 assimilation ( ), and GPP capacity with 2000 incoming photosynthetic photon flux density (GPP2000).

| Distributions of CWB and LSFS for needleleaf and broadleaf forests
Generally, CWB as a proxy for drought showed evident variation across sites ( Figure 4a) and a distinctly varied interquartile range and distribution across forest types (Figure 4c; Figure S4). The range of standard deviation for the CWB was from 0.25 to 2 mm ( Figure 4a). The distribution of CWB was wider in needle-leaf than in broadleaf forests (Figure 4c). LSFS showed large variations across sites (Figure 4b). The standard deviation for the LSFS ranged from 1.33 to 12.00. The regions with needle-leaf forests had more severe frost occurrences during the growing season than the ones with broadleaf forests ( Figure S4). Moreover, LSFS exhibited distribution values in needle-leaf forests were higher than broadleaf, with 25%, 50%, and 75% quartile values of 6.73, 9.05, and 11.81, respectively ( Figure 4c).

| Estimations of the light response curve
With a PPFD increase from 0 to 2000 mol ⋅ photon ⋅ m −2 ⋅ s −1 , the response of GPP to light level in needle-leaf forests had an average tendency to be saturated at a mean maximum GPP of ~20 mol ⋅ CO 2 ⋅ m −2 ⋅ s −1 (Figure 5a). By contrast, the GPP in broadleaf forests continued to rise even at the highest light levels ( Figure 5b). The light response curve (Equation (1)) used in this study showed robust estimations of GPP ( Figure S7) as well as its parameters (e. g. , , P max , and ). The overall performance of the light response curve simulation with the mean R 2 (coefficient of determination) and the mean RMSE (root mean square error) were 0.83 and 3.45 mol ⋅ CO 2 ⋅ m −2 ⋅ s −1 , respectively (Figure 5c). showed a similar distribution pattern within needle-leaf and broadleaf forests, whereas GPP2000 showed a single-peak distribution and higher values in broadleaf than in needle-leaf forests, which had a bimodal distribution, with 25%, 50%, and 75% quartile values of 25.04, 29.98, and 34.04 mol ⋅ CO 2 ⋅ m −2 ⋅ s −1 , respectively ( Figure 5d).

| Immediate and carry-over effects of drought and frost on ecosystem photosynthesis capacity
To assess the immediate effects of extreme events, linear model analysis was applied. A significant negative relationship between LSFS and GPP2000 was observed in both needle-leaf and broadleaf forests (p < .001) (Figure 6b). However, the significant impact of drought on GPP2000 was detected only in needle-leaf forests (p < .01) (Figure 6a). No interaction between CWB and LSFS was found in either needle-leaf or broadleaf forests (Figure 6c).
Additionally, by comparing Figure 6a,c, it revealed that drought impacts were unaffected by presence or absence of LSF. The linear model results lead to very similar results when SPEI was used as drought index instead of CWB, with the exception that the drought effect was not statistically significant in needle-leaf forests when LSF and drought occurred in the same year (Table S2).
There was a considerable immediate influence of LSF in the present year in needle-leaf forests (p < .001), whereas drought had no carry-over effect on GPP2000 in both needle-leaf and broadleaf F I G U R E 5 General pattern of the light response curve of GPP in needleleaf (a) and broadleaf forests (b). Overall performance of the light response curve simulation (c). The means comparisons of the distributions of the maximum quantum yield of CO 2 assimilation ( ) and GPP with 2000 incoming photosynthetic photon flux density (GPP2000) between needle-leaf forest and broadleaf forests by using violin plots by using violin plot (d), the horizontal dashed lines inside the violin plot are the 25%, 50%, and 75% quartiles of the variable distribution from the bottom-up, respectively. The distribution was estimated by kernel density function. RMSE is the root mean square error with unite of mol ⋅ CO 2 ⋅ m −2 ⋅ s −1 calculated from raw GPP and estimated GPP by Equation (1) in each moving windows across all sites (see Section 2). R 2 is the coefficient of determination, calculated from raw GPP and estimated GPP by Equation (1) in each moving windows across all sites (see Section 2).

F I G U R E 6
Results of selected linear models for showing the effects of climatic water balance (CWB) (a) and late-spring frost severity (LSFS) (b) separately and combined (c) on GPP with 2000 incoming photosynthetic photon flux density (GPP2000) in needle-leaf and broadleaf forests. Gray bar means standard error of estimated coefficient. "*" indicates a significance at p < .05; "**" indicates a significance at p < .01; "***" indicates a significance at p < .001.

(a) (b) (c)
forests (p > .05) (Figure 7a). The strong drought effect diminished the following year, suggesting that the needle-leaf forests recovered from the drought within a year (Figure 7a). In both needle-leaf and broadleaf forests, the strong impact of late frost was clear in the following year (p < .05). After 2 years, the effect lessened notably in both needle-leaf (p = .367) and broadleaf forests (p = .103). The linear models confirmed these results when SPEI was used as drought index (Table S3).

| Different impact paths of biophysical factors on ecosystem photosynthesis capacity
To For drought, T a , CWB, and canopy G c had significant direct positive effects on GPP2000 in needle-leaf forests. High T a , high VPD, and low CWB significantly depressed GPP2000 via G c (Figure 8c; Table S4), whereas none of them had an effect in broadleaf forests ( Figure 8c,d). Canopy G c , a major mediator (which was influenced by other factors), was more responsive to water stress in needleleaf than in broadleaf forests (Figure 8c,d). GPP capacity was more sensitive to variations of G c during drought in needle-leaf than in broadleaf forests (Table S4). The p-values of Fisher's test were greater than .05 in both needle-leaf and broadleaf forests, indicating path analysis for drought was reasonable. These results were also confirmed by the path analysis when we used SPEI as drought index ( Figure S5).

| Immediate effect of late-spring frost on GPP capacity
We detected distinct immediate negative effects on forest photosynthetic capacity across forest types. The rapidly suppressed GPP capacity of both needle-leaf and broadleaf forests in the current year indicates the low tolerance of trees to late frost during the growing period. These results are similar to other studies of the effects of LSF on conifer tree species (Gazol et al., 2019) and broadleaf tree species (Allevato, 2019;Bascietto et al., 2018;Rubio-Cuadrado et al., 2021).
The reasons for these outcomes have been extensively expatiated by prior studies, which indicate that trees are more sensitive to frost F I G U R E 7 Carry-over effects of drought (a) and late-spring frost (b) on GPP capacity in needle-leaf and broadleaf forests. Drought severity was measured by the climatic water balance (CWB). Late-spring frost severity (LSFS) was measured by minimum daily temperature, which is under the threshold of −2.2°C (for details, see Section 2). Gray bar means standard error of estimated coefficient. "*" indicates a significance at p < .05; "**" indicates a significance at p < .01; "***" indicates a significance at p < .001.

(a) (b)
damage during the growing season because new emerging leaves tend to be more susceptible to low temperatures than dormant buds during the winter (Chamberlain et al., 2021;Dittmar et al., 2006).
As a result, LSF may readily damage leaves and canopies, thereby reducing photosynthetic capacity throughout the entire year.
LSF may also hinder photosynthesis by causing photoinhibition (Tyystjärvi, 2013). Photoinhibition is usually fully reversible, as it is taken care of by the protective mechanisms of plants (Krause, 1988).
By contrast, frost damage can lead to tissue damage and cell death of living biomass, as well as necrosis of leaf and meristem tissues.
In our study, path analysis showed that suppression of LSF in forest GPP2000 was mostly due to direct effects rather than indirect effects (Figure 6a,b). This reveals that, rather than temporary photoinhibition, LSF may have induced tissue damage and prolonged the recovery period to years, which reduced GPP capacity after LSF events.
There was no interaction between LSF and drought, neither in needle-leaf nor in broadleaf forests (Figure 6c). This implies that the simultaneous occurrence of these two climatic extreme events had the same impact on growth as the individual occurrence of either event. These findings are consistent with a recent study of tree rings in Spain, which also reported no significant interaction between LSF and summer drought (Gazol et al., 2019). Although we have different definitions of LSF and drought events, comparable results of both studies highlight the importance of forest phenology. Phenological events (e.g., timing of leaf unfolding) have high variability between forest type and even within forest type, resulting in a wide range of LSF tolerance of species. The importance of determining LSF using flexible phenology calculation cannot be overstated, especially given that no situ-based phenology observations are available. In our case, we simultaneously determined LSF events and applied a method that coupled yearly specific growing phenology by daily GPP observation and a temperature-based threshold. As the LSF threshold, we used a moderate temperature threshold of −2.2°C (Ma et al., 2019).
Other studies have used a range of threshold temperatures to detect LSF and concluded that LSF can significantly depress the growth of plants (Dittmar et al., 2006;. Our sensitivity analysis for LSF also supports this conclusion ( Figure S6). It should be noted that conditions that predict the appearance of LSF damages are still debated, given that LSF tolerances vary from organ level to ecosystem level (Augspurger, 2009;Rubio-Cuadrado et al., 2021).

| Different responses of needle-leaf and broadleaf forests to growing season drought
Our study revealed that GPP capacity in needle-leaf and broadleaf forests responded differently to drought. Needle-leaf forests were more sensitive to drought than broadleaf forests (Figure 6a). The varied responses of drought between needle-leaf and broadleaf forests were also reported by Zhang et al. (2006), who indicated that GPP F I G U R E 8 Path analyses exploring the effects of late frost (a, b) and drought (c, d) with numerous factors on GPP2000 in needleleaf and broadleaf forests. LSFS, late-spring frost severity. G c , canopy conductance. , maximum quantum yield of CO 2 assimilation. T a , air temperature. CWB, Climatic water valance. VPD, vapor pressure deficit. Red and blue indicate the positive and negative effects of path coefficients, respectively. Solid and dashed line types indicate the significant effect and non-significant effect of path coefficients, respectively.  et al., 2008). By contrast, the response of GPP capacity in needleleaf forests in our dataset strongly regulates their stomatal conductance. The closing of stomata reduces hydraulic efficiency and, consequently, productivity, closely resembling isohydric behavior (Berkelhammer et al., 2020). However, this quick reaction prevents xylem conduits from embolization and maintains hydraulic safety, albeit at the expense of photosynthesis (Blackman et al., 2009). The significant control of canopy stomatal conductance to GPP was also reported by Ciais et al. (2005), who demonstrated that during the 2003 European drought event, 13 of 14 sites showed distinct GPP reduction due to significantly decreased canopy conductance, despite different site properties and drought intensity. Their study also clearly showed that the most significant GPP reduction in annual scale occurred in a needle-leaf forests site, whereas the slightest GPP reduction occurred in a broadleaf forests site.
Further, non-stomatal limitation should be considered as a factor to limit photosynthesis in needle-leaf forests, given that the path analysis for drought only explained 37% of GPP2000 (Figure 8c).
The results suggest that there may be more key processes that we missed in our analysis. Non-stomatal limitations include decreased electron transport capacity (Epron & Dreyer, 1992), reduced mesophyll conductance (Flexas et al., 2012), and lower Rubisco activity (Parry et al., 2002). The path analysis indicated that the direct effect of α was significant in needle-leaf and broadleaf forests; however, α in needle-leaf forests was suppressed by T a (Figure 8c). More importantly, a suppressed α may limit CO 2 assimilation rate by restricting electron transport capacity (Yamori et al., 2010). This pathway, therefore, implies that non-stomatal limitations may exist in needleleaf forests. Gourlez de la Motte et al. (2020) and Zhang et al. (2006) also reported that non-stomatal regulations during drought may play a key role in regulating forest photosynthesis. In addition to this, our results highlight that CWB directly impacted GPP2000 in needle-leaf forests (Figure 8c,d). There is a possibility that drought affects photosynthesis by directly restricting C metabolism and potentially influencing the integrity of cells in needle-leaf forests (Hasanuzzaman et al., 2013). Last, the discrepancies in availability of (deep) ground water between needle-leaf and broadleaf forests may also be a potential cause. Species with deep root systems could maintain stomata opening by taking advantage of deep soil water and thus, delay or even avoid the drought influence on plant growth (Phillips et al., 2016). Sturm et al. (2022) also reported that Norway spruce with shallow root depths is more sensitive to drought than European beech with deeper root system.

| Carry-over effect of late-spring frost and growing season drought
Our study highlighted the notable carry-over effects of LSF on forest photosynthesis capacity, whereas no significant carry-over effect of drought was detected (Figure 7a,b), suggesting that LSF has a more lasting impact on the photosynthetic capacity of forests. The reason is that for many trees, earlier budburst exposes leaves to LSF when their leaves are sensitive to frost damage, and spring frost can lead to long-term leaf damage with reductions in photosynthetic production (Dittmar et al., 2006). This is also reflected in our study. If new buds and leaves do not regrow promptly after LSF during the growing season, a significant decrease in productivity can be observed in the following years. Dittmar et al. (2006) reported a significant reduced growth of common beech forests in southern Germany in subsequent years after LSF. By collecting 2844 tree ring chronologies, Vitasse et al. (2019) found that the complete recovery of oak and beech growth was around 2 years in Switzerland. Additionally, when ice starts to melt, water-diluted gases form bubbles that spread through the xylem conduits, interrupting the continuous flow of water, resulting in cavitation, and delaying plant recovery. This secondary damage is thought to be a contributing factor in how efficiently forests recover (Choat et al., 2011).
The recovery of vegetation from drought is determined by the severity of the drought, the intrinsic recovery capacity of the plant, and the water supply (Choat et al., 2018). In our study, unlike the longlasting effects of LSF on both needle-leaf and broadleaf forests, rapid recovery from drought was observed in needle-leaf forests within a year ( Figure 7a). This finding is in line with that of Schwalm et al. (2017), who found that the recovery time after droughts for the major forest types ranges from 1 to 6 months globally, regardless of soil, fire, or land use. However, contrasting results have been observed where carry-over effects of drought existed for more than 1 year (Anderegg et al., 2015;Luo et al., 2022). The absence of carry-over effects of drought in this study might have been affected by the following two reasons. First, in our study, drought severity was measured by climatic water balance or minimum SPEI value during growing season to represent the accumulated water stress to the forest. However, the climatic water balance does not consider the extent of soil water storage nor the recharge process (Piedallu et al., 2013), which causes bias for the detection of drought events and the estimation of carry-over effects of drought. Second, the range of mean climatic water balance in our study was −200 to 600 mm (Figure 4), which can be considered mild droughts. In our study period, drought severity may not have been strong enough to cause carry-over effects, and forests recover quickly from minor droughts. The recovery time could be even shorter if sufficient water (e.g., rain events) was supplied after drought. This would, for one thing, reduce vulnerability to water deficit and, for another, boost the physiological activity of plants (Bréda et al., 2006). Carry-over effects are also related to the availability of stored nonstructural carbohydrates since they can contribute to ecosystem functional stability by helping plants adapt to environmental stress.
Nonstructural carbohydrates can be used for producing secondcohort leaves or repairing the xylem when photosynthates are insufficient after extreme events (Richardson et al., 2013). Needle-leaf species usually produce new leaves less frequently than broadleaf species. Moreover, for broadleaf species, second-cohort leaves may compensate productivity for late-spring frost damage during autumn (Zohner et al., 2019). Thus, after drought or frost damage, needle-leaf forests are expected to suffer more damage than broadleaf forests.
Instead, our results showed that needle-leaf forests and broadleaf forests were similarly affected by both immediate and carry-over effects during or after late-spring frost (Figures 6 and 7). We identified two potential reasons. First, although broadleaf species can regrow new leaves at the same growing season, the low level of reserved nonstructural carbohydrates cannot compensate the loss of new buds even if the reserved nonstructural carbohydrates are insufficient, which further limits the growth of plants and weakens the stability of forests (D'Andrea et al., 2021). Second, the growth of needle-leaf species was regarded as more dependent on the number of new buds from last year. However, needle-leaf species usually have evergreen leaves that persist for several seasons, and photosynthesis is, therefore, compensated by old leaves if new needles are damaged by LSF (Vitasse et al., 2019). In summary, the dynamic of nonstructural carbohydrates together with leaf phenology could lead to similar immediate and carry-over effects of LSF in the two forest types (Vitasse et al., 2019).

| CON CLUS IONS
This study highlighted the diverse response paths of GPP capacity to growing season drought in needle-leaf and broadleaf forests, with canopy conductance sensitivities across different forest types serving as the primary drivers. Broadleaf forests were more drought resistant by reducing the impact of stomatal regulation on photosynthesis, which is more comparable to anisophydric behavior.
Whereas needle-leaf forests were strongly dependent on stomatal conductance regulation, resembling isohydric behavior. However, both needle-leaf and broadleaf forests experienced comparable negative effects from LSF on GPP capacity. LSF not only immediately reduced photosynthetic capacity but also caused a more severe and long-lasting impact on forests. Nevertheless, the impact of drought persisted only for the prevailing year. Furthermore, when LSF and drought coexisted, the LSF effect did not alter the drought effect.
No interaction between LSF and drought was found in either needleleaf or broadleaf forests. Moreover, our path analysis indicated that more concurrent and long-term observations of phenology and environmental patterns are needed to further refine our knowledge of the requisite conditions that predict when extreme events will occur.
Overall, this study enriches our understanding of the mechanisms of forest response to LSF and drought across different forest types.

ACK N OWLED G M ENTS
We are grateful to all anonymous reviewers for their valuable com- Thematic Center and the OzFlux, ChinaFlux, and AsiaFlux offices.

CO N FLI C T O F I NTER E S T S TATEM ENT
All other authors declare that they have no competing interests.