Highlights

  • Tree growth and spectral indices decoupled during drought-triggered dieback.

  • Dieback is a consequence of recurrent, severe drought events.

  • Interactive tree and stand-level responses drive forest dieback.

Introduction

Climatic extremes can impact forest functioning at multiple scales and affect a wide range of processes, from tree physiology (Salmon and others 2019; Rubio-Cuadrado and others 2021) and phenology (He and others 2018), to growth (Seifert and others 2017; Abiyu and others 2018) and stand productivity (Anderegg and others 2019; Khoury and Coomes 2020). Despite increases in the magnitude and frequency of extreme events acting as one of the main drivers of forest damage and mortality (Anderegg and others 2020), low-intensity climatic events can also be a cause of forest dieback (sensu (Mueller-Dombois 1988), that is “the unseasonal loss of crown foliage (partial or complete) of many trees of a stand”) and mortality. These events can lead to cumulative and carryover legacy effects (Franklin and others 1987; Sánchez-Pinillos and others 2021) resulting in reduced biomass (Ma and others 2012).

Climate-driven die-off and mortality events occur because the cumulative impacts of environmental variation and lead to amplified and rapid impacts at the community level, such as changes in structure, shifts in species composition or even forest collapse (Brook and others 2013; Reyer and others 2015). In addition, forest response and resilience to climatic perturbations (Lloret and others 2011) are modulated by forest structure (Bottero and others 2017; Jump and others 2017; Navarro-Cerrillo and others 2019) and composition (De Keersmaecker and others 2015; Greenwood and others 2017; Gazol and others 2018; Li and others 2021; Liu and others 2021b). Understanding the mechanisms driving forest responses to climatic extremes requires analyses of processes operating at multiple spatio-temporal scales and levels of biological organization.

At the individual level, it is generally assumed that secondary radial growth is positively related to health status (Dobbertin 2005; Camarero and others 2015b). Tree-ring width data obtained through dendrochronological techniques have been widely used to describe the resilience of radial growth in response to climatic impacts (Drobyshev and others 2021; Marqués and others 2021). Yet their capability to characterize the impacts of drought on population and community dieback processes—for example, including plant to plant interactions across several spatial scales—remains more limited, largely due to lack of data. At the stand and landscape scales, remote sensing provides temporal series of continuous observations of forest vitality (Rogers and others 2018; Buras and others 2021; Moreno-Fernández and others 2021). These observations trace aggregated vegetation responses to the environment conditions through spectral indices associated with functional processes such as photosynthesis rates (that is, greenness indices) or water content (that is, wetness indices) (Tucker 1979; Gao 1996). Spectral indices, however, aggregate multiple forest components and processes because they include canopy dominant and understory layers (Ahl and others 2006; Ryu and others 2014).

The impact, severity and magnitude of the biotic and abiotic perturbations on temporal series of tree growth and spectral indices can be quantified through the forest resilience (Lloret and others 2011), sensu the engineering resilience or the ability of a system to return to its pre-disturbance stage and the capacity to absorb change and disturbance while maintaining similar feedback dynamics (please see Bone and others (2016) and Nikinmaa and others (2020) for other definitions of forest resilience). The comparison of tree-ring data and remote sensing approaches allow us for continuous and retrospective monitoring processes key to drought-induced responses, including individual tree responses (Camarero and others 2015b), functional and structural rapid adjustments as photosynthesis and leaf shedding (Aragones and others 2019; Wu and others 2021) and water (Gao 1996), opening a new avenue to further understand long-term dynamics. Few studies, however, have addressed simultaneously the impact of drought episodes on resilience indices based on tree ring and remote sensing information (Gazol and others 2018).

Several studies have assessed the strength of the correlation between temporal series of spectral indices and tree growth data. Many of them have found positive relationships between tree and stand-data, although the strength of the association varies across biomes or species (Vicente-Serrano and others 2014, 2020; Zhou and others 2020; Castellaneta and others 2022). Mechanisms of resilience can be understood by examining different metrics of response, for example Gazol and others (2018) found that resilience to drought was more pronounced in tree-ring width than satellite-based indices, suggesting that trees may reallocate resources to photosynthesis and repair canopy damage in detriment of radial growth to alleviate the impact of a drought episode (Kannenberg and others 2019). Both tree-ring data and multispectral indices have been able to detect forest dieback before the externalization of the symptoms, this is early warning signals of forest dieback (Camarero and others 2015b; Rogers and others 2018; Moreno-Fernández and others 2021; Valeriano and others 2021). Several issues deserve to be further investigated in the use of multiscale approaches combining long-term tree and stand responses to understand drought-induced mortality (Allen and others 2015), including successive drought events, which can have critical impacts on long-term dynamics leading to abrupt responses at tree and stand levels (Navarro-Cerrillo and others 2018; Chuste and others 2020; San-José and others 2021; Yang and others 2021). Furthermore, responses to mean climate and climatic extremes can differ and be mediated by stand-level competition (Marqués and others 2021; Serra-Maluquer and others 2021). However, the importance of recurrent droughts and extreme events on both tree growth and spectral indices have not been previously assessed simultaneously (Castellaneta and others 2022).

Here, we retrospectively investigated multiscale tree and stand-level responses to drought in a Mediterranean stone pine (Pinus pinea L.) forest suffering a recent and pronounced dieback process. Specifically, we (i) explored the degree of coupling between tree-ring width (that is, secondary growth) and greenness and wetness spectral indices (that is, related to stand primary productivity and biomass accumulation) during drought episodes; (ii) quantified tree growth and spectral responses to climate; and (iii) investigated resilience, resistance and recovery for both tree- and stand-level processes in response to drought events. Our results provide evidence of long-term forest responses to recurrent droughts at tree and stand-level, which could be helpful to further characterize and understand complementary ecological scales.

Material and Methods

Study Area, Field Sampling and Climatic Information

The study was conducted in a P. pinea forest of 1500 ha, with a widespread dieback process since 2019. The forest is located in Central Spain (40.27526 N, 4.36569 W) at an altitude ranging from 500 to 850 m a.s.l. on acid soils (Figure 1). For this species, the site quality of the area is moderate (Aguirre and others 2022). Despite P. pinea being the main species, Quercus ilex L. (holm oak) and Juniperus oxycedrus L. (cade juniper) are also present in the midstory, and Cistus ladanifer L. and aromatic dwarf shrubs (Lamiaceae), such as Salvia rosmarinus L. and Lavandula stoechas Lam., complete the understory layer.

Figure 1
figure 1

Location of our study area in central Spain (upper right); plots with low (green), moderate (blue) and high (red) dieback intensity within the area; and drone flight picture (June 2021. Author: Emily R. Lines).

Full size image

The forest is under a continental Mediterranean climate, with a mean annual temperature of 14.2 °C and a mean annual rainfall of 568 mm. The rainfall is irregularly distributed through the year with spring (March, April and May) and autumn (September, October and November) being the wettest months. Summer months are especially warm and dry, which may result in long periods of water deficit and extreme droughts. In 2019, forest rangers reported a strong and rapid dieback process in the pine forest, which was confirmed by analyzing recent orthophotos. Likely, there were no changes in forest management or pathogens present during the period of study, so the dieback process was associated with rapidly changing climate conditions and drought stress.

We installed thirty-one 17-m-radius circular plots (that is 908 m2), equivalent to the area of a 30 m Landsat pixel, during May–June 2021 aiming at covering the widest range of dieback. Before plot establishment, we visited the area with the forest rangers and managers and delimited areas according to dieback level. Then, we visually classified the study area into damage levels using a planet image (www.planet.com, 24/03/2021) with a spatial resolution of three meters using false color to identify the vegetation status (R/G/B as IR/R/G; see Figure S1 in Supplementary Material). Finally, we randomly established the sampling plots within each level.

We positioned the plot center with an EMLID REACH RS2 GPS receiver (Ng and others 2018) capable of sub-meter accuracy using Differential Global Positioning System techniques. Within each plot, we measured the diameter at breast height (dbh) and identified the species of each adult tree (trees with dbh ≥ 7.5 cm). Canopy defoliation, defined as the percentage of needle loss in a crown as compared to a reference, fully foliated tree, was evaluated at least by two observers as a proxy of forest dieback in classes of 5% following the ICP Forest Manual (ICP Forest 2016). We calculated the mean plot defoliation weighting the tree defoliation by the dbh as a proxy of crown size (Porté and others 2000; Moreno-Fernández and others 2021). Then, we classified the damage level of each plot according to their mean plot defoliation: low (defoliation < 50%), moderate (50% ≤ defoliation < 70%) and high (defoliation ≥ 70%, see Table 1).

Table 1 Mean Values of Forest Attributes, Basal Area Increment (BAI) and Spectral Indices Including NDVI and NDWI in the Plots with Low, Moderate and High Dieback Intensity
Full size table

We calculated several forest attributes including stand basal area (m2 ha−1), tree density (trees ha−1) and mean tree dbh (Table 1). We used analysis of variance to assess if there were significant differences in these variables among diebacks levels, and we did not find significant relationships between the dieback level and mean tree diameter (p = 0.6977), stand density (p = 0.7693) and basal area (p = 0.0868), suggesting that the stand variables are not linked to the dieback process. Similarly, we evaluated the effect of potential soil water availability on the dieback phenomenon using the topographic wetness index and we neither found a significant relationship (p = 0.7853) between both variables. Additionally, we collected cores from approx. six pine trees per plot at 1.3 m (one core per tree) using a 5-mm Pressler increment borer aiming at reconstructing radial growth across a wide range of sizes and dieback status in each plot. We selected the trees to core considering a good coverage of tree sizes and health status, aiming at sampling at least a dead and healthy tree per plot.

Aiming to characterize the temporal variations in climatic conditions, we derived the 12-month December Standardized Evapotranspiration Index (SPEI; https://monitordesequia.csic.es/) at a 1 km2 resolution (Vicente-Serrano and others 2017). Low values of SPEI point to a negative water cumulative balance as a function of precipitation and temperature at different time scales (Vicente-Serrano and others 2010). Years with SPEI values lower than −1 were considered as drought years, due to its classification as moderate or extreme drought (Alam and others 2017) and the reasonable cut-off for our Mediterranean data (Figure 2A). We also calculated the mean annual temperature and the cumulative annual precipitation over 1984 to 2020 (see Figure S2 in Supplementary Material) from the eight nearest climatic stations using the meteoland package (De Cáceres and others 2018) implemented in R 4.1.1 (R Core Team 2021).

Figure 2
figure 2

The 12-month December Standardized Precipitation Evapotranspiration Index (SPEI) recorded from 1984 to 2020. We considered a drought event when SPEI < −1 (A) and the representation of the smooth functions (Eqs. 4–5) for the basal area increment (BAI, in mm2 year.−1, (B)) and the two spectral indices (NDVI (C) and NDWI (D)) over the monitoring period (January 1984–December 2020) grouped by dieback level. The shaded areas indicate the standard errors. Solid vertical lines refer to the period delimitation (1984–1995, 1995–2011, 2011–2020, see "Temporal trends of tree growth data and spectral indices" section. for further information on periods), dashed black vertical lines correspond to years with SPEI < −1 and dotted gray vertical lines correspond to years with −1 ≤SPEI < 0.

Full size image

Multispectral Image Processing

We used Google Earth Engine to access the open-access multispectral images from the Landsat archive and we derived the two following indices for the pixels that contain the center of each plot:

$$ {\text{NDVI}} = \frac{{\left( {{\text{NIR}} - {\text{RED}}} \right)}}{{\left( {{\text{NIR}} + {\text{RED}}} \right)}} $$
(1)
$$ {\text{NDWI}} = \frac{{\left( {{\text{NIR}} - {\text{SWIR1}}} \right)}}{{\left( {{\text{NIR}} + {\text{SWIR1}}} \right)}} $$
(2)

NDVI (Normalized Difference Vegetation Index) is a greenness vegetation index, that is, it is based on the estimation of absorbed photosynthetically active radiation, commonly used to assess the vegetation health status (Tucker 1979). NDWI (Normalized Difference Water Index) is a vegetation index based on the near-infrared (NIR) and short-wave infrared (SWIR) regions and relates to the vegetation water content (Gao 1996).

For the calculation of these two spectral indices, we selected monthly and cloud-free observations from January 1984 to December 2020 and applied the topographic correction and inter-sensor harmonization. First, we normalized Tier 1 Surface Reflectance images from Landsat 5 TM, Landsat 7 ETM + datasets to Landsat 8 OLI datasets from a multilinear regression approach due to differences between spectral characteristics among sensors (Roy and others 2016). Then we applied the topographic correction SCS + C proposed by Soenen and others (2005), which is based on the Sun-Canopy- Sensor correction (Gu and Gillespie 1998), to remove the effect of the terrain slope. After applying these corrections, we generated monthly composites using a medoid selection process choosing the pixel closest to the median of the corresponding pixels among images (Flood 2013; Bright and others 2019).

We used the BEAST (Bayesian Estimator of Abrupt change, Seasonal change, and Trend; (Zhao and others 2019b)) to separate the spectral indices trend signal (Trend) from seasonality, that is, intra-annual variability associated with forest phenology, and noise. BEAST is an ensemble algorithm that fits individual models, measures their relative importance, and averages these individual models throughout a Bayesian framework instead of selecting a single best model. In contrast to other algorithms that derive only linear or piecewise linear trends, this routine fits linear and nonlinear trends by applying flexible basis functions (Zhao and others 2019b), so it is suitable for ecological applications where processes usually follow nonlinear and complex patterns over time instead of linear or piecewise linear patterns. This approach has been previously used to discriminate plots according to their forest health status (Moreno-Fernández and others 2021). Deeper statistical details of BEAST are given in Zhao and others (2019b). This routine is implemented in the Rbeast package (Zhao and others 2019a) in R 4.1.1 (R Core Team 2021).

Tree-Ring Analyses

The cores were air-dried, mounted on wooden supports, and then sanded with sandpapers of successively finer grain until tree rings were visible. Samples were visually cross-dated using marker rings (Yamaguchi 1991), and tree-ring widths were measured with a 0.001 mm resolution on scanned images (Epson Expression 10000XL scanner) using the CDendro and CooRecorder software (Larsson and Larsson 2017). The visual cross-dating was further checked using the COFECHA software, which calculates shifting correlations with a mean site series (Holmes 1983). We converted the ring-width data into basal area increments (BAI, mm2 year−1) assuming a circular outline of stems (Visser 1995) as follows:

$$ {\text{BAI}}_{t} = \pi \left( {r_{t}^{2} - r_{t - 1}^{2} } \right) $$
(3)

where r is the tree radius while t is the year of tree-ring formation. We used BAI because it eliminates the radial growth variations associated with increasing circumference and keeps both the long and short-term changes in tree growth (LeBlanc and others 1992). We fitted linear or exponential functions to BAI data to remove long-term BAI trends and obtained detrended BAI series by dividing observed by fitted values using the dplR package (Bunn 2010) in R 4.1.1 (R Core Team 2021).

It is worth noting that we cored alive and dead trees, which mainly died recently (40% during 2018–2019; Figure S3). The year of death was estimated considering the last tree ring. Despite alive trees exhibiting greater BAI than dead trees over the study period, the BAI trends followed a similar pattern (Figure S4), and, therefore, we performed the statistical analyses combining both alive and dead trees.

Temporal Trends of Tree Growth and Spectral Indices

To compare the trends of tree growth data and spectral indices, we modeled the temporal pattern of tree growth (that is, BAI) and the trend component of NDVI and NDWI (Trend) as a function of damage level, time, plot and tree, using additive mixed models and the R package mgcv (Wood 2017). Additive models allow complex and nonlinear relationships between response and predictors to be described, especially in ecological studies (Faraway 2006; Wood 2006). We proposed the following models for BAI (4) and Trend (5) data:

$$ {\text{BAI}}_{{{\text{itnk}}}} = {\text{Dieback}}_{k} + f^{k} \left( {{\text{Time}}_{t} } \right) + {\text{Plot}}_{i} + {\text{Tree}}_{n} + \varepsilon_{{{\text{itnk}}}} $$
(4)
$$ {\text{Trend}}_{{{\text{imk}}}} = {\text{Dieback}}_{k} + f^{k} \left( {{\text{Time}}_{m} } \right) + {\text{Plot}}_{i} + \varepsilon_{{{\text{imk}}}} $$
(5)

where Dieback is a factor referred to the tree and plot damage level (k = low, moderate and high), \( {f}^{k}\left(\text{Time}\right) \) is a damage level smoother via thin plate regression splines (Wood 2003) where the smoothing variable is the Time in years (subfix t) for (4) and in months (subfix m) for (5). This formulation allows each dieback level to be differently shaped over the study period (Pedersen and others 2019). Plot and Tree are random effects to account for the intra-plot variability (i = 1, …, 31) and intra-tree variability (n = 1, …, 145), and \(\varepsilon \) is the error term of each plot i, time (either year t or month m) and damage level k (Zuur and others 2009). Because the data sources differed in temporal length (from 1970 to 2020 for tree growth and from 1984 to 2021 for spectral indices), the three data sets (tree growth, NDVI and NDWI) were filtered bounding the same temporal range, that is, from 1984 to 2020, to allow temporal comparisons. Temporal autocorrelation was accounted for through autoregressive structures of errors.

Relationships of Climate Condition with Tree-Ring Growth Data and Spectral Indices

We used dispersion plots and the Spearman’s correlation coefficient to investigate the correlations between plot tree growth and spectral indices. In the case of tree growth, we considered the BAI chronologies for each plot as well as the detrended chronologies of BAI. We fitted polynomial (20-year long splines) to BAI data to remove BAI trends and obtained detrended BAI series by dividing observed by fitted values using the dplR package. In the case of the spectral indices, we averaged annual and summer (June, July and August) values for both raw and Trend values of both indices.

The relationships between climate conditions (temperature and precipitation) and tree growth and spectral indices between 1984 and 2020 were quantified with linear mixed models in the R package nlme (Pinheiro and others 2020). For tree growth, BAI was transformed as log(detrended BAI + 1) to meet the assumption of normality. We included a first-order autoregressive structure of errors to account for the temporal autocorrelation (Zuur and others 2009) and the tree id or plot id as a random intercept term. Temperature and Precipitation were centered and scaled to facilitate the interpretability of the regression coefficients and the interaction terms (Schielzeth 2010). We proposed the following model formulations:

$$ {\text{log}}\left( {{\text{BAI}}_{{{\text{itn}}}} + 1} \right) = {\text{Prep}}_{t} + {\text{Temp}}_{t} + {\text{Prep}} \times {\text{Temp}}_{t} + {\text{Plot}}_{i} + {\text{Tree}}_{n} + \varepsilon_{{{\text{itn}}}} $$
(6)
$$ {\text{Trend}}_{{{\text{it}}}} = {\text{Prep}}_{t} + {\text{Temp}}_{t} + {\text{Prep}} \times {\text{Temp}}_{t} + {\text{Plot}}_{i} + \varepsilon_{{{\text{it}}}} $$
(7)

Temp and Prep are the mean annual temperature and the cumulative annual precipitation, respectively. The climatic variables and the pairwise interaction were selected following a forward stepwise procedure via the likelihood ratio test and using the maximum likelihood (Zuur and others 2009). Finally, the restricted log-likelihood was selected to fit the final model.

Vegetation Responses to Drought: Resilience Metrics

For the drought years (SPEI < −1; Figure 2A), we evaluated the vegetation responses to drought through resistance (Rt), recovery (Rc) and resilience (Rs) metrics (Lloret and others 2011):

$$ {\text{Rt}} = \frac{{{\text{Dr}}}}{{{\text{PreDr}}}} $$
(8)
$$ {\text{Rc}} = \frac{{{\text{PostDr}}}}{{{\text{Dr}}}} $$
(9)
$$ {\text{Rs}} = \frac{{{\text{PostDr}}}}{{{\text{PreDr}}}} $$
(10)

where Dr is the growth or spectral value during a given drought event and PreDr and PostDr are the mean values of the four previous and posterior years of a given drought event (Rubio-Cuadrado and others 2018b). We evaluated the effect of the Dieback and drought events on the three resilience metrics using two-way repeated-measures analyses of variance for tree growth (BAI), greenness (NDVI) and wetness (NDWI) indices, that is, including 3 resilience metrics and 3 indices.

Results

Temporal Trends of Tree Growth Data and Spectral Indices

We found a significant effect of Dieback on tree growth (p > 0.05) according to the Wald test. Tree growth was greater in areas with low dieback levels (Table 1 and Figure 2B), but there were no corresponding differences for spectral indices (Table 1 and Figure 2C–D). However, the inclusion of the smoothing term \( {f}^{k}\left(Time\right) \) significantly improved all models, indicating that the temporal patterns of the dieback varied over time for the three response variables studied (tree growth, NDVI and NDWI). We visually defined three temporal periods (1984–1995, 1995–2011 and 2011–2020, see Figure 2) to facilitate the interpretation of the results.

In the first period (1984–1995), tree growth and spectral indices responses overlapped for the curves of the three dieback categories (Figure 2B–D and see Figure S5 for the arithmetic mean trends for tree growth). NDVI had a local minimum in 1992–1993 that SPEI analysis suggests was preceded by dry years (Figure 2C), but this minimum was less pronounced for NDWI (Figure 2D). Following this, NDVI and NDWI increased until the dry years of 1994–1995.

In the second period (1995–2011), spectral indices had a higher value in moderate and high dieback than in areas with low dieback, whereas the opposite trend was found for BAI (see Figure 2B vs. Figure 2C–D). The smooth functions of spectral indices for the low dieback level reached minima in 2000–2001 and 2009–2010 for the three dieback levels. The 2009–2010 minimum occurred a couple of years lagged behind the reduction in tree growth in 2008.

In the third period (2011–2020), drought frequency was highest (years 2012, 2015, 2017 and 2019; Figure 2A). Tree growth had the steepest drop after the 2012 drought, regardless of dieback level, but this drought seems to have a negligible impact on spectral indices. However, we found different patterns between NDVI and NDWI trends (Figure 2C and D). The rates of initial increase and subsequent decrease were larger for NDVI than NDWI, and NDWI peaked earlier (2014) during this period than NDVI (2017). Low dieback plots showed higher values in the spectral indices through this period. Finally, spectral indices and tree growth showed a decrease from 2017 to 2019, in line with the recurrent droughts over this time period. After 2018–2019 smooth functions for the three indices increased in areas of low dieback and decreased in areas of high dieback.

Relationships of Climate Conditions with Tree Growth and Spectral Indices

We did not find significant relationships between tree growth chronologies and the annual and summer spectral indices (Spearman’s correlation coefficient, p > 0.05, see Figure S6). The likelihood test indicated that precipitation was positively related to tree growth and NDVI, but we did not find a significant relationship with NDWI. However, the temperature was positively and significantly related to NDWI (Table 2).

Table 2 Coefficients of the Scaled Temperature, the Scaled Precipitation and Their Interaction for the Basal Area Increment (log(BAI + 1)) and the Two Spectral Indices (Eqs. 6 and 7) for the Final Models
Full size table

Vegetation Responses to Drought: Resilience Metrics

The SPEI series revealed droughts in 1994–1995, 2005, 2009, 2012, 2015, 2017 and 2019 (Figure 2A). We found a significant effect of Drought on the three resilience metrics for tree growth and spectral indices whereas Dieback did not have a significant impact on any of the resilience components (Eqs. 9–11) according to the likelihood ratio test. Resistance was similar for both spectral indices over the seven studied droughts, and did not show a clear linear trend, whereas Resistance for tree growth displayed a decreasing trend over the study period (Figure 3A). Resilience followed similar patterns for the three variables (tree growth, NDVI, NDWI), with the highest values found in 1994–1995 and 2012, and the lowest in 2005, 2015 and 2017 (Figure 3B). Recovery displayed relatively constant values for spectral indices, while values for tree growth reached the largest Recovery value during the drought of 2019 (Figure 3C).

Figure 3
figure 3

Boxplots and linear trends of A resistance, B resilience and C recovery indices for tree growth (BAI, basal area increment in mm2), NDVI and NDWI.

Full size image

Discussion

The Interplay Between Tree and Stand Responses to 1990s Drought Underlies Current Dieback

Our results indicate that recent forest dynamics are driven by a climatic legacy. Specifically, droughts during the first period (1995 the most remarkable) triggered a divergence in trajectories of stands in our study area. Stands already displaying intense and rapid canopy dieback did not exhibit a further drop in tree growth and spectral indices during the second period (middle 1990s). Therefore, we suggest that the differential tree (tree growth) and stand responses (spectral indices) in low vs medium–high dieback were partially due to differences in 1990s droughts. The strong effect of the 1990s drought has been already reported in other Iberian forests (Pacheco and others 2018; Moreno-Fernández and others 2021). Furthermore, low dieback level showed a marked reduction in spectral indices after the 1995 drought, and a rapid increase in tree growth. The drop in spectral indices can be explained by a decrease in biomass, that is, stocking or tree density (also namely stocking or biomass release) (Ogaya and others 2015; Zhu and Liu 2015), and therefore a reduction in competition that can favor post-drought tree growth (Gómez-Aparicio and others 2011; Moreno-Fernández and others 2013).

The drop in spectral indices occurring during the 1990s in low dieback plots could also have resulted in increased recruitment, and overstory stem turnover with resulting shrubs and herbs expansion (Rubio-Cuadrado and others 2018a; Batllori and others 2020) that attenuated the drought signal as measured by spectral indices (Ahl and others 2006; Ryu and others 2014). Under increasing aridity, companion tree species of P. pinea in this forest (Q. ilex and J. oxycedrus) may exhibit competitive advantage and relatively more adaptative traits to the new light and water availability environments (Mayoral and others 2015, 2016; de-Dios-García and others 2018; Férriz and others 2021; Pardos and Calama 2022). However, overstory composition did not differ significantly among dieback levels during the field survey of 2020, so we do not expect that changes in species composition over time are driving temporal patterns of spectral signals. Additionally, the basal area of other coexisting tree species was low (Table 1).

According to orthophotos interpretation and local experts (pers. comm), the external dieback symptoms (for example, canopy dieback and defoliation) were visible from 2019. The recurrent droughts from 2009 to 2020 (seven years with negative SPEI) seem to trigger the crown defoliation (dieback) and mortality processes in this forest. In concordance with our results, several studies have highlighted that recurrent droughts govern mortality and dieback phenomena in forests growing across different bioclimatic areas (Navarro-Cerrillo and others 2018; Chuste and others 2020; San-José and others 2021; Yang and others 2021). Moreover, we observed that both tree growth and spectral indices had higher values from 2019 in currently low than moderate and high plots. This agrees with previous findings linking tree growth and spectral indices with field survey health status (Camarero and others 2015b; Sangüesa-Barreda and others 2015; Moreno-Fernández and others 2021).

In the absence of silvicultural treatments reported in our study area, the observed differences between low and moderate/high damaged plots could be due to climate and microsite conditions modulating canopy openness such as leaf shedding (Janssen and others 2021) and tree density that can lead to decreased resource availability and drought-induced responses (Brunet-Navarro and others 2016; Panayotov and others 2016; Jump and others 2017). In fact, Cavin and others (2013) observed a growth increase after a drought associated with a competition release, which could be similar to the effects of anthropogenic stand reductions via thinning (Giuggiola and others 2013; Moreno-Fernández and others 2013). However, other factors can also affect the productivity of Mediterranean pines including site conditions (Calama and others 2019), structural and functional heterogeneity (Ratcliffe and others 2016). For example, access to the water table during drought periods has been suggested as a key underlying driver of growth reduction in P. pinea (Mazza and Sarris 2021) and could underlie dieback on sandy or shallow soils. Regardless, we suggest that the biotic and abiotic differences in the 1990s between areas with low and moderate/high damage are leading to the current divergent dieback and mortality patterns.

Currently low damaged plots show higher tree growth since the middle 1990s (second period), almost 30 years before the emergence of currently visible dieback symptoms in 2019. This may be explained by variability in unaccounted predisposing factors (for example, site or stand conditions, soil features, genotype) making some stands more prone to drought-induced damage, and to lags between canopy and growth responses to drought with defoliation being preceded by cambial death (Pedersen 1998). All of this confirms that dendrochronology is a key tool to provide early warning measurements in forests sensitive to drought (Camarero and others 2015b, 2015a; Pellizzari and others 2016; Férriz and others 2021). In contrast to our results, other studies reported early warning signals from spectral indices (Anderegg and others 2019; Moreno-Fernández and others 2021), but in these studies the lower recovery capacity of damaged stands after the initial drought period seems to be behind the dieback (Liu and others 2021a).

Complementing Spectral Indices and Tree Growth can Better Understand Long-Term Vegetation Responses to Drought

We found differential responses for greenness and wetness indices (that is, NDVI and NDWI). Several authors investigated which type of indices are most suitable to detect forest dieback, concluding that moisture indices are preferable to those capturing only greenness (Marusig and others 2020; Moreno-Fernández and others 2021); here we did not find conclusive evidence regarding their robustness for drought-related dieback detection. However, the fact the NDWI peaked earlier (2014) than NDVI (2017) suggests that NDWI was more sensitive to the early signs of dieback than NDVI (Moreno-Fernández and others 2021). Conversely, Pascual and others (2022) reported that NDVI overperformed wetness indices in the evaluation of drought-induced mortality in a tropical Eucalyptus spp. forest. Similarly, Liu and others (2021b) found a faster decrease in greenness than wetness indices in response to drought in the forest-steppe ecotone, suggesting a lower chlorophyll content but stable water content to cope with drought. This variation in the performance of spectral indices among studies could reflect functionally divergent responses to seasonal drought (for example, stomatal closure, leaf shedding, etc.), and we suggest that the combined use of greenness and wetness indices is most appropriate to better understand vegetation responses to drought. Furthermore, more studies with field data are needed to disentangle what greenness and wetness indices detect.

We found low temporal correlation between spectral indices and tree growth. Previous results are inconclusive, some observing low correlations (Camarero and others 2015a; Correa-Díaz and others 2021) while other authors found coupling between tree growth data and NDVI (Vicente-Serrano and others 2016, 2020; Coulthard and others 2017; Castellaneta and others 2022). The fact that this forest is relatively open and the spectral indices also capture the signal from the understory layer (Ahl and others 2006) could be a handicap for the detection of significant relationships between remote-sensed indices and tree growth. Vicente-Serrano and others (2020), however, found stronger relationships between both data sources our study. This could be due to the fact that some species such as P. pinea often form open forests (see Moreno-Fernández and others 2020 for stand attributes for P. pinea at the national level). Despite this, spectral indices derived from satellite data have been successfully used to monitor drought-induced forests responses (Anderegg and others 2019; Moreno-Fernández and others 2021; Pascual and others 2022), and the legacy effect of drought can be less noticeable from remotely sensed data than from tree growth series (Gazol and others 2018; Kannenberg and others 2019). In the case of photosynthesis indices, this could lead to an underestimation of drought impacts associated with bias regarding soil moisture (Stocker and others 2019). In addition, tree radial growth is a low-priority sink for resource allocation, so may be very responsive to drought (Camarero and others 2015b), as trees may rapidly reallocate resources to photosynthesis and to repair canopy damage at the expense of radial growth (Kannenberg and others 2019).

Water Stress is a Key Constraint on Tree Growth and Spectral Responses

Our findings indicate that precipitation is the main climatic variable correlated with P. pinea growth and spectral responses, especially for tree growth and NDVI, whereas the main constraint for NDWI was temperature. Similar to our results, several studies found that water stress and precipitation have a larger impact on P. pinea radial growth than temperature (Mazza and Manetti 2013; Calama and others 2019; Mazza and Sarris 2021). Shestakova and others (2021) found that radial growth of this pine is positively related to the precipitation from previous October to current May and to the temperature from previous December to current February. These authors also found a negative association between growth and May to July temperature indicating drought was magnified by higher temperature and evapotranspiration rates, and the effect of temperature on NDWI responses can be interpreted through the magnification of water demand. Overall, the heat tolerance of P. pinea has been suggested to be a breeding trait (Loewe Muñoz and others 2015) although, according to the above-mentioned results, the radial growth of this species could be more strongly affected by extreme drought periods.

Recurrent Droughts are Triggering the Drought-Induced Mortality

The fact that spectral resilience remained more constant over the study period than growth resilience suggests that tree growth resilience is more sensitive to drought than spectral resilience, which is in concordance with previous results (Gazol and others 2018). The greater sensitivity of tree growth than spectral indices to drought could lead to an underestimation of drought impacts if spectral indices are used alone (Stocker and others 2019). Additionally, the differential processes captured across spatial scales, that is, from the individual tree for tree growth to integrated information at the pixel scale for spectral indices, may be governing divergent interpretations of vegetation response to droughts. Tree growth measures may highlight individuals’ strong responses to stress, whereas spectral information integrate vegetation responses from different species and vegetation types.

Our results indicate that during the monitoring period, tree growth resilience to drought events was relatively constant through time, whereas tree resistance strongly decreased and tree recovery increased with marked differences among drought years, depending on resilience and resistance trends. Based on resilience trends, the 2005 and 2015 droughts produced the lowest tree growth resilience. The 2005 drought is recognized to be one of the major drought episodes in the Iberian Peninsula and, resulted in widespread mortality events and a reduction in forest productivity (Galiano and others 2010; Sangüesa-Barreda and others 2015; Rodriguez-Vallejo and Navarro-Cerrillo 2019). The 2015 drought was accompanied by heatwaves (Orth and others 2016) affecting strongly tree growth in central European forests (Dietrich and others 2019) and here we show its strong effect on the tree growth resilience in Mediterranean pine forests. Based on resistance, the lowest value of resistance that occurred during the 2019 drought event could be explaining the rapid dieback process observed in the study area. Tree growth decreased in resistance over time thus suggesting that recurrent drought events could have a cumulative impact on forest health, causing crown defoliation and mortality in the long-term (Serra-Maluquer and others 2021). This also suggests that resistance is the metric most closely related and useful metric for identifying the dieback and mortality processes. In contrast, DeSoto and others (2020) argued that the drought-related mortality risk in gymnosperms is linked to low recovery capacity. Moreover, the recovery and resilience metrics for the 2019 drought should be analyzed with caution, as the post-drought period was calculated with values of one year instead of four.

Conclusions

Climate change, and, specifically, increased aridity (bringing rising temperatures and more frequent and intense dry spells) underlies the observed dieback processes in a drought-prone Mediterranean P. pinea forest. We compared two different organization scales—tree growth at the individual level and remote sensing at the stand level—across dieback levels.

The differential response among dieback levels could be due to a hypothesized biomass reduction (stocking) during the middle 1990s in low damaged plots evaluated through spectral indices that could have resulted in a tree growth release. In contrast, currently damaged stands did not experience a biomass reduction and, therefore, they did not show a subsequent tree growth enhancement from competitive release. The 1990s droughts initiated a divergence in trajectories of stands in our study area, with forest succession modulating the interplay of cumulative drought impacts and neighborhood competition. Our study highlights the importance of long-term retrospective multiscale analyses to disentangle mechanisms triggering and driving forest dieback. Understanding these mechanisms and monitoring post-dieback forests is critical for reducing vulnerability, forecasting changes in forest composition and structure and increasing resilience by targeting anticipatory adaptation measures.