The genus Cedrus may have once been present in coastal mountains of the former Tethys Sea from north-western Africa and southern Europe to western Asia, including the rising Himalayas. The Pleistocene ice ages and emerging deserts isolated the western, central, and eastern populations, leading to speciation and caused the extinction of the genus in Europe. Currently, this genus includes three extant species native from the Mediterranean mountains and one species from the Himalaya: Cedrus atlantica Manetti, from Algeria and Morocco; Cedrus libani Rich., from Asia Minor; Cedrus brevifolia (Hooker fil.) Henry, from Cyprus; and Cedrus deodara Don, from Afghanistan and southern slopes of western Himalaya. The Atlas cedar forests (Figure 1) cover an area of over 130,000 ha [36] distributed in Morocco (Rif, Middle Atlas, and north-eastern High Atlas) and Algeria (Aurès, Belezma, Hodna, Djbel Babor, Djurdjura, Blida and Ouarsenis). Atlas cedar occurs at elevations of 1,300 to 2,600 m a.s.l., where the amount of annual rainfall ranges from 500 to 2,000 mm and the minimum temperature of the coldest month ranges between −1 and −8 °C [35,46]. The Middle Atlas (northern Morocco) contains about 80% of the Atlas cedar forest surface area (ca. 100,000 ha). Middle Atlas cedar forests contain several evergreen (holm oak, Quercus rotundifolia Lam.; prickly juniper, Juniperus oxycedrus L.; European holly, Ilex aquifolium L.) and deciduous (Acer opalus Mill., Crataegus laciniata Ucria) tree and shrub species.

The most abundant tree species in the stands studied was Q. rotundifolia, mostly regenerated as sprouts which were distinguishable from seedlings because they lacked scars of cotyledonary leaves. In this region, the Atlas cedar mortality appears to be taking place mainly in old trees on dry sites such as steep slopes with southern exposure and rocky soils [37].

The study was conducted during 2009 and 2010 in Atlas cedar stands located at the Ifrane-Azrou Atlas cedar forests (Ifrane National Park, northern Morocco; Figure 1). For the description and quantification of the stand structure and spatial patterns of Atlas cedar, four study plots with contrasting structure and composition were selected based on an extensive field survey (area delimited by the coordinates latitude 33°24′40″−33°24′60″N; longitude 5°03′00″−5°08′40″W; elevation range 1,830−1,890 m a.s.l.). We sampled two plots in old growth stands, apparently subjected to intense logging and grazing (thereafter plots M. ATLAS 1 and 2) and two plots composed by an agradative closed canopy, apparently subjected to lower perturbation regimes (thereafter plots M. ATLAS 3 and 4). We avoided sampling stands with very high Atlas cedar mortality due to difficulties in obtaining and cross-dating wood cores from dead trees. Within each plot (ca. 0.13 ha in size; located about 1 km apart from each other), all trees with a diameter greater than 3 cm at 1.3 m from the base (diameter at breast height; dbh) were identified, tagged, and mapped, and their diameter was measured. All stands were located on limestone substrates.

The tree-to-tree competition intensity (CI) of each tree cored for dendrochronological analyses (focal tree thereafter) was estimated by calculating a distance-dependent competition index that takes into account the number, size, and distance of adjacent trees [47]. The degree of competition experienced by the focal tree (i) was calculated as the sum over all j neighboring trees within a radius R of the quotients between the basal-area ratio BA_j/BA_i and the distance between the focal tree i and the corresponding neighboring tree j (dist_ij). The threshold radius above which neighbors were regarded as not competing was 8 m (see [26]).

We used local climatic data (annual and monthly mean temperature and total precipitation) from the Ifrane meteorological station located 14 km from the studied plots (33°32′N, 5°07′W, 1,630 m, period 1958–2003) in order to quantify temperature and precipitation trends over the second half of the 20th century. We also used these data to identify extreme drought events, defined as years with total precipitation at least two standard deviations below the mean [9]. For the estimation of long-term trends, the raw climatic data were smoothed by polynomial regression and weights computed from the Gaussian density function (LOESS). Trends were then estimated over the smoothed series by least-squares linear regression, as well as the average slope per decade [9,48].

To quantify the growth patterns of Atlas cedar trees, we randomly selected five to ten dominant and co-dominant trees per plot (a total of 32 trees were sampled; 10, 5, 9, 8 trees for plots 1, 2, 3 and 4, respectively) and took two to three cores at breast height (1.3 m) on the cross-slope sides of the trunk using a Pressler increment borer. Most cedars were cored with a borer 40 cm long, but a 60 cm length borer was used for large trees. All cored trees were examined for symptoms corresponding to crown defoliation (expressed as percentage of total crown), presence of processionary caterpillar, bark stripping by Barbary Macaques, and the presence of bark beetles. Sampling included only living trees, both healthy (defoliation lower than 50%) and declining trees (defoliation greater than 50%, abundant dead branches, abundant brown and red needles), throughout the size (dbh) range of dominant and co-dominant trees in each plot.

All cores were sanded until tree rings were clearly visible under a binocular microscope and samples were visually cross-dated. Tree-ring widths were measured to the nearest 0.001 mm using a LINTAB measuring device (F. Rinntech, Heidelberg, Germany), and cross-dating quality was checked using COFECHA [49]. The trend due to the geometrical constraint of adding a volume of wood to a stem of increasing radius was corrected by converting tree-ring widths into basal area increments (BAI).

When a wood core did not reach the pith, we calculated basal area increment values as follows: first, we calculated the tree stem radius (Rad) from its perimeter value (Per) measured at 1.3 m from the base, as Rad = Per/2π; then, we obtained the radius inside bark (RadIB) as the difference between Rad and the bark thickness measured in the field. The difference between RadIB and the core length (CL) gives the radius of the stem just prior to the formation of the first tree-ring captured by the borer. Thus, the stem basal area at that time (BA0) can be estimated as: (2) BA 0 = π × ( RadIB − CL ) 2

The stem basal area by the end of the following year (the year of formation of the oldest tree-ring in the core) can be calculated as: (3) BA 1 = π × ( RadIB − CL + TRW 1 ) 2

where TRW1 is the measured width of the oldest ring in the core. Therefore, the first basal area increment value in our time series is:

BAI1= BA1 − BA0 = π × (RadIB − CL + TRW1)² − π × (RadIB − CL)²; the second one is:

BAI trends were calculated for different time periods by least-squares linear regression [9]. Drought sensitivity was defined as the BAI variance explained by the Palmer Drought Severity Index (PDSI). We used the PDSI reconstructed by Esper et al. (2007) for the same study area [50]. We expected a high correlation between the PDSI and BAI data, since PDSI was reconstructed using as a proxy Atlas cedar tree-ring widths. However, Esper et al. (2007) focused on old growth cedar trees and used standardized chronologies [50], while the present study sampled dominant and co-dominant trees and used raw BAI data. Based on this difference, it would be expected to obtain contrasting correlations. Therefore, we assumed that those periods when the correlations between PDSI and BAI are low may correspond to growth suppressions or releases induced by local non-climatic perturbations, such as human logging and pruning, and/or insect outbreaks. Correlations analyses were performed for the following time spans: the entire 20th century, the second half of the 20th century, the period 1970–2009, and the period 1990–2009.

To disentangle the relative contribution of tree features and stand structure to the recent Atlas cedar growth trends and decline symptoms (defoliation) we conducted a partial redundancy analysis (pRDA; see [48]. The mean 1990 to 2009 BAI of each cored tree and the canopy defoliation (%) were included in the RDA as response variables. The explanatory variables were divided into two groups: (1) tree features, which included the tree age at coring height; tree dbh; tree basal area; and mean BAI for the 1900–1990 span, as an indicator of long-term tree growth rate; (2) stand structure-related variables, which included the competition index; the surrounding basal area of neighboring (within a 8 m radius) living cedar trees; the surrounding basal area of living holm oak (Q. rotundifolia) trees; and the surrounding basal area of cut stumps; the latter was used as an indicator of human logging intensity.

Variance partitioning was performed to determine the total variance explained, the pure tree features effects, the pure stand structure effects, the shared effects of both groups and the amount of residual variation. We then followed a sequence of steps for variance partitioning in RDA: first, we applied an RDA on the response variables against tree features and stand structure together; second, we applied an RDA on the response variables against tree features alone; third, we applied an RDA on the response variables against stand structure alone; fourth, we applied a partial RDA on the response variables against tree features, using stand structure as covariates; and fifth, we applied a partial RDA on the response variables against stand structure, using tree features as covariates. Using the total sum of canonical eigenvalues of each RDA analysis (which is similar to R² in regression), the pure tree and stand structure effects, the shared information and the residual variation can all be explained as a percentage of the total inertia (variation).

Finally, to test which of the single explanatory variables was the most important, we performed a forward selection in the partial RDAs. For these, the sum of all canonical eigenvalues was used as a tool to assess the marginal effects (i.e., how well does a specific selection of explanatory variables explains the variance) for mean 1990–2009 BAI and the defoliation percentage of each cored tree. RDA was implemented using the Brodgar v2.4.1 statistical package (Highland Statistics Ltd., Newburgh, UK).

Linear regression analysis performed over the smoothed series for total annual precipitation showed a significant negative trend (−364 mm on average for the 1958–2003 span; Figure 2), which represents an approximate 28% reduction in the total annual precipitation. The greatest precipitation decrease was found in early spring (March and April, data not shown). The strength of the extreme drought events worsened towards the last decades of the 20th century (Figure 2). Drought events registered in 1993, 1995 and 1999 represent the lowest annual precipitation recorded for the second half of the 20th century. Linear regression analyses performed over the smoothed series for mean annual temperature showed a total increase of +1.6 °C, in which the highest temperature increases were observed from March to June (data not shown).

The agradative stands (plots M. ATLAS 3 and 4) were dominated by living Atlas cedar trees with mean dbh between 10 and 20 cm and scattered large trees at very low density. Living holm oak and prickly juniper (J. oxycedrus) showed higher values of the basal area in the plot selected in old-growth open stand (plot 2), which showed smaller living cedar basal area (Figure 3; Table 1) and presented cut cedar stumps of large diameter, although this plot had the lowest stump density. Deciduous tree species (A. opalus, C. laciniata) had an overall low density and were small trees with diameters of less than 10 cm thick. Dead Atlas cedars were mainly trees of dbh between 6 and 10 cm, and they were more abundant in the stands with high cedar density (plots 3 and 4) than elsewhere.

Within the cored trees the higher values of competition index were found in the stands with higher cedar density (about 1,100 trees per hectare) and lower mean age at coring height (Table 2). These stands composed by a closed cedar trees canopy had the lower density of holm oak. In contrast, the open stand (plot M. ATLAS 2), composed of isolated old trees and large cut stumps of Atlas cedar, showed the higher density of holm oak (Table 1).

Annual BAI data for the study area showed both growth releases (1990, 1996 and 2003) and years characterized by sudden growth reductions (1985, 1994 and 1999) superimposed over a long-term declining trend from the onset of the 1980s onwards (Figure 4). Correlations between BAI and PDSI over a 20 year span was significant for two time periods located around the middle 1940s and from the beginning of the 1980s onwards. The studied trees showed increasing drought sensitivity throughout the 20th century, having the highest correlations between the PDSI and the BAI from 1990 onwards (Figure 5). However, plot M. ATLAS 2 deviated from this pattern of increasing drought sensitivity over time, since there from 1990 onwards the correlation between PDSI and BAI decreased.

Results of RDA and partial RDA analyses for tree features and stand structure as explanatory variables of 1990–2009 mean BAI and tree defoliation appear in Table 3. The pure tree-features effect was equal to 33.1% of the variation, while the variance accounted for by stand structure-related variables was 5.8% (Table 4). The shared amount of variation was 19.4%, indicating that tree variables strongly covary with stand-structure variables. Variance decomposition indicated that tree age and tree mean growth had a greater effect on BAI and defoliation than stand structure.

The mean 20th-century BAI accounted for 14.7% of the variance and was positively related to the mean 1990–2009 BAI and negatively related to the percentage of defoliation of the studied trees. Tree age accounted for 13.5% of the variance and was negatively related to the mean 1990–2009 BAI and positively related to defoliation (Table 5). When the effects of tree variables were removed (see the variance partitioning procedure at material and methods section), the surrounding living holm oak trees basal area was negatively correlated to recent Atlas cedar BAI, providing a marginally significant effect (Table 5). In contrast, cedar and stumps neighborhoods and competition index were not significant. Both tree age and mean 20th century BAI showed a highly significant effect when they were included in the model, which indicates that old trees (independently of the mean growth rate) and trees with low mean growth rate (independent of the age) had the maximum likelihood of decline.

Our results support the contention that rising spring-to-summer temperatures and decreasing rainfall from winter to early summer will magnify Atlas cedar drought stress and growth decline [39,50-52]. Dendroclimatic reconstructions of drought in North Africa confirm that during the early 2000s this area experienced the most severe droughts since at least the middle of the 15th century [53]. Such retrospective assessment based on tree-ring Atlas cedar chronologies is also consistent with climate data and model projections which respectively indicate and forecast a trend of increasing aridity in the Mediterranean Basin for the past and current centuries [54-56].

Significant radial-growth reductions in response to severe droughts have been reported in other Mediterranean mountain conifers. Climate warming and increasing water stress appear to be the main causes of growth decline in Pinus sylvestris [57,58], Pinus nigra [9], and Abies alba [59,60] near the southern limit of the species distribution range, and have also caused a severe BAI reduction in the relict fir (A. pinsapo) [61]. Climatic stress influences tree physiological status and may determine whether a tree can withstand, for instance, severe defoliations or bark stripping by Barbary Macaques. In turn, climate change appears to exert a significant effect on herbivore behavior [42,62]. Recently, cedar bark stripping by Barbary Macaques appears to have increased the incidence of decline in some Atlas cedar forests [62]. The Barbary Macaques strip the bark of cedar branches and tree apices in search of water and sugars, particularly as water resources are limited and protected from monkeys [42]. In the studied stands, we detected no significant bark stripping, although previous researchers have suggested that water scarcity and monkey exclusion from previously available permanent water sources are linked to intense cedar bark-stripping behavior by Barbary Macaques [42,62,63]. This fact increases the complexity of the multifactorial Atlas cedar decline, since Barbary Macaques are a protected species and a major faunal component of Atlas cedar forests and a valuable tourist attraction [64].

In the Mediterranean Basin, human demographic pressure and livestock grazing have widely disturbed the main forest ecosystems which show a continual regression in terms of surface area in the most xeric areas [41]. The Atlas cedar forests of the Middle Atlas in Morocco have been severely affected by drought, by overgrazing involving mixed herds of goat and sheep, as well as by excessive logging for timber, firewood, and livestock fodder [64]. Human disturbances have strongly altered Atlas cedar forest structure and holm oak stand densities, which were formerly coppiced through frequent cutting cycles, and the resulting open areas serve for grazing (see holm oak and stump basal area in Table 1). Ongoing decline of open Atlas cedar stands in intensively logged and overgrazed areas are leading to their replacement by holm oak and prickly juniper. High-frequency cutting and worsening drought could be encouraging the extension of high-density holm oak coppices and decline of the Atlas cedar overstory (Figure 3). Further exclusion of cedar regeneration by more competitive oak seedlings and the replacement of this conifer by this and other angiosperm species with a more efficient hydraulic system might accelerate the postulated vegetation shift [12].

Atlas cedar decline symptoms have attracted attention mainly in 2001 but records go back several decades to the 1940s when a severe drought caused death of the cedars presumably with a comparable dieback extent [37,40,64]. In our study area, an estimated 23% of the Atlas forest surface area (3,400 of 15,000 hectares) showed some decline symptoms after the 1999–2000 drought events [64]. Many factors may be involved in this decline episode such as drought and high-temperature extremes, silvicultural practices, overgrazing, removal of lower branches for winter fodder, and pests and disease [37]. In the Moroccan Middle Atlas, the cedar decline seems to occur mainly in old trees. However, the Atlas cedar mortality pattern appeared highly heterogeneous, showing both clumped and scattered spatial patterns, and trees located near declining individuals frequently exhibited no symptoms.

The decline of the Atlas cedar forests in the Middle Atlas could affect both carbon and water cycles at a regional scale. Successive replacement of cedars and deciduous drought-vulnerable tree species by better drought-adapted evergreen species could cause a reduction in net primary production, because the Atlas cedar is the species that reaches the greatest biomass in the Middle Atlas forests [37]. Decline and mortality imply a net input of organic carbon to the soil, which could enhance the soil heterotrophic respiration and the decomposition of organic matter, thus increasing carbon emissions [65]. Total evapotranspiration may be higher in heavily logged and overgrazed open stands than in the closed cedar forest during the summer months, as a result of higher soil and air temperatures, and greater vapor pressure deficits in the open canopies [24,66]. Available soil moisture should in addition become limited by the erosion losses of the upper organic soil layer, after the taller trees were felled for timber, or killed by severe drought events [12].

We hypothesized that heavily logged and overgrazed stands would show stronger Atlas cedar growth decline in response to climate stress. However, stand structure alone does not explain the differences in BAI decline with increasing climate stress. Our results suggest that Atlas cedar growth decline is induced mainly by recent climate change [67,68], with tree age and prior growth rate playing the role of a predisposing decline factor. However, human logging seems to promote the replacement of the closed cedar forest by open mixed stands dominated by cedar and holm oak.