Wildfires are one of the most important environmental problems at present. In the European context, Spain is one of the countries which registers the highest fire incidence in number and surface burned [1
]. These events cause a great amount of damages (i.e., degradation of soils, water and biodiversity), which may further lead to economic and even human costs. From an environmental point of view, wildfires cause alterations of landcover, loss of carbon reserves, and changes in the soil composition and its hydrogeomorphologic behavior [4
]. Nevertheless, Mediterranean vegetation is quite adapted to fire disturbance. Mediterranean species have post-fire ecological strategies, like the resprout ability, the seed bank persistency, or the growth or dispersal ability [4
]. In this sense, the analysis of Large Forest Fires (LFF) acquires a special interest because the effects they provoke are devastating. For this reason, the application of an adequate restoration policy requires the exhaustive study of the physical environment and dynamic evolution of the affected area [5
Over the last few years, several studies have underlined the importance of remote sensing in analyzing ecological dynamics following fire and studying post-fire vegetation regeneration [5
]. Compared to field surveying, satellite images offer a less expensive alternative and provide broader information of burned areas by obtaining biophysical variables of wildfires [7
Medium-resolution optical sensors such as MSS, TM, ETM+, and OLI of Landsat series enables the monitoring of vegetation in burned areas for more than 40 years [8
]. There are also researches studying vegetation regeneration using other sensors such as AVIRIS [11
], AVHRR [12
], MODIS [13
], SPOT-VEGETATION [14
], or RADAR [15
]. According to Hirschmugl et al. [16
], time series analysis is the most used approach in the monitoring of forest disturbances and degradation, as well as regeneration processes with the objective to analyze spectral variations in the forest cover [8
The factors which define the vegetation regeneration rate after a wildfire are multiple, and their complete identification or modeling is difficult. These factors are related to wildfire characteristics, environmental conditions, and the life history of plant species [18
]. Some studies applied to northern regions [10
] and Mediterranean regions [8
] have pointed out some of the main drivers of vegetation regeneration following forest fires. Fire severity levels, topography (elevation, slope, and orientation), post-fire climate, or vegetation cover class are the most used in regeneration estimates. Other studies have emphasized the influence of solar radiation on the water availability for vegetation growth [8
] or the effects of applying different restoration models [9
]. Several studies have revealed that the influence of environmental factors on regeneration can vary across vegetation types [8
In order to evaluate fire severity, defined as the magnitude of the ecological change produced by fire [6
], different methods have been proposed. Some studies have successfully applied vegetation indices such as the Normalized Difference Vegetation Index
], due to the relationship between the amount of vegetation consumed and fire severity [7
]. On the other hand, specific indices have been developed which record with greater spectral contrast the fire effects as the Normalized Burn Ratio
] subsequently modified [25
]. Another method based on parametrized variables to estimate fire effects in vertical strata is the Composite Burn Index
], widely and effectively applied with good results (correlation between CBI and dNBR R2
= 0.83) [27
] and modified by De Santis and Chuvieco [28
To assess vegetation regeneration using time series, several methodologies have also been proposed. The most common is to monitor the vegetation state from spectral indices. Among the most used indices are the Normalized Difference Vegetation Index
], the Regeneration Index
], the Normalized Difference Infrared Index
], and the Soil-Adjusted Vegetation Index
]. Spectral Mixture Analysis
has also effectively been applied in that context [8
Although several studies have evaluated fire severity and vegetation regeneration using remote sensing data, few have integrated both aspects in a single study. Several studies have identified the significant influence of fire severity [21
] and environmental factors, such as topography and climate [8
], in vegetation regeneration. Therefore, the need to integrate both analyses to improve the predictive models of these dynamics has arisen. In recent years, more studies have been undertaken to investigate the factors that determine post-fire regeneration patterns with satisfactory results [19
In this study, we hypothesize that fire severity and environmental variables such as vegetation type, meteorology, and topography determine the post-fire vegetation regeneration. Therefore, regeneration patterns on burned surfaces will vary between areas presenting different severity levels. Furthermore, meteorological post-fire conditions and topography will have different impacts on the regeneration of different forest covers.
The general objective will be to model the short-term vegetation regeneration (five years after fire) in three large forest fires in the Mediterranean region of the Iberian Peninsula, by knowing the fire severity and the interacting environmental variables. In this case, we will model the forest regeneration of the genus Pinus
, which is widely extended in Mediterranean forests. In addition, Pinus
communities are one of the most affected by fires due to the high content of resins that promote both fire start and spread [33
Specific objectives of the study area include: (1) To model the evolution of post-fire vegetation to obtain regrowth patterns through spectral indices; (2) Generate environmental variables involved in regeneration and (3) Identify the variables which are most relevant in explaining the short-term regeneration using multiple regression models and to estimate the short-term regeneration of Pinus species.
3.1. Regeneration Evolution According to Fire Severity
First analysis has focused on the evolution of the short-term regeneration of the selected Pinus
species according to the severity level. According to fire severity level, Pinus halepensis
evolution on the one hand, and other Pinus
) on the other, from the mean NDVI values, has been represented (Figure 3
). This separation is based on the unequal germination capacity of the species, being higher in Pinus halepensis
Results show a significant decrease in the post-fire values and a later progressive recovery, being faster in the early years in the case of Pinus nigra and pinaster, while in Pinus halepensis, recovery is slower. Fire severity level affects Pinus regeneration significantly, being higher NDVI values associated with lower fire severity. Pre-fire values are different between fire severity levels due to the influence of other environmental variables according to the zone. In addition, areas which were drier before fire may have burned more intensely. Some fluctuations in the short-term regeneration may be caused by variations in the images acquisition dates after summer. This could explain the fall in NDVI values in Pinus nigra and Pinus pinaster between 1997 and 1998.
3.2. Multiple Linear Regression with Ordinary Least Squares
An exploratory regression analysis with OLS allowed for the elimination of some variables initially defined because of multicollinearity problems (spatial autocorrelation) or the low significance for the predictive models.
The average of the climatic anomalies between 1994 and 1999 were dismissed because of their high multicollinearity with post-fire climatic anomalies (1994 and 1995). Moreover, the low spatial resolution of the meteorological data implies a reduced variability of precipitation and temperature data, and there is a high multicollinearity between maximum and minimum temperatures. Therefore, two models were created for each group of species, with one model having maximum temperature, while the other model used minimum temperature. The resulting models were then evaluated and the model with best fit selected. Additionally, the explanatory variables included are the most consistent, that is, those that were significant a greater number of times (>80) when executing the regression with OLS 100 times (Table 4
). In this case, the orientation variables have not been significant and have been removed.
and Table 6
contain the first regression models results. The models that have shown the best results have been those in which the minimum temperatures have not been included because they were redundant (high value of the Variance Inflation Factor—VIF).
and Adjusted R2
values show a good fit for Pinus halepensis
model (Table 5
). In order to assess in detail each explanatory variable, the coefficient, t
-statistic, robust probability, and VIF were used.
Each explanatory variable coefficient shows the relationship between each explanatory and dependent variable. In this case, slope, severity levels, and Tmax anomalies have a negative influence. High slope slows down rooting vegetation while abnormally high temperatures can cause thermal stress and limit regeneration. As anticipated, short-term regeneration rate was lower when severity was higher. Nevertheless, the severity seems to influence less strongly because sample pixels used coincided with high severity values. Consequently, the influence of a low or medium severity on regeneration is not collected, leaving this variable weak against others. On the other hand, the variable with a greater positive impact is the state of the vegetation after fire (NDVI+1). That is, short-term regeneration rate was higher when post-fire vegetation greenness was higher. There is also a positive influence on elevation and precipitation. In the Mediterranean region, above-average rainfall can help the growth of vegetation by increasing soil moisture. The elevation could be related to lower temperatures avoiding thermal stress in summer.
Standard errors allow assessment of the coefficients obtained. Low standard errors across all variables indicate coefficients are consistent. On the other hand, the robust probability or p-value showed that all variables are statistically significant and, therefore, important for the regression model. VIF provided information on the possible redundancy in explanatory variables. In this case, the relatively high values for climatic anomalies suggest removing one of the two variables could increase the model fit.
As regards to the regression model for Pinus nigra
), the model shows lower Multiple R2
and Adjusted R2
values due to differences in the vegetative cycle between modelled species. Again, variables with inverse relationships are fire severity and slope, whereas the post-fire vegetation state (greeness) is the variable with more positive influence. Certain above-average rainfall and temperatures higher than usual show a positive coefficient that could be due to compensations in evapotranspiration. Standard errors are reduced and VIF values confirm the absence of variable redundancy in this model.
However, in both models, the significance of Koenker statistic (p-value less than 0.05 for a 95% confidence level) indicated biased standard errors due to heteroscedasticity. In addition, the significant p-value in the Jarque-Bera statistic showed that the residual values deviated slightly from a normal distribution. Finally, an analysis of the residuals was carried out to study spatial autocorrelation using the Moran Index. In both, autocorrelation has been positive, with values around 0.3. Therefore, it was suitable to run the Geographically Weighted Regression (GWR) analysis.
3.3. Multiple Linear Regression with Geographically Weighted Regression (GWR)
Regression with GWR allows for the improvement of the adjustments and for the neutralizing of the spatial dependence in residual values. In comparison with a global regression, the coefficients in GWR are functions of spatial location [56
]. Regression analyses were carried out using a defined kernel with a two-square function in which the bandwidth was determined by an optimal number of neighbors. The optimal number of neighbors has been 500, where the value of the Akaike Information Criterion (AIC), which is a relative quality indicator of a model, is the lowest. In final models, anomalies in maximum temperatures were removed because of local multicollinearity problems.
and adjusted R2
values obtained in the models using GWR show a significant improvement over the OLS model. In this case, the variables used provide a model explaining 80% of Pinus halepensis
regeneration and 78% for Pinus nigra
and Pinus pinaster
). It is noticed that explanatory variable forces and influences have changed over OLS models for Pinus nigra
and Pinus pinaster
. Higher slope shows a slightly positive coefficient, as in the study by Meng et al. [20
] for spruce forests. In contrast, the elevation has a negative coefficient that could be related to temperatures too low, which can limit regeneration. However, forces and influences are practically maintained in the Pinus halepensis
regression model. Higher slope and higher fire severity have negative influence, while increasing the elevation and the anomalies in precipitations have a positive influence.
However, the regression models do not have an equal predictive capacity for all areas, so there are regional variations (Figure 4
). Models show a better fit for Pinus halepensis
regeneration in Moratalla and Pinus pinaster
in Castrocontrigo. These spatial variations could be related to the sampling point distribution by a smaller number of neighbors, because in Uncastillo sampling points are distributed in a more dispersed way, and by the local conditions effects (microclimate, edaphic composition).
Finally, the relative importance of the GWR explanatory variables were analyzed from the t
-statistic (Figure 5
). As expected, the post-fire vegetation state (NDVI+1) is the most important variable to the model. In contrast, fire severity was found to have a weak influence on the other variables in both models. Elevation is the variable that displayed the most difference among species, and this was attributed to species adaptation to temperatures decreasing with elevations. This decrease can be positive to avoid abnormally high temperatures in summer in Moratalla (excess evapotranspiration), but can act as a restriction in Castrocontrigo in case of too low temperatures (reduced plant activity).
Moreover, analysis of the residual values also shows better results in GWR than in OLS. The Moran Index values are much closer to the expected values with values below 0.05, also showing a lower variance and greater probability of random distribution (p value and Z score).
In order to validate the models, we used the root-mean-square error (RMSE) which measures the average of the squared errors (errors among NDVI values on 1999 and the values predicted by OLS and GWR) (Table 8
). Results show that GWR models have smaller error, which implies a better adjustment than OLS models. Taking into account that NDVI values fluctuate among −1 and 1, errors are low. Therefore, predictions obtained using GWR could be considered valid.
Several studies have researched post-fire vegetation regeneration based on growth patterns according to species, or on the relationship among fire severity and regeneration dynamics. Nevertheless, other previous studies have supported the importance of delving into the analysis of factors which determine regeneration [10
]. Therefore, in spite of the difficulties to obtain information to generate certain environmental variables, in this study it has been considered important to model the regeneration considering post-fire climate and topography.
The results show a clear relationship between fire severity levels and regeneration rate values, which are slower when fire severity level is higher [19
]. Meng et al. [20
] and Ireland & Petropoulos [19
] agree on the importance that solar radiation levels have on vegetation growth rate. In Mediterranean contexts, north aspect and high slopes present higher regeneration rates due to less evapotranspiration and higher humidity content. However, this phenomenon has not been analyzed because orientation variables have not been significant for the model in the cases studied. With regards to elevation, it has been shown that it could act positively to avoid high temperatures and periods of post-fire drought [53
] in more arid environments, such as Moratalla. This is in agreement with Meng et al. [20
] for coniferous forests in Sierra Nevada (USA).
With regards to the influence of post-fire weather, a positive relationship between short-term regeneration and above-average precipitation in the months after fire has been found. This relationship was also found for Pinus halepensis
and Pinus nigra
in Catalonia [21
] and Pinus halepensis
in Ayora in Valencia [8
]. Nevertheless, climatic data have a lower spatial resolution compared to Landsat images and topographic data. Consequently, multicollinearity problems have not allowed for a robust analysis of the temperatures or precipitation impact on regeneration. In addition, it would be interesting to include drought indices, which would require greater spatial and temporal resolution data.
Modeling results could also be improved by incorporating new variables that capture local conditions such as landscape configuration, local microclimate, and hydrological processes which can determine vegetation recovery after fire [4
]. In the study carried out by Chu et al. [10
], water soil content showed a positive relationship with the regeneration of the larch forest, being the second conditioning factor in the recovery of this species after fire. Moreover, Röder et al. [8
] concluded that water availability, in relation to slope and edaphic composition, is the most important factor that limits vegetation recovery.
In addition, it should be noted that short-term regeneration measured, through the NDVI, represents the relative vegetation cover. A direct measure of Pinus
forests regeneration could be obtained also considering the vegetation structure using LiDAR techniques [57
Our study provides advances in the analysis of the impact of fire severity and environmental variables on the short-term vegetation regeneration in Mediterranean regions.
Regeneration measurement for different species of Pinus from the NDVI has been related to the fire severity levels with good results. Severity degree measured by the CBI indicates that short-term regeneration was slowed down when severity was higher. In addition, the immediate NDVI values after the fire allowed us to verify that when the less damage produced in the vegetation, the short-term regeneration is greater.
From the multiple linear regression models generated, the explanatory capacity of the environmental variables of topography (elevation and slope) and post-fire climate (anomalies in precipitation) in post-fire vegetation recovery after fire has been verified/tested. The impact that each has on regeneration is closely related to the environmental characteristics to which each species is adapted. Thus, the elevation can be a driving factor for the Pinus nigra regeneration linked to temperatures too low or, in contrast, can favor Pinus halepensis growth, by avoiding high temperatures in summer in the southeast of the Iberian Peninsula.
In contrast to OLS, it has been confirmed that GWR is an important and valid local regression technique to explore spatial heterogeneity in the relations of explanatory variables. With this methodology, it has been possible to model the regeneration obtaining high adjustment values, with adjusted R2 values of 0.80 for Pinus halepensis and 0.77 for Pinus nigra and Pinus pinaster. The models’ improvement should focus on the generation of more precise environmental variables and considering new factors that can increase it explanatory power: lithologic characteristics, alteration of edaphic composition, solar radiation, time elapsed since the last fire, vegetation physical characteristics, etc.
Moreover, the proposed method is an approximation to the modelling of the short-term regeneration for the Iberian Peninsula but exportable to other territories with input variables transformation. The results obtained are useful in improving knowledge about the factors which determine the post-fire regeneration patterns of a forest ecosystem under different environmental and climatic conditions. Therefore, these advances could help decision-makers in determining which areas vegetation will not regenerate naturally after large fires and thus requires the implementation of specific restoration programs.