Modeling post-fire mortality in pure and mixed forest stands in Portugal—A forest planning-oriented model

sustainability Article Modeling Post-Fire Mortality in Pure and Mixed Forest Stands in Portugal—A Forest Planning-Oriented Model Brigite Botequim 1,*, Manuel Arias-Rodil 2, Jordi Garcia-Gonzalo 1,3, Andreia Silva 1, Susete Marques 1, JoséG. Borges 1, Maria Manuela Oliveira 4and Margarida Tomé1 1Forest Research Center, School of Agriculture, University of Lisbon, Tapada da Ajuda, 1349-017 Lisboa, Portugal; [email protected] (J.G.-G.); [email protected] (A.S.); smar[email protected] (S.M.); [email protected] (J.G.B.); [email protected] (M.T.) 2Unidad de Gestión Forestal Sostenible (UXFS), Departamento de Ingeniería Agroforestal, Escuela Politécnica Superior Universidad de Santiago de Compostela, Campus Universitario, 27002 Lugo, Spain; [email protected] 3Forest Sciences Centre of Catalonia (CEMFOR-CTFC), Ctra de St. Llorenç de Morunys Km 2, E25280 Solsona, Spain 4Research Centre in Mathematics and Applications, Colégio Luís Verney, University of Évora, Rua Romão Ramanho, 59, 7000-671 Évora, Portugal; [email protected] *Correspondence: [email protected]; Tel.: +351-2-1365-3343 Academic Editor: Marc A. Rosen Received: 31 December 2016; Accepted: 1 March 2017; Published: 7 March 2017 Abstract: Assessing impacts of management strategies may allow designing more resistant forests to wildfires. Planning-oriented models to predict the effect of stand structure and forest composition on mortality for supporting fire-smart management decisions, and allowing its inclusion in forest management optimization systems were developed. Post-fire mortality was modeled as a function of measurable forest inventory data and projections over time in 165 pure and 76 mixed forest stands in Portugal, collected by the 5th National Forest Inventory plots (NFI) plus other sample plots from ForFireS project, intercepted within 2006–2008 wildfire perimeters’ data. Presence and tree survival were obtained by examining 2450 trees from 16 species one year after the wildfire occurrence. A set of logistic regression models were developed under a three-stage modeling system: firstly multiple fixed-effects at stand-level that comprises a sub-model to predict mortality from wildfire; and another for the proportion of dead trees on stands killed by fire. At tree-level due to the nested structure of the data analyzed (trees within stands), a mixed-effect model was developed to estimate mortality among trees in a fire event. The results imply that the variation of tree mortality decreases when tree diameter at breast height increases. Moreover, the relative mortality increases with stand density, higher altitude and steeper slopes. In the same conditions, conifers are more prone to die than eucalyptus and broadleaves. Pure stands of broadleaves exhibit noticeably higher fire resistance than mixed stands of broadleaves and others species composition. Keywords: pre and post-fire management decision-making; post-fire mortality; stand structure; forest heterogeneity; fire-adapted silviculture 1. Introduction Forests cover more than one third of Portugal (3.2 million of hectares), ranking eighth in Europe as highest country with forestlands, where the forest sector represents 3% of gross domestic product (GDP) and 12% of exports, and they are a key element in the Portuguese landscape pattern [ 1 ]. Nonetheless, Portugal has the highest incidence of wildfire events in the Mediterranean basin [2]. Sustainability 2017,9, 390; doi:10.3390/su9030390 www.mdpi.com/journal/sustainability Sustainability 2017,9, 390 2 of 23 Fire hazard and spread depend on both the canopy fuels and the understory fuel structure (e.g., [3,4]). In terms of the most affected areas nationwide, statistics (1996–2012) indicate that eucalyptus (Eucalyptus globulus Labill) and maritime pine (Pinus pinaster Ait) stands are flammable forest cover types that dominate northern and central Portugal, accounting 35.9% and 41.3% of forest burned, respectively. In fact, industrial plantations are highly combustible when compared to evergreen oak woodland called “montados” (i.e., 5.2% cork oak—Quercus suber L., and 2.9% holm oak (Quercus rotundifolia L .) that predominate in the south-west and souteast of Portugal, respectively) [ 5 ]. Pine forest cover results in high fire severity, whereas the mixed forest corresponds to more desirable pre-fire stand conditions [ 6 ]. In Portuguese mainland, with the monospecific plantations being problematic in terms of risk of fires and pathogenic problems, the introduction of species diversity may be important for the forest sustainability [ 7 – 9 ]. In enhancing the heterogeneity of forest, burn severity may be substantially reduced [ 10 ]. However, mixed forests in Portugal are a small part of the forest area (5.13 × 10 3 ha, 16% of the total forest area). In fact, limited effort has been devoted to promote mixed-species stands in Portugal as compared with the effort devoted to quantifying productivity on single-species stands [11]. This information could be an important tool for forest managers’ options. Understanding and predicting forest growth and yield requires information on the drivers of tree mortality. Catastrophic disturbances have been estimated either by using fire behavior simulators (e.g., [ 12 , 13 ]) or by using fire-severity descriptors that are based on measurements of tree tissue damage, which use two main categories of observable indicators to assess tree mortality: crown and bole injury (e.g., [ 14 – 16 ]). However, the above approaches require data (e.g., tree tissue damage, fire intensity or specific meteorological conditions at the time of the fire event) that seldom are available for forest managers beforehand, and thus restrict its applicability for predicting the long-term consequences of forest management alternatives [ 17 , 18 ]. In the other hand, many studies demonstrate that variables “controllable” by the manager (e.g., mean diameter and stand density) are related with fire damage (e.g., [ 19 – 21 ]). Stand structure is related to fire intensity [ 22 ], fire severity [ 23 ] and with damage/mortality [ 18 , 24 – 26 ]. Furthermore, stand-level prescriptions provide the biological framework for managing the stands in conditions of lower fire hazard [24,27–29]. Most of the post-fire mortality models developed in Portugal have addressed fire-effects on pure stands of maritime pine (e.g., [ 30 – 32 ]) and cork oak [ 33 ]. Information on mortality prediction in uneven-aged stands is comparatively scarce. However, individual tree mortality has been predicted both from fire injury indicators and tree characteristics [ 34 , 35 ]. A set of post-fire mortality models within a forest planning context has been presented by [ 18 , 26 ] in pure eucalyptus and maritime pine stands, respectively, as a function of variables that may be “controlled” by forest managers, and none of these apply to mixed-forests. Nevertheless, the introduction of ecosystems with mixed-species stands might contribute to important goals of sustainable forest management, such as higher biological diversity, more resistance and resilience to disturbances, and restoration of Portuguese forests [ 7 , 36 ]. Promoting mixed forest has also been identified as an adaptation strategy in forest management to cope with climate change [ 37 ]. Therefore, understanding the way trees respond to fire is crucial for understanding forest dynamics in Mediterranean ecosystems [38]. In this research, a three-stage modeling technique was used to predict mortality in pure and mixed forest structure in Portugal. The inclusion of these steps will provide adequate predictions to properly examine mortality among plots, quantify and distribute among trees [ 39 ]. Specifically, a set of stand and tree models were developed: (1) to predict whether mortality occurs in a specific stand if a wildfire takes place; (2) to quantify the proportion of dead trees in stands where mortality occurs; and (3) to estimate the probability of a tree to die if fire occurs. One hypothesis is that stand mortality is affected by stand structure and stand composition, as well as by site conditions. Considering the tree mortality rate, our hypothesis is that it is affected by the effects of species, individual size and competition factors. For this purpose, the modeling was restricted to include only explanatory variables that are directly available or easily measurable by practical inventories (i.e., representing forest structure and topographic characteristics) and easily projected over the planning horizon. This enables some control Sustainability 2017,9, 390 3 of 23 in mortality. Indirect data that may be related to fire behavior (e.g., slope) were also included in the analysis. In addition, we evaluated which stand composition (pure or heterogeneous) has lower mortality, especially when taking into account the risk of wildfires. These models are planning-oriented by nature as they enable the manager to quantify effects of different management options on the expected wildfire damage. In addition, such models bears relevant management implications with the identification of forest cover types that are more fire damage-prone as they provide a critical insight into wildfire risk assessment, and also may help reverse the current mixed-forest trends in Portugal. Finally, these models are instrumental to address risk and uncertainty in an adaptive framework and may be applied in forest management optimization systems [40,41]. 2. Materials and Methods 2.1. Data Collection 2.1.1. Wildfire Perimeters and Inventory Plots Status Three major forest species cover Portugal forestlands: eucalyptus, cork oak, maritime pine, encompassing 25.8% (8.12 × 10 3 ha), 23.7% (7.37 × 10 3 ha) and 23.4% (7.14 × 10 3 ha), respectively. The remaining area is occupied by holm oak (10.5%), stone pine (Pinus pinea, 6%) and other broad-leaved tree species and conifer species (17%) [ 1 ]. These forests encompass a wide variety of ecosystems ranging from intensive silviculture plantations for wood production to plantations for coastal dune protection and agroforestry systems. Portugal is the European country most affected by wildfire with a mean annual fire incidence of 3% of its forest and wildland surface area in the 2000–2011 period, corresponding to an annual average of 1.4 × 10 5 ha [ 42 ]. Therefore, the development of preventive forest management including fire hazard reduction strategies should be of primary importance. This study used data of 3436 fire events (about 125,000 hectares burned) in several sites that burned in the three fire seasons from 2006 to 2008 in Portugal, comprising “official” Portuguese wildfire polygons larger than 5 ha (Figure 1a), and the available forest inventory stand data. Burned area mapping in this period was obtained by automated classification of medium-resolution remote sensing data (i.e., Landsat Multi-Spectral Scanner (MSS), Landsat Thematic Mapper (TM) and Landsat Enhanced TM+). First, GIS tools (ArcGIS ® 9.3 software, Esri, Redlands, CA, USA) were used to identify the plots from the 5th National Forest Inventory (NFI, carried out in 2005–2006 period) where a wildfire occurred between 2006 and 2008. The wildfire perimeters and the 12,258 plots of the NFI (with 5267 plots of forest stands) were overlaid for that purpose (Figure 1b). In total, 38 NFI plots with 795 trees had been measured prior to wildfire occurrence (pre-fire inventory). In the same period, 203 additional burned plots distributed all over the country from the framework of ForFireS Project [ 43 ], and encompassing 1725 trees were further measured (post-fire inventory in 2007 and 2008) (Figure 1d). All these plots were sampled in areas where the fire perimeter was known and trees had not been harvested. In total 241 plots (165 pure and 76 mixed forest stands) representing 2520 individuals containing tree-level data for 16 species (2 oaks, 8 other broadleaves, 1 eucalyptus, and 5 conifers) from a network of plots distributed across the country with different site conditions and stand characteristics, representing the environmental variation of the forest type, were considered for the study (Figure 2). Both the individual tree evaluation in the plots and mortality due to fire (i.e., classification of whether a tree died or not), was checked only once (roughly 1 year after the wildfire occurrence). Trees were considered dead when no green foliage was present regardless of its location on the tree [ 43 , 44 ]. Status (Survival “S” or dead “D”) was recorded for each tree sampled (Table 1). The pre-fire inventories encompassed evaluating and collecting all the information available in the corresponding NFI database. The post-fire inventories encompassed the measurement of biometric variables (e.g., tree height (h), m; and diameter at the breast height 1.30 m (dbh), cm) for trees larger Sustainability 2017,9, 390 4 of 23 than 7.5 cm (5 cm in the case of eucalyptus trees), along with topographic characteristics of the plot (elevation, aspect, slope). Site slopes ranged from 12 to 32 degrees, and Southwest and Northwest orientations were prevalent in the study plots. Since the objective of the model was to predict post-fire mortality over long horizons within a planning-oriented framework, biometric variables tested for the model were restricted to easily measurable tree (Table 1) and stand characteristics (Table 2, pure stands; and Table 3, mixed stands) without considering cambium injury variables. The former variables allow the forest manager to predict the effect of changes in stand structure and species compositions on the expected mortality. Sustainability 2017, 9, 390 4 of 23 of the plot (elevation, aspect, slope). Site slopes ranged from 12 to 32 degrees, and Southwest and Northwest orientations were prevalent in the study plots. Since the objective of the model was to predict post-fire mortality over long horizons within a planning-oriented framework, biometric variables tested for the model were restricted to easily measurable tree (Table 1) and stand characteristics (Table 2, pure stands; and Table 3, mixed stands) without considering cambium injury variables. The former variables allow the forest manager to predict the effect of changes in stand structure and species compositions on the expected mortality. Figure 1. Inventory plots location for data acquisition: (a) the map displays the distribution of fire perimeters occurred in Portugal between 2006–2008 larger than 5 ha; (b) the 5th National Forest Inventory (NFI) plots covering the forest area of Portugal in a systematic 2 × 2 km grid (12,258 plots); (c) the overlaid from fire perimeters and plots of the 5th NFI; and (d) the sample plot 38 NFI (795 trees) from the post-fire inventory of plots in 2007 and 2008 and additional 203 plots from ForFireS Project (1725 trees). Figure 1. Inventory plots location for data acquisition: ( a ) the map displays the distribution of fire perimeters occurred in Portugal between 2006–2008 larger than 5 ha; ( b ) the 5th National Forest Inventory (NFI) plots covering the forest area of Portugal in a systematic 2 × 2 km grid (12,258 plots); ( c ) the overlaid from fire perimeters and plots of the 5th NFI; and ( d ) the sample plot 38 NFI (795 trees) from the post-fire inventory of plots in 2007 and 2008 and additional 203 plots from ForFireS Project (1725 trees). Sustainability 2017,9, 390 5 of 23 Table 1. Descriptive statistics for tree level data and individual tree mortality by species (Tree status: D, dead; S, Survival,) tested as model predictors in the range of study (n= 2520, 16 species), concerning the period of analyses (2006–2008). Tree Species Tree Status ndbh (cm) h(m) g(m2) Mean Range Mean Range Mean Range Ec S 177 15.24 5.20–59.30 15.13 5.40–30.40 0.022 0.002–0.276 D 762 10.06 5.00–46.30 12.24 1.40–28.20 0.008 0.002–0.168 OB S 37 10.88 7.00–21.00 6.92 4.90–11.60 0.010 0.004–0.035 D 43 10.19 7.00–23.00 5.97 3.12–9.90 0.009 0.004–0.042 OC S 11 25.09 7.00–59.00 15.09 7.98–26.90 0.073 0.004–0.273 D 5 13.40 9.00–21.00 9.79 8.22–11.44 0.015 0.006–0.035 Pp S 263 19.17 6.00–51.00 13.89 5.50–28.50 0.035 0.003–0.204 D 981 14.63 7.00–51.00 11.09 3.80–28.50 0.021 0.004–0.149 Ppi S 7 31.23 18.00–43.00 17.45 7.35–24.40 0.086 0.025–0.145 D 4 13.70 8.50–22.60 7.22 3.97–8.90 0.017 0.006–0.040 Qr S 20 16.82 8.00–39.00 5.09 3.83–7.60 0.028 0.005–0.119 D 10 19.08 7.00–64.00 4.99 3.40–7.50 0.051 0.004–0.321 Qs S 58 20.01 4.30–62.00 6.72 2.70–13.90 0.044 0.001–0.302 D 48 17.53 4.50–59.80 6.34 2.95–10.70 0.042 0.002–0.281 Qsp S 42 12.21 7.00–24.00 8.61 5.00–13.80 0.013 0.004–0.045 D 52 10.09 7.00–21.30 7.25 4.20–12.50 0.009 0.004–0.035 where dbh is the tree diameter at breast height (cm); his the total tree height (m); gis basal area of the tree (m 2 ). Ec: Eucalyptus globulus; OB: Others Broadleaves; OC: Others conifers; Pp: Pinus pinaster; Ppi: Pinus pinea; Qr: Quercus rotundifolia; Qs: Quercus suber; Qsp: Other oak trees. Range: minimum–maximum. 2.1.2. Reverse Engineering to Rebuild the Tree Characteristics In the case of the 203 additional plots that had not been measured before wildfire (post-fire inventory), to take advantage of available inventory data, reverse engineering was used to re-build the forest before the fire [ 44 , 45 ]. This was especially needed to estimate tree height before wildfire, as fire often destroys part of the crown thus complicating height measurements. In the case of plots with standing burned trees, pre-fire diameter was assumed to be unaffected by fire. Pre-fire height was estimated using equations developed by Tomé, M. et al. [ 46 ] for maritime pine, [ 47 ] for eucalyptus, and [46] for the rest of species i.e., cork oak, holm oak and stone pine. 2.2. Post-Fire Model Description The three-stage modeling system developed in this study comprises three models: the first two at stand level and the remaining one at tree level. For the two former cases, the information required for model development considers one observation per plot, and it is reasonable to assume that they are independent between each other, as the plots are located without following any specific pattern. Under this assumption, the multiple fixed-effects logistic regression seems an adequate fitting procedure. However, the data used for the development of the latter model (at tree level) presents a hierarchical structure, i.e., several trees were measured in each sample plot. This means that a within-plot correlation (between trees of the same plot) in the response variable is expected. Accordingly, we have considered multiple mixed-effects logistic regression to develop this model, as this technique allows the inclusion of random effects in some (or all) parameters to account for the expected correlation between trees of the same plot. Finally, we did not quantify the probability of a stand being impacted by fire (fire occurrence), we only assessed what happens when a stand is impacted by fire (i.e., if mortality would occur, what would be the mortality and which trees would die). Sustainability 2017,9, 390 6 of 23 Table 2. Descriptive statistics for variables tested as model predictors at stand level for pure stands. Stands with Dead Trees (n= 96) Stands without Dead Trees (n= 68) Eucalyptus Conifers Other Broadleaves Eucalyptus Conifers Other Broadleaves Variable (Code) Range Mean (S.d.) Range Mean (S.d.) Range Mean (S.d.) Range Mean (S.d.) Range Mean (S.d.) Range Mean (S.d.) Avgdbh 2.92–14.50 9.03 (3) 7–26.25 13.75 (5.81) 7.5–38.03 18.85 (9.05) 7–27.45 11.49 (3.63) 13–32.8 20.18 (6.76) 8–73.36 30.11 (17) Avgh 3.49–15.24 10.18 (3.35) 3.8–19.37 9.78 (4.08) 3.93–9.4 6.45 (1.53) 6.5–21.9 13.84 (3.68) 3.47–19.73 12.01 (4.42) 4.7–18.23 8.12 (4.23) dg 5.59–26.03 8.98 (4.17) 5.14–45.09 18.64 (11.9) 13.6–45.09 28.77 (10.14) 6.03–45.09 13.54 (8.13) 10.08–45.09 27.16 (15.78) 17.04–45.09 34.14 (9.96) G0.31–11.04 4.91 (2.57) 0.077–38.16 7.15 (9.46) 0.08–26.51 3.9 (6.54) 0.08–29.73 7.62 (7.2) 0.27–33.13 8.2 (11.01) 0.1–13.81 4.09 (3.75) G/dg 0.012–1.80 0.66 (0.42) 0.0017–3.78 0.78 (1.1) 0.002–1.95 0.23 (0.48) 0.002–3.4 0.8 (0.84) 0.006–3.2 0.71 (1.1) 0.002–0.81 0.15 (0.19) N60–1299 691.48 (347.55) 20–1539 318.7 (353.86) 20–220 72 (57.09) 20–1811 617.4 (429.62) 20–623 181.62 (213.82) 20–140 46.96 (33.91) Sd 0.78–5.37 3.28 (1.3) 0–12.41 3.63 (3.13) 0–26.02 7.31 (8.67) 0–7.91 2.88 (1.89) 0–11.72 4.3 (4.27) 0–19.8 4.58 (6.01) Sh 0.4–6.96 3.43 (1.98) 0–5.46 1.57 (1.38) 0–3.81 0.96 (1.08) 0–5.07 2.2 (1.23) 0–4.09 1.34 (1.46) 0–5.91 0.75 (1.28) Sd/dg 0.03–0.88 0.45 (0.25) 0–1.34 0.33 (0.35) 0–1.26 0.34 (0.43) 0–1.07 0.27 (0.21) 0–1.13 0.33 (0.37) 0–0.93 0.18 (0.25) Altitude 0–272 179 (81) 0–893 330.42 (171.77) 76–800 296.55 (152.55) 0–491 192 (150.53) 106–931 441 (331.3) 0–861 313.39 (224.44) Slope 0–26.6 10.27 (6.2) 0–29 13.3 (7.92) 0–22.8 8.27 (6.53) 0.6–32 13.09 (8.68) 1.8–27 11.48 (6.8) 0–25.2 9.38 (6.27) Where Altitude is measured in meters (m), Avgdbh, mean tree diameter at breast height of the stand; Avgh, the average tree height; dg, the quadratic mean diameter (cm); G, stand basal area (m 2 /ha); G/dg, non-linearly a density measure related to the number of trees per hectare, N, tree density number of trees per ha; G, stand basal area (m 2 /ha); sd, standard deviation of tree diameters; sh, standard deviation of tree heights of the trees in the stand. The predictor Sd/dg expresses the relative variability of tree diameters. The variable is close to “1” in rather uneven stands and approaches “0” in homogeneous stands (0 < Sd/dg < 1); Slope, in degrees (◦), Range: minimum–maximum. (S.d.): Standard deviation. Sustainability 2017,9, 390 7 of 23 Table 3. Descriptive statistics for variables tested as model predictors at stand level for mixed stands. Stands with Dead Trees Stands without Dead Trees (n= 56) (n= 20) Variable (Code) Range Mean Range Mean (S.d.) (S.d.) Altitude 0–940 337.78 0–919 268.15 (232.79) (285.51) Avgdbh 7.64–25.67 15.41 7.99–26.88 14.7 (4.84) (5.54) Avgh 3.72–22.7 10.99 4.92–17.06 11.15 (4.19) (3.48) dg 5.64–31.88 13.04 10.08–26.03 13.66 (5.31) (4.93) G0.55–30.52 9.29 0.93–33.18 8.81 (7.89) (8.76) G/dg 0.02–3.33 0.95 0.051–3.29 0.79 (0.94) (0.88) N40–1279 398.02 60–1769 486.3 (297.73) (433.8) PBr 0–0.98 0.27 0–0.82 0.25 (0.33) (0.29) PCon 0–0.99 0.56 0–0.95 0.4 (0.35) (0.29) PEc 0–0.97 0.16 0–0.95 0.35 (0.3) (0.41) Sd 0.7–15.4 6.57 1.42–16.96 6.54 (3.31) (4.53) Sh 0.55–6.49 3.13 0.69–6.53 3.37 (1.58) (1.82) Sd/dg 0.06–1.54 0.58 0.09–1.02 0.49 (0.37) (0.3) Altitude 0–940 337.78 0–919 268.15 (232.79) (285.51) Slope 0–32 13.53 0.6–22.6 13.7 (7.77) (6.06) PBr is the proportion of other broadleaved trees in the stand (“1” indicating that stand is purely occupied by broadleaves), PCon is the proportion of conifers and PEc is the proportion of eucalyptus in the stand. For the other parameter definitions, see Table 2. Range: minimum–maximum; (S.d.): Standard deviation. 2.2.1. Statistical Approach Fixed-Effects Logistic Regression Although most cumulative distribution functions would work in modeling mortality, the most used is the logistic (e.g., [ 48 , 49 ]) because it is mathematically flexible and has a meaningful Sustainability 2017,9, 390 8 of 23 interpretation [ 50 ]. The logistic function predicts a probability of an occurrence ranging continuously between 0 and 1. The dependent variable is dichotomous (e.g., death and survival as event/non-event, respectively). A cut-off point may be defined in order to assign “1” to the death event and a “0” to the non-death event [48]. Y=1 1+e−(β0+β1x1+...+βpxp)(1) where Yis the dependent variable, representing a measure of total contributions of all the independent variables used in the model, x 1 to x p are independent variables, β0 is the intercept, and β1 to βp are parameters to be estimated or regression coefficients. The logistic function was used to model stand-level mortality due to wildfires. The glm function of the R statistical software [ 51 ] that estimates the parameters of the logistic equation through maximum likelihood was used in the first two stages of the proposed approach to develop the post-fire stand mortality models. Different combinations of biometric variables from groups pertaining to tree size as an indicator of tree vigor (e.g., tree diameter at breast height (dbh)), competition expressed by a variety of stand density measures (e.g., basal area (G)) and easy-to-calculate competition indices (e.g., basal area of the trees larger than the given tree, BAL), and average variables (quadratic mean diameter (dg)) were tested to express the mortality rate [ 48 , 52 , 53 ]. However, indirect measurements that may be related to fire behavior (e.g., slope) were included in the analysis. Mixed-Effects Logistic Regression Mixed-effects modeling is based on the fact that some (or all) parameters are formed by a fixed part, common to the whole population, and a random effect, specific for each subject (in this case, each plot). For the special case of logistic regression, the model becomes a generalized linear mixed-effects model with a logit link function. The expression is equivalent to the Equation (1), except that in this case the vector of parameters β ( β0 , and β1 to βp ) is now expanded with a vector of random effects u : not all the parameters need to include a random effect but at least one of them should include a fixed and a random part to be considered a mixed-effects model. The vector u is assumed to follow a multivariate normal distribution with mean zero and a variance-covariance matrix to be estimated in the fitting procedure, in addition to the fixed-effects parameters and the error variance. The fitting procedure is based on a log-likelihood maximization and was performed with the glmer function of the lme4 package [ 54 ] of the R statistical software [ 51 ]. Within the mixed-effects modeling context, we have evaluated the predictions made only with fixed-effects parameters, i.e., the mean prediction. Therefore, note that mixed-effects modeling was only considered in the fitting procedure in order to account for the existing hierarchical data structure rather than for estimating random effects whose aim would be the improvement in the model prediction. Due to the nested structure of the data (trees inside plots), parameters associated with plot variables (e.g., slope, G, and N) are considered only fixed parameters because these variables are already accounting for the variability between plots. However, the parameters associated to tree-level variables (e.g., group of species, BAL, dbh or h) might have a random effect that accounts for the variability between plots. 2.2.2. Stand-Level Modeling To predict mortality in a stand after a wildfire, a plot-level binary variable indicating whether mortality did occur in the inventoried stands was created. As suggested by [ 55 ], this variable takes the value “1” if some mortality occurred within the stand, i.e., at least one tree bigger than a certain dbh died (5 cm for eucalyptus or 7.5 cm for the remaining species), and the value “0” if no death occurred in the stand. This modeling approach thus provides information to differentiate the stands where some mortality occurs from the other stands. Thus, the data used to develop this sub-model included all sample plots (e.g., plots with and without occurrence of mortality). A number of stand-level features Sustainability 2017,9, 390 9 of 23 defining site conditions, composition and biometric variables as described in Tables 2and 3were used for estimating the mortality of wildfire (PsDead). For modeling purposes and because there are tree species that tend to share similar traits, a preliminary analysis was performed to decide how to group the species present in the inventory. The stands were classified according to the proportion (0 ≤ Pcovertype ≤ 1) of each cover type in the stand, i.e., the proportion of “eucalyptus” in the stand (PEc), proportion of “conifers” (PC) (including Pinus pinaster,Pinus pinea and short-needled conifers such as Pinus sylvestris, Pseudotsuga menziesii and other conifers) and the proportion of “other broadleaves” including Mediterranean evergreen broadleaves such as cork and holm trees (PBr). Oak stands (including Quercus suber and Quercus rotundifolia) and deciduous and evergreen species (including Acacia spp., Alnus glutinosa,Betula spp., Castanea sativa,Fraxinus spp., Quercus pyrenaica,Quercus robur) were grouped within the “other broadleaves” cover type. This was done for two main reasons: (i) statistical analysis showed that oak mortality did not differ from broadleaved stands; and (ii) the relatively small number of oak stands (cork and holm) observations detected in the dataset to be included in an “oak” cover type composition. Eucalyptus stands were left out of the “other broadleaves” group because of its specific stand characteristics and the relevance as an economic resource that supports a large and significant pulp and paper industry, one of the key businesses in Portugal (Figure 2). As a direct follow-up, to quantify post-fire mortality in 153 stands where mortality did occur one plot-level variable was created, such as: proportion of events—the number of trees that died as a consequence of a wildfire within a stand (which is calculated from the number of dead trees and the total number of trees in the plot). The glm function of R used this information to apply the logistic regression. Therefore, this sub-model (PdMort) described fire damage as the proportion of trees that died in the stand per hectare as a consequence of a wildfire, i.e., number of dead trees divided by total trees in the stand (trees · ha −1 ). In our analysis, the mortality of the Mediterranean evergreen broadleaves (cork and holm) stands did not differ from other broadleaved stands. Thus, as in the previous sub-model PsDead, for modeling purposes the stands were classified according to the proportions (0 ≤ Pcovertype ≤ 1) of three cover types (“eucalyptus”, “conifers” and “other broadleaves”). A number of stand-level variables related to topography (slope, altitude, aspect), biometric variables (e.g., mean tree diameter at breast height of the stand, Avgdbh) and structure (e.g., standard deviation of tree heights, sh) were tested. 2.2.3. Tree-Level Modeling In this third stage, the predicted variable was the probability of tree death in the event of fire. Thus, only trees present in stands where mortality was predicted with the stand-level model PdMort were used to fit the tree mortality model. For modeling purposes, a tree-level binary categorical variable was created. This variable takes the value “1” if the tree dies, and “0” if the tree survives. The model was parameterized for 16 of the most Portuguese common species, using 2520 individual tree record in total inventoried in burned plots, of which 1905 were present in stands where mortality was predicted as a consequence of fire. For modeling purposes, to take advantage of available inventory data of trees per each species and individual tree characteristics, single trees were reclassified in four classes of cover types: i.e., “eucalyptus” (n Ec = 939), “other broadleaves” (n Broad = 174), “conifers” (nCon = 1271) and “oak trees” (nOak = 136). In this model we distinguish the oak trees (cork and holm oak) from the other broadleaves. This was done because previous studies focusing on the factors influencing post-fire tree survival and regeneration capacity showed that fire injury and individual tree characteristics are the main factors [ 21 , 56 ], which depends on fire-resistance or avoidance mechanisms, and in flammability of species [ 22 ]. The model PdTree includes dummy variables for the different cover types. In the case the tree is from one of the cover types the value of the dummy variable is “1”. Here a logistic regression mixed-effects model that assigns a probability of mortality to an individual tree given its size (e.g., stand height (h), basal area of the tree (g), and quadratic mean diameter of trees (dg)) (Table 1), species identity (represented by classes of cover types), competitive environment Sustainability 2017,9, 390 16 of 23 4. Discussion The results suggest that the set of models are logical, and may provide reasonable estimates of post-fire mortality over long-term planning horizons. The logistic modeling approach used in this study has been used earlier for predicting tree-mortality as a consequence of regular mortality (mortality due to competition between trees) and irregular mortality, i.e., mortality due to wind damage [ 61 , 62 ], prescribed fire [ 63 , 64 ] and wildfire [ 17 , 18 , 21 , 64 , 65 ]. In addition, the modeling approach using different steps has been used to model natural tree mortality [ 55 , 66 , 67 ]. Here we use the Portuguese Forest Inventory and Project ForFireS dataset to parameterize, for a total of 16 common Portuguese tree species, a logistic regression model that assigns post-fire mortality. The advantage of the three-stage modeling technique used compared to other approaches is the possibility of detecting stands where no mortality occurs. Otherwise, traditional models always generate some mortality for all plots [55]. Post-fire mortality has been studied using a variety of direct and indirect methods (e.g., [ 12 , 68 ]). In the literature we can find two main types of models, the ones based on variables reflecting fire injury indicators and those based on biometric information. Here, while the proposed approach employs some concepts that featured in earlier work (e.g., [ 17 , 18 ]) on the whole, the selected post-fire mortality models tend to differ significantly from other well-known classic approaches for fire-induced mortality (e.g., [ 34 , 35 , 69 ]). However, the use of the latter methods in long-term forest management planning is constrained by the difficulty to predict accurately the fire injury/severity variables they use (e.g., crown-kill, bole-kill or fire intensity are information rarely available to forest beforehand). In fire-prone ecosystems, active fuel management may be used to preserve forest stands from fire damage [ 17 , 23 , 27 ]. Indeed, damage and survival depended on tree and stand characteristics that may be changed through forest management enabling some control on fire mortality. This is the first attempt to develop post-fire mortality models for any stand structure and species composition in Portugal as a function of explanatory variables that are directly available or measurable/predicted by practical inventories and easily projected over the planning horizon. The role fire meteorology was not included in the model mainly because the weather before and/or during a specific fire event is very difficult to generate reliable meteorological data over long periods to be used in a long-term forest management planning context (e.g., 60 years planning horizons). Nevertheless, indirect variables that may be related to fire behavior such as the slope and the vertical structure of the stands were included in the modeling analysis. As stressed above, the set of three models does not use fire-related variables, but information on species composition and forest structure. In this sense, the models provide: (i) a quantitative framework to address long-term planning periods; (ii) insight into different behavior or post-fire mortality between pure and mixed stands, helping to understand the performance of any forest structure; and (iii) information for making pre-fire management decisions over a wide range of tree sizes and species and in a variety of stand conditions in Portugal, which has been lacking in previous discussions. 4.1. Forest Composition, Heterogeneity and Structure The models show that forest cover composition has an impact on the proportion of dead trees after a wildfire. The post-fire mortality models developed indicate that the conifers are the most prone to suffer mortality after wildfire, whereas the broadleaves are more resistant to fire (Figure 5). This confirms the generally low fire susceptibility observed in most broadleaved trees compared with other forest types, probably explained by differences in fuel load composition, moisture content, flammability, and broadleaves resprouting ability (e.g., [ 3 , 34 , 35 , 70 , 71 ]). Previous studies argued that the conifers have higher flammability because their contents in resins and oils [ 32 , 70 ]. Other authors attribute the difference between broadleaves and conifers to the fact that the forest management regimens applied to both of the forest tree types is different, being broadleaves more resistant to the fire occurrence (e.g., [ 36 ]). For instance, in Portugal, eucalyptus and maritime pine stands are usually Sustainability 2017,9, 390 17 of 23 industrial plantations with high stand densities, and highly prone to fire [ 22 , 72 ]. In fact, commercial eucalypt plantations are very flammable due to the nature and accumulation of their litter and bark fuels with high levels of biomass, and thus prone to intense wildfires [ 20 ]. In the case of mixed eucalyptus and pine stands, fire hazard may be even higher, when compared with pure stands of eucalyptus or pines [ 72 ]. On the other hand, many of the oak stands (i.e., Q. Suber and Q. rotundifolia) are classified as “montado”, which is an agroforestry management regime, with low tree stocking, which reduces considerably the fire occurrence and propagation [ 4 ] (proportion of dead trees, Table 7). One explanation may be the high resprouting capability of some Mediterranean Quercus species such as Q. rotundifolia,Q. suber and Q. coccifera [ 73 , 74 ], which can lead to Quercus dominated forest in places with high fire recurrence. Thus, understanding the relationships between composition of the forest cover and burn severity is important for developing management guidelines leading to fire-resilient forests. Forest heterogeneity (mixed stands) was shown to be associated with low proportion of dead trees in the stand (Figure 4). This is in accordance with results from [ 10 ] who explored the relationship between forest heterogeneity and burn severity at the stand-level concluding that pure conifer stands present more severe fires than mixed stands. Thus, forest managers may consider enhancing the heterogeneity of forests when implementing fuel treatment schemes giving more importance to mixed stands. In fact, higher mortality is expected in coniferous stands rather than in broadleaved species, because most species from the former group are not able to resprout when the entire canopy is burned [34]. In general, biometric variables that impacted post-fire mortality included tree diameter (dbh of the tree, average dbh of the stand and variability of tree diameters Sd/dg), and indicators of density such as basal area (G) and competition index (dbh/dg). The coefficients of biometric variables regarding stand structure are in concordance with findings from international studies (e.g., [ 17 , 75 ]) and studies focused on Portuguese conditions [ 18 , 26 , 34 ]. The stand-level model indicates that even-aged stands with higher tree diameters have lower probabilities of mortality after a fire than irregular stands with smaller-sized trees (model PsDead). These small and dominated trees will be more exposed to a given intensity fire than larger trees, especially for low to moderate severity fires [21]. Further, eucalyptus stands with a reduced variability of tree diameters (Sd/dg) results in lower probability of mortality than conifers stands (Figure 3). Mortality in stands with higher densities is expected to be higher (PdMort). These results agree with findings of [ 22 , 76 ] who stated that the level of injury and mortality for a given species is a combined outcome of fire behavior, tree size and stand structure. Extensive model testing led to the rejection of other biometric variables as predictors of stand-level damage after a wildfire. Regarding the individual tree mortality PdTree the coefficients of biometric variables also confirmed the findings of previous studies. For tree-level mortality prediction of forest species in Portugal, best-fitting model PdTree1 indicates that stand basal area (G) and dbh traits affected the mortality rate (basal area (G) was negatively related with tree mortality). This is in concordance with other studies [ 17 , 18 , 20 , 33 ]. In addition, size (dbh) has a greatest effect on mortality rate at the level of the individual tree, with larger trees size exhibiting mortality rates much lower than smaller trees, consistent with a decrease in mortality reported by several studies [ 18 , 26 , 71 ]. These two variables are related, so for a fixed value of basal area (G) if the stand density increases, it means that average tree diameters are smaller. This means that fire would more likely be more intense and more trees are likely to die. Indeed, fire intensity varies and this could result in different probability curves depending on species and intensity [70] The results suggest that both slope and stand location (i.e., altitude) also impact mortality. In the stand-level model (PdMort), steeper slopes increase the probability of death to occur in a stand and also the expected damage (Figure 4). This is in concordance with other studies developed in Portugal [ 2 , 5 , 20 ]. This effect may be explained as slope increases fire spread rate and consequently increases fire intensity [ 77 ]. Additionally, altitude has a significant impact on fire behavior possibly Sustainability 2017,9, 390 18 of 23 due difference in fuel moisture present in a given stand, and weather conditions, in particular, with temperature and precipitation and the compounding effects of fuels [ 78 ]. In our case, altitude correlates positively with the degree of mortality in burned areas. Thus, forest managers may want to consider avoiding severe topographic conditions (e.g., steeper slopes when planning new plantations), in order to increase resilience to fire and management of wildfire damage. Validation of the models was done through studies of the performance of the functions. No specific validation datasets were set-aside and later used for that purpose. Two main reasons justify this decision. Firstly, the relatively small number of observations in the stand dataset used. Secondly, we were more interested in obtaining the best possible parameter estimates. There are advantages and disadvantages of splitting the dataset for model validation purposes as discussed by [ 60 ]. However, they concluded that cross validation by data splitting and double cross validation may provide little, if any, additional information in the process of evaluating regression models. 4.2. Pre and Post-Fire Smart Management: Applicability This research estimated a set of three models to predict mortality that present a real step forward for managing mixed stands under fire risk conditions, and identifying management options that reduce the expected losses due to fire. This may be critical to design prescriptions that may reduce ecological and economic wildfire damage through adequate planning, as a management objective in numerical planning calculations or easy integration of post-fire mortality in forest simulators and optimization systems, particularly in ecosystems where wildfires are a recurrent disturbance. It is useful to interpret the current study in the broader context of fuel management, whereas stand management strategies are widely accepted as effective means to establish less flammable and more resilient forests and landscapes in recently burned area. In the framework of forest management planning, the stand-level models PsDead and PdMort are useful for planning when the growth and yield model used does not provide the dimensions of each of the trees. First, model PsDead may be used to predict whether mortality may occur in a stand after a wildfire (i.e., there is at least one dead tree in the stand). If the stand presents mortality, model PdMort gives the proportion of dead trees in the stand (i.e., the total number of trees that will die is provided). For this reason no threshold value is needed to convert the estimated probability into a dichotomous variable (e.g., death or survival). Because PdMort cannot be used for a specific tree, the model PdTree to predict the probability of a tree to die should be applied. Thus, the selected model at tree-level PdTree may be used to predict the probability of mortality of each tree in the stand and to build a list of all trees in the stand ordered according to this probability (which means that the trees should simply be sorted from higher to smaller mortality probability, trees with higher probability are ranked first in the list). The management planning model may then select the trees that will be assumed to die for planning purposes by going down the list and stopping when the number of trees that are estimated to die by model PdMort is reached. Another way to use these set of models in forest management scheduling and scenarios analyses is in the framework of a full risk profile, especially when first generate fire occurrence with an earlier model for pure and mixed forest stands [ 44 , 79 , 80 ], after which the degree of damage can be predicted with the second model PdMort. Thus, the quality of the models is dependent on the quality of the equations used for that purpose. In a management planning context, the PdTree post-fire mortality model is very important when the growth and yield simulation uses an individual tree model (which means that every tree may have different characteristics). 5. Conclusions This research encompassed the development of post-fire stand and tree mortality management-oriented models for improved pure and mixed forest planning in Portugal. Given the general uncertainty related to large-scale forestry scenario analyses, and the uncertainty related to mortality as a phenomenon, the selected mortality models were considered to have an appropriate Sustainability 2017,9, 390 19 of 23 level of reliability. Indeed, the set of models perform similarly to other planning-oriented models developed for pure stands in Portugal. The predicted proportion of dead trees in pure maritime pine stands using the stand-level mortality model PdMort presents a very similar response pattern than the model developed by Garcia-Gonzalo et al. [ 18 ]. If model PdMort is used for pure eucalyptus stands, it also gives similar results as the model developed by Marques et al. [26]. The advantage of the current post-fire models system is that they are suitable over a wide range of tree sizes and species (i.e., forest with mixed and pure species composition) and in a variety of stand conditions (i.e., even-aged and uneven-aged forests) in Portugal. Considering all the species in the same model and recognizing the important differences among forest cover types is instrumental to incorporate fire-smart management considerations into forest planning. In addition, these models can easily be implemented in decision support systems that may allow the manager to minimize the expected losses due to wildfires when developing management plans either at stand-level [17,28,29,40,81,82] or at landscape-level [ 27 , 83 ]. Moreover, they help to reduce the uncertainty by predicting the outcomes of different management alternatives [ 84 ]. Thus, this set of models confirmed the potential of the proposed approach, and can be valuable to integrate effectively long-term fire-effects into any forest management planning. Future research efforts on this topic would benefit from the existence of information from permanent plots, measured over time, and where detailed data on the understory (evolution of fuel load) and overstory composition and structure are collected. Supplementary Materials: The following are available online at www.mdpi.com/2071-1050/9/3/390/s1. Acknowledgments: This research was supported by Project UID/AGR/00239/2013, PTDC/AGR-CFL/64146/2006 “Decision support tools for integrating fire and forest management planning” and project FIRE-ENGINE “Flexible Design of Forest Fire Management Systems” (MIT/FSE/0064/2009), both funded by the Portuguese Science Foundation (FCT), and contributes to the activities of the ALTERFOR Project “Alternative models and robust decision-making for future forest management”—H2020-ISIB-2015-2/grant agreement No. 67654, funded by European Union Seventh Framework Programme. This research has received also funding from the European Union’s H2020 research and innovation program under the Marie Sklodowska-Curie grant agreement No. 691149 (SuFoRun). The authors would like to thank the Portuguese Science Foundation for funding the doctoral scholarships of Brigite Botequim (SFRH/ BD/44830/2008) and the Post Doc grant SFRH/BPD/96806/2013 of Susete Marques. Researcher Jordi Garcia-Gonzalo was supported by a “Ramon y Cajal” research contract from the MINECO (Ref. RYC-2013-14262) and has received funding from CERCA Programme / Generalitat de Catalunya. In addition, the authors wish to acknowledge the Portuguese Forest Service (ICNF) for supplying the perimeters of wildfires and NFI Databases and ForFireS Project for providing the inventory Databases. Author Contributions: All authors have cooperated in the preparation of this research and contributed to produce the present paper. The individual contribution and responsibilities of the authors are as follows: Brigite Botequim and Jordi Garcia-Gonzalo designed the experiments; JoséG. Borges coordinate the research tasks; Brigite Botequim, Andreia Silva and Susete Marques collected and analyzed the data; Brigite Botequim, and Manuel Arias-Rodil performed the statistical treatments for fixed-effects modeling; Manuel Arias-Rodil performed the statistical treatments for mixed-effects modeling; Maria Manuela Oliveira contributed with statistical methods; Brigite Botequim, Manuel Arias-Rodil and Margarida Toméanalyzed and discussed statistical outcomes; Brigite Botequim, Manuel Arias-Rodil and Jordi Garcia-Gonzalo wrote the paper; and a final review, including final manuscript revisions, was performed by JoséG. Borges. Conflicts of Interest: The authors declare no conflict of interest. References 1. ICNF. Áreas Dos Usos Do Solo E Das Espécies Florestais de Portugal Continental. Resultados Preliminares, 6 º Inventário Florestal Nacional. 2013, p. 34. Available online: http://www.icnf.pt/portal/florestas/ifn/ resource/ficheiros/ifn/ifn6-res-prelimv1-1 (accessed on 8 August 2016). (In Portuguese) 2. Pereira, M.G.; Aranha, J.; Amraoui, M. Land cover fire proneness in Europe. For. Syst. 2014 ,23, 598–610. [CrossRef] 3. Rothermel, R.C. How to Predict the Spread and Intensity of Forest and Range Fires; U.S. Department of Agriculture, Forest Service, Intermountain Forest and Range Experiment Station: Ogden, UT, USA, 1983. 4. Silva, J.S.; Moreira, F.; Vaz, P.; Catry, F.; Godinho-Ferreira, P. Assessing the Relative Fire Proneness of Different Forest Types in Portugal. Plant Biosyst. 2009,143, 597–608. [CrossRef] Sustainability 2017,9, 390 20 of 23 5. Mateus, P.; Fernandes, P.M. Forest Fires in Portugal: Dynamics, Causes and Policies. In Forest Context and Policies in Portugal; Springer: Berlin, Germany, 2014; pp. 97–115. 6. Catry, F.X.; Pausas, J.G.; Moreira, F.; Fernandes, P.M.; Rego, F. Post-Fire Response Variability in Mediterranean Basin Tree Species in Portugal. Int. J. Wildland Fire 2013,22, 919–932. [CrossRef] 7. Del Río, M.; Pretzsch, H.; Alberdi, I.; Bielak, K.; Bravo, F.; Brunner, A.; Condés, S.; Ducey, M.J.; Fonseca, T.; von Lüpke, N. Characterization of the Structure, Dynamics, and Productivity of Mixed-Species Stands: Review and Perspectives. Eur. J. For. Res. 2016,135, 23–49. [CrossRef] 8. Lee, S.-W.; Lee, M.-B.; Lee, Y.-G.; Won, M.-S.; Kim, J.J.; Hong, S. Relationship between Landscape Structure and Burn Severity at the Landscape and Class Levels in Samchuck, South Korea. For. Ecol. Manag. 2009 ,258, 1594–1604. [CrossRef] 9. Kint, V.; Geudens, G.; Mohren, G.M.J.; Lust, N. Silvicultural Interpretation of Natural Vegetation Dynamics in Ageing Scots Pine Stands for Their Conversion into Mixed Broadleaved Stands. For. Ecol. Manag. 2006 , 223, 363–370. [CrossRef] 10. Jactel, H.; Nicoll, B.C.; Branco, M.; Gonzalez-Olabarria, J.R.; Grodzki, W.; Långström, B.; Moreira, F.; Netherer, S.; Orazio, C.; Piou, D. The Influences of Forest Stand Management on Biotic and Abiotic Risks of Damage. Ann. For. Sci. 2009,66, 1–18. [CrossRef] 11. Nunes, L.; Lopes, D.; Rego, F.C.; Gower, S.T. Aboveground Biomass and Net Primary Production of Pine, Oak and Mixed Pine-oak Forests on the Vila Real District, Portugal. For. Ecol. Manag. 2013 ,305, 38–47. [CrossRef] 12. Finney, M.A.; Andrews, P.L. FARSITE—A Program for Fire Growth Simulation. Fire Manag. Notes 1999 ,59, 13–15. 13. Finney, M.A. An Overview of FlamMap Fire Modeling Capabilities. Fuels Management—How to Measure Success. USDA Forest Service Proceedings RMRS-P-41. Andrews, P.L., Butler, B.W., Eds.; pp. 213–220. Available online: https://www.fs.fed.us/rm/pubs/rmrs_p041.pdf (accessed on 6 March 2017). 14. Ryan, K.C. Evaluating Potential Tree Mortality from Prescribed Burning. In Proceedings of the Site Preparation and Fuels Management on Steep Terrain, Pullman, WA, USA, 15–17 February 1982; pp. 167–179. 15. Keyser, T.L.; Smith, F.W.; Lentile, L.B.; Shepperd, W.D. Modeling Postfire Mortality of Ponderosa Pine Following a Mixed-Severity Wildfire in the Black Hills: The Role of Tree Morphology and Direct Fire Effects. For. Sci. 2006,52, 530–539. 16. Hull Sieg, C.; McMillin, J.D.; Fowler, J.F.; Allen, K.K.; Negron, J.F.; Wadleigh, L.L.; Anhold, J.A.; Gibson, K.E. Best Predictors for Postfire Mortality of Ponderosa Pine Trees in the Intermountain West. For. Sci. 2006 ,52, 718–728. 17. González, J.R.; Trasobares, A.; Palahi, M.; Pukkala, T. Predicting Stand Damage and Tree Survival in Burned Forests in Catalonia (North-East Spain). Ann. For. Sci. 2007,64, 733–742. [CrossRef] 18. Garcia-Gonzalo, J.; Marques, S.; Borges, J.G.; Botequim, B.; Oliveira, M.M.; Tomé, J.; Tomé, M. A Three-Step Approach to Post-Fire Mortality Modelling in Maritime Pine (Pinus pinaster Ait) Stands for Enhanced Forest Planning in Portugal. Forestry 2011,84, 197–206. [CrossRef] 19. Linder, P.; Jonsson, P.; Niklasson, M. Tree Mortality after Prescribed Burning in an Old-Growth Scots Pine Forest in Northern Sweden. Silva Fenn. 1998,32, 339–349. [CrossRef] 20. Hély, C.; Flannigan, M.; Bergeron, Y. Modeling Tree Mortality Following Wildfire in the Southeastern Canadian Mixed-Wood Boreal Forest. For. Sci. 2003,49, 566–576. 21. McHugh, C.W.; Kolb, T.E. Corrigendum to: Ponderosa Pine Mortality Following Fire in Northern Arizona. Int. J. Wildland Fire 2003,12, 245. [CrossRef] 22. Fernandes, P.M. Combining Forest Structure Data and Fuel Modelling to Classify Fire Hazard in Portugal. Ann. For. Sci. 2009,66, 1–9. [CrossRef] 23. Fernandes, P.M.; Luz, A.; Loureiro, C. Changes in Wildfire Severity from Maritime Pine Woodland to Contiguous Forest Types in the Mountains of Northwestern Portugal. For. Ecol. Manag. 2010 ,260, 883–892. [CrossRef] 24. Agee, J.K.; Skinner, C.N. Basic Principles of Forest Fuel Reduction Treatments. For. Ecol. Manag. 2005 ,211, 83–96. [CrossRef] 25. Gonzalez, J.R.; Palahi, M.; Trasobares, A.; Pukkala, T. A Fire Probability Model for Forest Stands in Catalonia (North-East Spain). Ann. For. Sci. 2006,63, 169–176. [CrossRef] Sustainability 2017,9, 390 21 of 23 26. Marques, S.; Garcia-Gonzalo, J.; Borges, J.G.; Botequim, B.; Oliveira, M.M.; Tomé, J.; Tomé, M. Developing Post-Fire Eucalyptus globulus Stand Damage and Tree Mortality Models for Enhanced Forest Planning in Portugal. Silva Fenn. 2011,45, 69–83. [CrossRef] 27. Gonzalez, J.R.; Palahí, M.; Pukkala, T. Integrating Fire Risk Considerations in Forest Management Planning in Spain—A Landscape Level Perspective. Landsc. Ecol. 2005,20, 957–970. [CrossRef] 28. Garcia-Gonzalo, J.; Pukkala, T.; Borges, J.G. Integrating Fire Risk in Stand Management Scheduling. An Application to Maritime Pine Stands in Portugal. Ann. Oper. Res. 2011,219, 379–395. [CrossRef] 29. Pasalodos-Tato, M.; Pukkala, T.; Alboreca, A.R. Optimal Management of Pinus pinaster in Galicia (Spain) under Risk of Fire. Int. J. Wildland Fire 2010,19, 937–948. [CrossRef] 30. Botelho, H.S.; Rego, F.C.; Ryan, K.C. Tree Mortality Models for Pinus Pinaster of Northern Portugal. In Proceedings of the 13th Conference on Fire and Forest Meteorology; IAWF: Lorne, Australia, 1998; pp. 235–240. 31. Fernandes, P.A.; Loureiro, C.A.; Botelho, H.S. Fire Behaviour and Severity in a Maritime Pine Stand under Differing Fuel Conditions. Ann. For. Sci. 2004,61, 537–544. [CrossRef] 32. Fernandes, P.M.; Rigolot, E. The Fire Ecology and Management of Maritime Pine (Pinus pinaster Ait.). For. Ecol. Manag. 2007,241, 1–13. [CrossRef] 33. Catry, F.X.; Moreira, F.; Duarte, I.; Acácio, V. Factors Affecting Post-Fire Crown Regeneration in Cork Oak (Quercus suber L.) Trees. Eur. J. For. Res. 2009,128, 231–240. [CrossRef] 34. Catry, F.X.; Rego, F.; Moreira, F.; Fernandes, P.M.; Pausas, J.G. Post-Fire Tree Mortality in Mixed Forests of Central Portugal. For. Ecol. Manag. 2010,260, 1184–1192. [CrossRef] 35. Catry, F.X.; Moreira, F.; Tujeira, R.; Silva, J.S. Post-Fire Survival and Regeneration of Eucalyptus globulus in Forest Plantations in Portugal. For. Ecol. Manag. 2013,310, 194–203. [CrossRef] 36. Nunes, L.; Coutinho, J.; Nunes, L.F.; Rego, F.C.; Lopes, D. Growth, Soil Properties and Foliage Chemical Analysis Comparison. For. Syst. 2011,20, 496–507. [CrossRef] 37. Kolström, M.; Lindner, M.; Vilén, T.; Maroschek, M.; Seidl, R.; Lexer, M.J.; Netherer, S.; Kremer, A.; Delzon, S.; Barbati, A. Reviewing the Science and Implementation of Climate Change Adaptation Measures in European Forestry. Forests 2011,2, 961–982. [CrossRef] 38. Keeley, J.E.; Bond, W.J.; Bradstock, R.A.; Pausas, J.G.; Rundel, P.W. Fire in Mediterranean Ecosystems: Ecology, Evolution and Management; Cambridge University Press: New York, NY, USA, 2012. 39. Naeslund, B.A. Simulation of Damage and Mortality in Young Stands and Associated Stand Development Effects [Incl. Betula Pendula, Betula Verrucosa, Betula Pubescens]; Sveriges lantbruksuniv.: Uppsala, Sweden, 1986; ISBN: 9789157626493. 40. Ferreira, L.; Constantino, M.; Borges, J.G. Optimizing Management of Pinus pinaster Ait Stands, under Risk of Fire. An Application in Portugal. Ann. Oper. Res. 2011,219, 359–377. [CrossRef] 41. Ferreira, L.; Constantino, M.F.; Borges, J.G.; Garcia-Gonzalo, J. A Stochastic Dynamic Programming Approach to Optimize Short-Rotation Coppice Systems Management Scheduling: An Application to Eucalypt Plantations under Wildfire Risk in Portugal. For. Sci. 2012,58, 353–365. [CrossRef] 42. Oliveira, S.L.; Pereira, J.M.; Carreiras, J.M. Fire Frequency Analysis in Portugal (1975–2005), Using Landsat-Based Burnt Area Maps. Int. J. Wildland Fire 2012,21, 48–60. [CrossRef] 43. Bergeot, F.; Oliveira, T. Development of a Simple and Efficient Method for Field Assessment of Fire Severity; (ForFireS) Contract Number 382167 FISC, Final Report; Inventaire Forestier National: Maurin, France, 2008. 44. Botequim, B.; Garcia-Gonzalo, J.; Marques, S.; Ricardo, A.; Borges, J.G.; Oliveira, M.M.; Tomé, J.; Tomé, M. Assessing wildfire risk probability in Eucalyptus globulus Labill stands in Portugal. iForest 2013 ,6, 217–227. [CrossRef] 45. Dieguez, U.; Barrio, M.; Castedo, F.; Balboa, M. Estimacion Del Diametro Normal Y Del Volumen Del Tronco a Partir de Las Dimensiones Del Tocon Para Seis Especies Forestales Comerciales de Galicia; Instituto Nacional de Investigacion y Tecnologia Agraria y Alimentaria, Madrid (España). Sist. Recur. For. 2003 ,12, 131–139. 46. Tomé, M.; Meyer, A.; Ramos, T.; Barreiro, S.; Faias, S.P.; Cortiçada, A. Relações Hipsométricas E Equações de Diâmetro Da Copa Desenvolvidas No Âmbito Do Tratamento Dos Dados Do Inventário Florestal Nacional 2005–2006; Publicações GIMREF RT, Universidade Técnica de Lisboa; Instituto Superior de Agronomia: Lisboa, Portugal, 2007; p. 23. (In Portuguese) 47. Soares, P.; Tomé, M. Height-diameter Equation for First Rotation Eucalypt Plantations in Portugal. For. Ecol. Manag. 2002,166, 99–109. [CrossRef] Sustainability 2017,9, 390 22 of 23 48. Monserud, R.A.; Sterba, H. Modeling Individual Tree Mortality for Austrian Forest Species. For. Ecol. Manag. 1999,113, 109–123. [CrossRef] 49. Vanclay, J.K. Modelling Forest Growth and Yield: Applications to Mixed Tropical Forests; School of Environment, Science and Engineering Papers; CAB International: Wallingford, UK, 1994; p. 537. 50. Hosmer, D.W.; Lemeshow, S. Applied Logistic Regression, 2nd ed.; John Wiley and Sons: New York, NY, USA, 2000. 51. R Core Team. R: A Language and Environment for Statistical Computing; R Core Team: Vienna, Austria, 2016. 52. Hamilton, D.A. A Logistic Model of Mortality in Thinned and Unthinned Mixed Conifer Stands of Northern Idaho. For. Sci. 1986,32, 989–1000. 53. Das, A.J.; Battles, J.J.; Stephenson, N.L.; Van Mantgem, P.J. The Relationship between Tree Growth Patterns and Likelihood of Mortality: A Study of Two Tree Species in the Sierra Nevada. Can. J. For. Res. 2007 ,37, 580–597. [CrossRef] 54. Bates, D.M.; Maechler, M.; Bolker, B.; Walker, S. Fitting linear mixed-effects models using lme4. J. Stat. Softw. 2015,67, 1–48. [CrossRef] 55. Fridman, J.; Ståhl, G. A Three-Step Approach for Modelling Tree Mortality in Swedish Forests. Scand. J. For. Res. 2001,16, 455–466. [CrossRef] 56. DeBano, L.F.; Neary, D.G.; Ffolliott, P.F. Fire Effects on Ecosystems; John Wiley & Sons: New York, NY, USA, 1998. 57. Freire, J.P.A. Modelação Do Crescimento E Da Produção de Pinha No Pinheiro Manso; Tese de Doutoramento, Instituto Superior de agronomia, Universidade de Lisboa: Lisboa, Portugal, 2009; p. 215. (In Portuguese) 58. Kleinbaum, D.G. Logistic Regression: A Self-Learning Text. S. 5: 103–104. Stat. Methods Med. Res. 1994 ,5, 103–104. 59. Ryan, T.P. Modern Regression Methods; John Wiley & Sons: New York, NY, USA, 1997. 60. Kozak, A.; Kozak, R. Does cross validation provide additional information in the evaluation of regression models? Can. J. For. Res. 2003,33, 976–987. [CrossRef] 61. Jalkanen, A.; Mattila, U. Logistic Regression Models for Wind and Snow Damage in Northern Finland Based on the National Forest Inventory Data. For. Ecol. Manag. 2000,135, 315–330. [CrossRef] 62. Beverly, J.L.; Martell, D.L. Modeling Pinus Strobus Mortality Following Prescribed Fire in Quetico Provincial Park, Northwestern Ontario. Can. J. For. Res. 2003,33, 740–751. [CrossRef] 63. Botelho, H.S.; Fernandes, P.M.; Ruas, L.L.S. Modeling Pinus pinaster Induced by up-Slope Wind Driven Prescribed Fires in Northern Portugal. In Proceedings of the 13th Fire and Forest Meteorology Conference, Lorne, Australia, 27–31 October 1996; pp. 473–476. 64. Regelbrugge, J.C.; Conard, S.G. Modeling Tree Mortality Following Wildfire in Pinus Ponderosa Forests in the Central Sierra-Nevada of California. Int. J. Wildland Fire 1993,3, 139–148. [CrossRef] 65. Rigolot, E. Predicting Postfire Mortality of Pinus halepensis Mill. and Pinus pinea L. Plant Ecol. 2004 ,171, 139–151. [CrossRef] 66. Eid, T.; Øyen, B.-H. Models for Prediction of Mortality in Even-aged Forest. Scand. J. For. Res. 2003 ,18, 64–77. 67. González, J.G.Á.; Dorado, F.C.; González, A.D.R.; Sánchez, C.A.L.; Von Gadow, K. A Two-Step Mortality Model for Even-Aged Stands of Pinus Radiata D. Don in Galicia (Northwestern Spain). Ann. For. Sci. 2004 , 61, 439–448. [CrossRef] 68. Peterson, D.L.; Ryan, K.C. Modeling Postfire Conifer Mortality for Long-Range Planning. Environ. Manag. 1986,10, 797–808. [CrossRef] 69. Lohmander, P.; Helles, F. Windthrow Probability as a Function of Stand Characteristics and Shelter. Scand. J. For. Res. 1987,2, 227–238. [CrossRef] 70. Bond, W.J.; van Wilgen, B.W. Fire and Plants—Population and Community Biology Series 14; Springer: Berlin, Germany, 1996; p. 263. ISBN: 978-94-009-1499-5. 71. Moreira, F.; Ferreira, P.G.; Rego, F.C.; Bunting, S. Landscape Changes and Breeding Bird Assemblages in Northwestern Portugal: The Role of Fire. Landsc. Ecol. 2001,16, 175–187. [CrossRef] 72. Moreira, F.; Vaz, P.; Catry, F.; Silva, J.S. Regional Variations in Wildfire Susceptibility of Land-Cover Types in Portugal: Implications for Landscape Management to Minimize Fire Hazard. Int. J. Wildland Fire 2009 ,18, 563–574. [CrossRef] 73. Pausas, J.G. Resprouting of Quercus suber in NE Spain after Fire. J. Veg. Sci. 1997,8, 703–706. [CrossRef] Sustainability 2017,9, 390 23 of 23 74. Retana, J.; Riba, M.; Castell, C.; Espelta, J.M. Regeneration by Sprouting of Holm-Oak (Quercus Ilex) Stands Exploited by Selection Thinning. In Quercus ilex L. Ecosystems: Function, Dynamics and Management; Springer: Berlin, Germany, 1992; pp. 355–364. 75. Pollet, J.; Omi, P.N. Effect of Thinning and Prescribed Burning on Crown Fire Severity in Ponderosa Pine Forests. Int. J. Wildland Fire 2002,11, 1–10. [CrossRef] 76. Catry, F.X.; Moreira, F.; Pausas, J.G.; Fernandes, P.M.; Rego, F.; Cardillo, E.; Curt, T. Cork Oak Vulnerability to Fire: The Role of Bark Harvesting, Tree Characteristics and Abiotic Factors. PLoS ONE 2012 ,7, e39810. [CrossRef] [PubMed] 77. Pyne, S.J.; Andrews, P.L.; Laven, R.D. Introduction to Wildland Fire; John Wiley and Sons: Hoboken, NJ, USA, 1996. 78. Ventura, J.; Vasconcelos, M.J. O Fogo Como Processo Físico-Químico E Ecológico. Incêndios Florestais em Portugal Impactes E Prevenção; Instituto Superior de Agronomia: Lisboa, Portugal, 2006; pp. 93–113. 79. Garcia-Gonzalo, J.; Zubizarreta-Gerendiain, A.; Ricardo, A.; Marques, S.; Botequim, B.; Borges, J.G.; Oliveira, M.M.; Tomé, M.; Pereira, J.M.C. Modelling wildfire risk in pure and mixed forest stands in Portugal. Allgemeine Forst und Jagdzeitung (AFJZ) 2012,183, 238–248. 80. Marques, S.; Garcia-Gonzalo, J.; Botequim, B.; Ricardo, A.; Borges, J.G.; Tomé, M.; Oliveira, M.M. Assessing wildfire risk probability in Pinus pinaster Ait. stands in Portugal. For. Syst. 2011,21, 111–120. 81. González, J.R.; Pukkala, T.; Palahí, M. Optimising the Management of Pinus sylvestris L. Stand under Risk of Fire in Catalonia (North-East of Spain). Ann. For. Sci. 2005,62, 493–501. [CrossRef] 82. González-Olabarría, J.R.; Palahí, M.; Pukkala, T.; Trasobares, A. Optimising the Management of Pinus nigra Arn. Stands under Endogenous Risk of Fire in Catalonia. For. Syst. 2008,17, 10–17. [CrossRef] 83. González-Olabarria, J.-R.; Pukkala, T. Integrating Fire Risk Considerations in Landscape-Level Forest Planning. For. Ecol. Manag. 2011,261, 278–287. [CrossRef] 84. Von Gadow, K. Evaluating Risk in Forest Planning Models. Silva Fenn. 2000,34, 181–191. © 2017 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).