International Journal of Applied Earth Observations and Geoinformation 108 (2022) 102754

Contents lists available at ScienceDirect

International Journal of Applied Earth

Observations and Geoinformation
journal homepage: www.elsevier.com/locate/jag

Assessing post-fire forest structure recovery by combining LiDAR data and

Landsat time series in Mediterranean pine forests
Alba Viana-Soto *, Mariano García , Inmaculada Aguado , Javier Salas
Universidad de Alcalá. Departamento de Geología, Geografía y Medio Ambiente. Environmental Remote Sensing Research Group, Calle Colegios 2, 28801 Alcalá de
Henares, Spain

A R T I C L E I N F O A B S T R A C T

Keywords: Understanding post-fire recovery dynamics is critical for effective management that enhance forest resilience to
Post-fire recovery fire. Mediterranean pine forests have been largely affected by wildfires, but the impacts of both changes in land
Canopy cover use and climate endanger their capacity to naturally recover. Multispectral imagery is commonly used to esti­
Height
mate post-fire recovery, yet changes in forest structure must be considered for a comprehensive evaluation of
LiDAR
Landsat
forest recovery. In this research, we combine Light Detection And Ranging (LiDAR) with Landsat imagery to
support vector regression (SVR) extrapolate forest structure variables over a 30-year period (1990–2020) to provide insights on how forest
structure has recovered after fire in Mediterranean pine forests. Forest recovery was evaluated attending to
vegetation cover (VC), tree cover (TC), mean height (MH) and heterogeneity (CVH). Structure variables were
derived from two LiDAR acquisitions from 2016 and 2009, for calibration and independent spatial and temporal
validation. A Support Vector Regression model (SVR) was calibrated to extrapolate LiDAR-derived variables
using a series of Landsat imagery, achieving an R2 of 0.78, 0.64, 0.70 and 0.63, and a relative RMSE of 24.4%,
30.2%, 36.5% and 27.4% for VC, TC, MH and CVH, respectively. Models showed to be consistent in the temporal
validation, although a wider variability was observed, with R2 ranging from 0.51 to 0.74. A different response to
fire was revealed attending to forest cover and height since vegetation cover recovered to a pre-fire state but
mean height did not 26-years after fire. Less than 50% of the area completely recovered to the pre-fire structure
within 26 years, and the area subjected to fire recurrence showed signs of greater difficulty in initiating the
recovery. Our results provide valuable information on forest structure recovery, which can support the imple­
mentation of mitigation and adaptation strategies that enhance fire resilience.

1. Introduction frequency and intensity of fires, as a result of drier and longer droughts
(Seidl et al., 2017; Turco et al., 2018). Consequently, the vulnerability of
Forest fires are one of the recurrent disturbances in Mediterranean these ecosystems to not naturally recover is likely to increase, requiring
biomes, shaping forest structure and composition (Keeley et al., 2012). forest management to focus on their resilience to fire (Ibáñez et al.,
Fire activity is particularly intense in Southern Europe, where half a 2019).
million hectares are burnt every year (Rigo et al., 2017) causing In this forthcoming situation, post-fire recovery has emerged as an
ecological, economic and human health impacts (Bowman et al., 2017). important indicator of ecosystem resilience (González-De Vega et al.,
Since fire is an integral component of the Mediterranean ecosystems 2016; Nikinmaa et al., 2020), defined as the time the system needs to
(Pausas and Keeley, 2014), plant species have developed diverse post- reach the pre-disturbance values. Remote sensing offers a powerful
fire ecological strategies to recover including resprouting, seed bank alternative to field methods for monitoring forest recovery based on the
persistence and dispersal capacity (de las Heras et al., 2012). Nonethe­ spectral response of vegetation at multiple spatial and temporal scales
less, impacts of both changes in land use and climate, and the resulting (Pérez-Cabello et al., 2021). The Landsat archive, with more than four
alterations in fire regimes, may endanger their natural capability to decades of observations, has become the most widely used imagery
recover (Pausas and Keeley, 2014). Therefore, Mediterranean forests source to monitor forest disturbances and subsequent recovery (Chu­
must adapt to new induced fire regimes with an increase in the vieco et al., 2020). These datasets have both temporal and spatial

* Corresponding author.
E-mail address: [email protected] (A. Viana-Soto).

https://doi.org/10.1016/j.jag.2022.102754
Received 24 January 2022; Received in revised form 1 March 2022; Accepted 21 March 2022
Available online 25 March 2022
1569-8432/© 2022 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
A. Viana-Soto et al. International Journal of Applied Earth Observation and Geoinformation 108 (2022) 102754

characteristics appropriate to analyse dynamic processes as recovery, i) to assess the feasibility of temporally extrapolating LiDAR-based
giving an overview of successional forest changes following a distur­ cover, height and canopy heterogeneity using Landsat data based
bance (Kennedy et al., 2014). Spectral indices and Tasseled Cap Trans­ on a support vector regression (SVR) approach
formations (TCT) have long been used to estimate parameters related to ii) to generate a 30-year series of forest structural variables based on
vegetation recovery (Hislop et al., 2018; Pickell et al., 2016; White et al., the temporal extrapolation of LiDAR based estimates using
2017). More recently, time-series analysis has become increasingly Landsat imagery
important to characterise post-fire vegetation recovery, taking advan­ iii) to characterize post-fire forest structure recovery from the
tage of the development of change detection algorithms (Zhu, 2017) and generated time-series of structure variables
cloud computing capabilities through the Google Earth Platform (Gor­
elick et al., 2017). Particularly, the use of recovery trajectories from 2. Study area
time-series analysis provides essential information to identify abrupt
disturbances and subsequent recovery processes (Bright et al., 2019; The study site is located in the Mediterranean semi-arid region from
Frazier et al., 2015; Nguyen et al., 2018; Viana-Soto et al., 2020). southeastern Spain (Fig. 1). The Yeste Fire occurred in the summer of
Notwithstanding, changes in forest structure in combination to the 1994 (7th August), which burned 11,685 ha of pine dominated. Average
spectral signal of recovery must be considered for a comprehensive annual rainfall is 600–700 mm and the average annual temperature is
evaluation of forest resilience to fire (Bartels et al., 2016; Kennedy et al., around 15 ◦ C, subjected to soil hydrological deficit in summer
2014). LiDAR sensors are capable of penetrating through vegetation and (Ninyerola et al., 2005). Pinus halepensis Mill. and Pinus pinaster Ait.
recording forest structural characteristics (Lefsky et al., 2002). Previous dominate in those relatively young semi-natural forests resulting from
studies have successfully used LiDAR data for estimating forest attri­ plantations carried out in the mid-20th century (Pausas et al., 2004).
butes (Bottalico et al., 2017; Zhao et al., 2011), quantify post- Both species are obligate seeders, requiring a minimum of 15 years and
disturbance structural characteristics (Bolton et al., 2015; McCarley up to 30 years to reach reproductive maturity. The understory is mainly
et al., 2017), fire severity assessment (García et al., 2020; Montealegre composed by shrub species such as Thymus L., Rosmarinus L. or Juniperus
et al., 2014) and vegetation recovery analysis (Gordon et al., 2017; L., but Quercus ilex L. occurs in the understory at some areas. Fire
Martín-Alcón et al., 2015). The use of low-density LiDAR data has also recurrence in the northwest sector in 2017 (27th July, with approxi­
proved its ability to estimate forest attributes in Mediterranean forests mately 3,000 ha burned) makes this site of particular interest. Patches
(Gelabert et al., 2020; Tijerín et al., 2022), providing a suitable repre­ dominated by P. halepensis and P. pinaster in the overstory were selected
sentation of the post-fire forest conditions attending to both vertical according to the Forest Map of Spain (MFE) (1:50.000), including those
structure and horizontal continuity of vegetation (Kane et al., 2010). forested pixels that were burned as well as unburned from a buffer of 2
Estimates of vegetation cover and height as well as structural hetero­ km from the fire perimeter, which represented the natural dynamics of
geneity provide valuable information on how the dominance of shrub the ecosystem. Other cover types present in the scene were masked.
and tree strata is shifting within the recovery process (Bartels et al.,
2016). 3. Data and methods
Considering that airborne LiDAR data availability is limited in space
and time, combining it with Landsat imagery to temporally and spatially The main methodological steps are shown in Fig. 2. Firstly, we
extrapolate LiDAR-derived structural variables emerges as a promising computed a set of forest structure variables from LiDAR data for two
approach to model post-fire recovery regarding forest structure (Ahmed different dates, 2009 and 2016. Then, we extracted a series of predictor
et al., 2015; García et al., 2017; Matasci et al., 2018; Pflugmacher et al., variables (Landsat spectral bands, indices and transformations, and
2014; Senf et al., 2019; Zald et al., 2016). The lack of a direct rela­ topographic variables) to model the LiDAR-derived structure variables
tionship between structure and reflectance can be solved by non- from a SVR approach. The SVR model was calibrated and applied to a
parametric machine learning methods (Ahmed et al., 2014; García 30-year series of Landsat imagery to generate a time-series of forest
et al., 2017, 2011). In fact, the use of machine learning algorithms for structure attributes. We validated the estimations both temporally and
combining multisource remotely sensed data have greatly increased due spatially from an independent sample of LiDAR-derived estimates.
to their ability to model complex relationships between the dependent
and independent variables and no assumption of data distribution (Lary 3.1. Data acquisition and processing
et al., 2016), even in cases of high dimensionality and relatively small
sample sizes. A diversity of machine learning regression algorithms has 3.1.1. LiDAR data
been used to address the extrapolation procedure, such as Random Airborne Laser Scanner (ALS) data were acquired in 2009
Forests (Martín-Alcón et al., 2015; Senf et al., 2019), k-Nearest Neigh­ (September - November) and 2016 (August - September) by the National
bour (Ahmed et al., 2015; Matasci et al., 2018) or Support Vector Plan for Aerial Orthophotography of Spain (PNOA) and distributed by
Regression (SVR) (García et al., 2017). From these, SVR has been proved the National Geographic Institute of Spain (IGN). Data were acquired
an effective tool to capture complex non-linear relationships between using small-footprint discrete-return airborne sensors (LEICA ALS50 and
the dependent and the explanatory variables, and thus enabling the ALS60), with a maximum scan angle of 50◦ , at a scanning frequency of
extrapolation of LiDAR-derived canopy cover and height based on 70 Hz (minimum of 40 Hz), and a minimum pulse frequency of 45 kHz.
Landsat images (García et al., 2017; Zhao et al., 2011). However, at­ The maximum flight height was 3000 m above ground level. The point
tempts to extrapolate over time are limited due to the difficulty to density of the study area was 0.5 pts/m2, with a vertical accuracy of
achieve a stability of the estimates among sensors and across different 0.30 m (2009) and 0.20 m (2016). Data processing and subsequent
environmental conditions (Matasci et al., 2018). Our understanding on statistical metrics computation were carried out using the U.S. Forest
how LiDAR-to-Landsat extrapolation will perform in high structurally Service’s FUSION software v3.80 (McGaughey, 2018).
heterogeneous and opened landscapes as the Mediterranean forests We calculated a set of forest structural variables from LiDAR data at
using LiDAR collected at low point densities remains incomplete. 30-m resolution to match the spatial resolution of the Landsat data:
In this current research, we aimed at combining LiDAR data and vegetation cover, tree cover, mean height, and coefficient of variation of
Landsat imagery to provide insights on how forest structure has recov­ height. Cover metrics were extracted directly from the cloud point using
ered after fire in Mediterranean pine forests over a 30-year period. The the “GridMetrics” command at a 30-m spatial resolution using all the
specific objectives were: returns (Morsdorf et al., 2006). Vegetation Cover (VC), defined as the
percentage of ground covered by the vertical projection of vegetation for
a given pixel (Jennings et al., 1999), was calculated as the ratio of

2
A. Viana-Soto et al. International Journal of Applied Earth Observation and Geoinformation 108 (2022) 102754

Fig. 1. Location of the study area: a) post-fire Landsat images (RGB composition: SWIR2, NIR, Blue) from August 1994 (left) and August 2017 (right); b) Examples of
Pinus forest from unburned and burned plots (photographs were taken in March 2021). (For interpretation of the references to colour in this figure legend, the reader
is referred to the web version of this article.)

vegetation returns above 0.3 m to all returns (Equation (1)). This height et al., 2016; García et al., 2011).
threshold was selected to avoid ground points in relation to the vertical Apart from the Landsat spectral bands, we computed a set of spectral
accuracy (Gelabert et al., 2020). Similarly, the Tree Cover (TC) was indices to encompass the diversity of vegetation characteristics: the
computed using a 2 m height threshold (Shi et al., 2018), which was Normalized Difference Vegetation Index (NDVI) (Tucker, 1979), the
enough to exclude shrubs and herbs. Enhanced Vegetation Index (EVI) (Huete et al., 2002), the Normalized
Difference Water Index (NWMI) (Gao, 1996), and the Normalised Burn
Returns above Hthreshold
Cover = × 100 (1) Ratio (NBR) (Key and Benson, 2006). Tasselled Cap Transformations
Total all returns
(Brightness, Greenness and Wetness) (Crist, 1985) were calculated using
Metrics of forest height were derived from the Canopy Height Model the coefficients derived by Baig et al. (2014) and then the Tasseled Cap
(CHM) constructed by subtracting the terrain heights (DEM) to the Angle (TCA) (Powell et al., 2010) and Tasseled Cap Distance (TCD)
points. The DEM was previously generated by interpolating the points (Duane et al., 2010) were computed from brightness and greenness
classified as ground. Canopy height for each 30-m cell was computed by components. TCA is sensitive to the gradient of vegetation cover (Gómez
averaging the contained pixels of the original CHM computed at 2-m et al., 2011), and TCD is related to vegetation composition and structure
spatial resolution. Subsequently, we calculated the coefficient of varia­ (Pflugmacher et al., 2014).
tion of height (CVH) by dividing the standard deviation of height
(STDH) to the mean height (MH) to assess the vertical structural 3.1.3. Topographic data
heterogeneity. Topographic variables were derived from the LiDAR-based Digital
Elevation Model (DEM) (25-m spatial resolution) from the National
3.1.2. Landsat data Geographic Institute (IGN) of Spain, as they are known to be related to
Annual Landsat composites for the period 1990–2020 were gener­ forest vegetation distribution: elevation, slope and transformed aspect
ated through the Google Earth Engine platform (Gorelick et al., 2017) (TRASP, Roberts and Cooper, 1989) (Equation (2)). The post-processing
using a medoid selection process choosing the pixel closest to the me­ of this DEM included the removal of possible artefacts and gap filling.
dian among images (Flood, 2013). Landsat TM, ETM + and OLI Surface
Reflectance images were selected according to a target period (31st TRASP =
1 − (cos(aspect− 30))
(2)
August − 15th October) to minimize the effect of phenological changes, 2
filtering by cloud cover (less than 20%). Inter-sensor harmonization was
where aspect is in degrees. TRASP ranges from 0 to 1: values of
carried out due to differences among the spectral characteristics of
0 correspond to cooler, wetter north-northeast aspects, whereas values
Landsat ETM + and OLI datasets (Vogeler et al., 2018). We normalised
of 1 correspond to hotter, dryer south-southwestern aspects.
Landsat TM and ETM + to Landsat OLI from a multilinear regression
approach (Roy et al., 2016). We applied the topographic correction SCS
+ C (Soenen et al., 2005) to remove the effect of the terrain slope. The 3.2. Predictor variables and sample selection
SCS + C is based on the Sun-Canopy-Sensor correction proposed by (Gu
and Gillespie, 1998), specifically designed for forested areas (Chance LiDAR-derived variables to describe structural recovery (Vegetation

3
A. Viana-Soto et al. International Journal of Applied Earth Observation and Geoinformation 108 (2022) 102754

Fig. 2. Workflow of the methodology for the modelling forest cover and height using Landsat imagery and LiDAR data.

Cover–VC-, Tree Cover-TC-, Mean Height –MH- and Coefficient of (training data) less than a value ε while being as flat as possible (Smola
Variation of Height –CVH-) were set as response variables, and Landsat and Schölkopf, 2004). SVR was selected for the extrapolation as it has
variables (bands and spectral indices), topography and position were set been demonstrated an effective tool to model forest canopy height
as predictors. A sample of 10,000 pixels was selected for the model (García et al., 2018; Pourshamsi et al., 2021) with a good generalization
training and validation by a stratified random sampling method. Out­ capability (Awad and Khanna, 2015). Input data were scaled to prevent
liers were previously removed to provide meaningful training and variables of higher numerical ranges dominating variables of lower
validation pixels considering 3 standard deviations from the mean value numerical ranges. We used the radial basis function kernel and per­
of the target variable. We stratified the dataset depending on the formed a grid search using 5-fold cross validation for the parameter
response variable histogram distribution to ensure all strata were rep­ tuning (Xu and Goodacre, 2018).
resented in the train-test and validation datasets. Mean height was
stratified in equal intervals of 0.5 m in a range of 0 to 15 m while cover 3.4. Model evaluation and performance
variables were stratified in equal intervals of 5% in a range of 0 to 100
%. We then carried out a spatial random selection within each strata The performance of the SVR models was evaluated both temporally
avoiding forested edge pixels considering a distance of 100 m from the and spatially. The spatial assessment was conducted using the reserved
borders. The 2016 data sample was split into 60% for calibration of the validation sample from the reference year (2016) (4,000 samples), and
model (i.e. 6,000 pixels) and 40% for independent validation. In addi­ the temporal assessment using a sample from the independent LiDAR
tion, to evaluate the temporal robustness of the model, we extracted an acquisition collected in 2009 (4,000 samples). Models were assessed by
independent sample from the 2009 dataset for the temporal validation, comparing predicted to observed values using a series of goodness of fit
with a size equal to the sample used for the spatial validation. measures, including the coefficient of determination R2, mean absolute
error (MAE), root mean squared error (RMSE), relative root mean square
error (rRMSE) and bias (i.e., average of predicted minus observed
3.3. Support vector regression modelling
values) for each response variable.
We calibrated a Support Vector Regression model (SVR) (Smola and 1 ∑N

Schölkopf, 2004) to estimate forest cover and height from predictor MAE = |Predictedi − Observed i | (3)
N i=1
variables using the module of Scikit-learn in Python (Pedregosa et al.,
2011). SVR is a supervised learning algorithm designed to solve √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√
regression problems, which can deal with non -linearity by transforming √1 ∑ N
RMSE = √ (Predictedi − Observedi )2 (4)
the data into higher-dimensional space using a kernel function. SVR N i=1
attempts to find an optimal loss function deviating from the target

4
A. Viana-Soto et al. International Journal of Applied Earth Observation and Geoinformation 108 (2022) 102754

RMSE
rRMSE = 100 (5)
MeanObservedi

3.5. Characterisation of post-fire structural recovery

Post-fire recovery was quantified using a relative indicate of recovery

within each of the estimated structural variables (VC, TC, MH and CVH).
Similarly to Kennedy et al. (2012) and White et al. (2017), the Recovery
Indicator (RI) was defined as the ratio of the post-fire recovery (Δ Re­
covery) by the magnitude of the fire (Δ Disturbance) (Fig. 3):
ΔRecovery
RI (VC,TC,MH,CVH) = (6)
ΔDisturbance

where Δ Recovery is the change in the forest structural variable value at

a defined year following fire:
ΔRecovery(VC,TC,MH,CVH) = Valuey+i − Valuey (7)

and Δ Disturbance is defined as the difference in the forest structural

variable value from the year before the fire and the year of the fire:
ΔDisturbance(VC,TC,MH,CVH) = Valuey− 1 − Valuey (8)

Recovery Indicator (RI) was calculated at different successional

stages to assess post-fire structural conditions immediately after fire (y
+ 1, 1995), in the short term (y + 6, 2000), mid-term (y + 11 and y + 16,
− 2005 and 2010-), and long-term (y + 21 and y + 26, − 2015 and 2020-
).

4. Results

4.1. Model assessment

Comparing forest structure variables estimated from Landsat to the

reference dataset of LiDAR estimates, we found the overall best agree­
ment for VC, followed by MH, TC and CVH (Fig. 4). Models explain a
good percentage of the variance, with model R2 ranging from 0.63 to
0.78, and rRMSE below 30% for all variables except for the mean height
(2016). Models from the temporal extrapolation (2009) maintain a good
performance, although R2 decreases (0.51–0.74) and the errors increase
(rRMSE ranging from 31 to 44%). VC showed the highest agreement
close to the 1:1 line with a MAE of 9.6% for 2016 and 11.1% for 2009.
The agreement of TC estimates was lower (MAE of 11.3% for 2016 and
12.6% for 2009) displaying an over- and underestimation in the lowest
and highest TC values, respectively. Mean height (MH) estimations
showed moderate agreement with a MAE of 1.25 m (2016) and 1.5 m
(2009) while Coefficient of Variation of height (CVH) estimations ach­ Fig. 4. Scatterplots between predicted and observed values for canopy cover
ieved a less strong correlation to the reference dataset. Positive bias in (VC and TC) and height (MH and CVH) response variables on the independent
2009 indicates that the model generally overestimated cover and height validation samples (2016 and 2009).
values, while we observed an underestimation of cover and height
values in 2016.

4.2. Estimations of annual forest cover and height

Fig. 5 contains the time-series of VC, TC, MH and CVH variables for
the burned pixels in 1994, burned pixels in both 1994 and 2017, and
unburned forested pixels for the period 1990–2020. In those unburned
areas, forest cover and height variables remained stable over the three
decades, with a slight upward trend. Time-series of burned areas shows a
loss of cover and height in the year of the fire that even extends into the
following post-fire year. Most noticeable is that burned areas did not
return to its pre-fire value in terms of Tree Cover and Mean Height
within the temporal window of the time-series (26 years post-fire),
although they remain close. Those patches with fire recurrence in
Fig. 3. Schema of recovery metrics: Δ Recovery, Δ Disturbance, pre-fire value 2017 were burned before reaching the previous conditions, returning to
and Recovery Indicator (RI). a starting point of the recovery process. Attending to the evolution of the

5
A. Viana-Soto et al. International Journal of Applied Earth Observation and Geoinformation 108 (2022) 102754

Fig. 5. Time-series of mean values of forest structure variables for burned and unburned pixels. Shading around the mean values indicate the standard deviation.
Vertical red lines indicate the year of the fire. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of
this article.)

CVH, the variability increases after the fire at the pixel level due to the values with a slight increase.
presence of a mixture of soil, remnant vegetation and dead wood. In the
following years after the fire, variability tends to be reduced by the
regrowth of vegetation. Nonetheless, given the high heterogeneity of 4.3. Quantifying post-fire structural recovery
forests, even in unburned areas, the CVH remained above 0.5 m.
These temporal patterns are also observed in the detailed maps of the From a spatio-temporal perspective, we examined the annual maps
estimated variables for the years 1990, 1995, 2005 and 2020 (Fig. 6). to get a detailed wall-to wall analysis of the forest structure recovery
Vegetation cover, tree cover and mean height values are clearly more from the fire occurrence until 2020. Fig. 7 shows the difference between
homogeneous throughout the area before the fire occurs (1990) with 2020 and 1994 (immediately after fire) of the estimated forest structural
intermediate values of Coefficient of Variation of height. One year after attributes over burned and unburned forested areas. Vegetation cover,
fire (1995), vegetation cover, tree cover and mean height values drop tree cover and mean height remained relatively stable in unburned
sharply, increasing the CVH. The heterogeneity of the landscape in terms forests with a slight increase over the period (10–20% in cover, 2–4 m in
of cover and height is well visible in those burned areas in the mid-term height). CVH also showed small changes in magnitude (-0.2), associated
and long-term (2005, 2020) while the unburned sector maintains stable to the natural successional process of the forest.
Burned areas display different percentages of increases in vegetation

6
A. Viana-Soto et al. International Journal of Applied Earth Observation and Geoinformation 108 (2022) 102754

Fig. 6. Maps of Vegetation Cover, Tree Cover, Mean Height and Coefficient of Variation of Height at four successional stages: 1990 (y-4), 1995 (y + 1), 2005 (y + 11)
and 2020 (y + 26). White spaces indicate masked areas.

cover but also patches with non-recovery in tree cover and mean height. the area burned in 2017 that was not disturbed in 1994. The area sub­
Largest increases occur in terms of vegetation cover indicating a quick jected to fire-recurrence, meaning burned in 1994 and 2017, show low
recovery of vegetation in the horizontal plane, whereas tree cover shows percentages of vegetation cover and mean height 3 years after the sec­
considerably lower values. Burned areas also show positive increases in ond fire. The variability of height decreased with respect to the year of
mean height values throughout the study area, with higher regrowth the fire as the recovery process moves forward and vegetation starts to
along on north-facing valley slopes. Most negative values correspond to encroach the ground.

7
A. Viana-Soto et al. International Journal of Applied Earth Observation and Geoinformation 108 (2022) 102754

Fig. 7. Difference (2020–1994) maps of post-fire Vegetation Cover, Tree Cover, Mean Height and Coefficient of Variation of Height. White spaces indicate
masked areas.

Results from the Recovery Indicator (RI) of the forest structure var­ A quantitative overview of the recovery at the successional stages is
iables at 5-yearly intervals (Fig. 8) provide a measure of the immediately provided by the percentage of pixels that recovered the pre-fire values
post-fire conditions (y + 1, 1995) and the state of recovery in the short regarding the four forest structure variables at 5-year intervals. We
term (y + 6, 2000), in the mid-term (y + 11-y + 16, 2005–2010) and, in considered two scenarios to account for a completely recovery
the long term (y + 21-y + 26, 2015–2020). The evolution of the RI shows (threshold of 100%) (Fig. 9a) and for an advanced state of the recovery
a gradual recovery for all variables, although only vegetation cover (threshold of 80%) (Fig. 9b). Considering a recovery threshold of 100%,
reaches pre-fire values 26 years after fire. RI values reached the less than 10% of the pixels completely recovered to its pre-fire structure
threshold of 0.5 (i.e. half of pre-fire state) in terms of vegetation cover in before 6 years after fire and around 15–20% recovered in the mid-term
the short term, but remain below 0.5 according to the mean height, (16 years post-fire). Approximately 40% of pixels recovered the pre-fire
indicating an initial recovery of pioneer herbs and shrubs and an initial values in terms of tree cover 26 years after fire, while less than 30% of
reestablishment of remnant trees. Over the mid-term to long-term (from pixels were recovered attending to mean height. Nevertheless, consid­
2010 onwards), tree cover almost recovered to pre-fire values, while ering a recovery threshold of 80% from the pre-fire values, the amount
mean height continues to increase at a slower pace. CVH shows the of recovered pixels increase to 73 % on vegetation cover, 58.6 % on tree
sharpest increase in the short-term and stabilizes thereafter. cover, and 43.8% on mean height 26 years after fire, meaning a

8
A. Viana-Soto et al. International Journal of Applied Earth Observation and Geoinformation 108 (2022) 102754

Fig. 8. Violin plots and boxplots of the Recovery Indicator for each forest structure variable. Boxes indicate the 25, 50 and 75 percentiles, and the whiskers show
confidence intervals (95%). Vertical line indicates a full recovery (RI = 1).

Fig. 9. Statistics of recovered pixels from burned area in 1994 at a) 100% and b) 80% of recovery attending to the four structure variables in 1995 (y + 1), 2000 (y +
6), 2005 (y + 11), 2010 (y + 16), 2015 (y + 21) and 2020 (y + 26).

relatively advanced state of the recovery. unique opportunity to extrapolate information on forest structure over
time for a comprehensive evaluation of forest resilience to fire. Here we
5. Discussion extrapolated LiDAR-derived cover and height variables over a 30-year
window using Landsat imagery and topography variables through a
5.1. Performance of Landsat-based forest structure extrapolation SVR model. Our results confirm previous studies in boreal and temperate
forests showing that Landsat data are well suited for mapping LiDAR-
Post-fire forest recovery dynamics are highly complex due to their derived forest cover and height (Ahmed et al., 2015; Matasci et al.,
high variability and processes undergone in changing environmental 2018; Senf et al., 2019; Zald et al., 2016). We assessed the performance
conditions. LiDAR data in combination with Landsat imagery provides a of the models both spatially and temporally, thereby ensuring their

9
A. Viana-Soto et al. International Journal of Applied Earth Observation and Geoinformation 108 (2022) 102754

capability to extrapolate LiDAR-based information on forest variables cover mean. The relatively good temporal performance of the models is
(Matasci et al., 2018; Zald et al., 2016). crucial for the generation of time series of structural variables that
The models performed better for the cover estimations, as cover enable the assessment of vegetation recovery. Differences in the vege­
metrics are assumed to be more directly related to the spectral response tative conditions of ecosystems require the model to adapt to these
of the pixel (García et al., 2017; Zhao et al., 2011), resulting in RMSE of differences to accurately capture forest structure changes over the time-
12.8% (2016) and 14.3% (2009) for VC, and RMSE of 14.8% (2016) and series (Matasci et al., 2018). Moreover, the temporal validation between
16.5% (2009) for TC. In this way, Matasci et al. (2018) better estimated two years of Landsat data from different sensors (TM in 2009 and OLI in
annual forest cover than forest height by combining Landsat composites 2016) ensures a temporal stability of the estimates despite the change in
and LiDAR plots using a nearest neighbour imputation approach. Out of sensor, making it possible to extend the series in the future with the
the cover variables, they also obtained higher errors for the cover above ongoing Landsat 8 data acquisitions and the recently launch of Landsat 9
2 m (RMSE ranging from 13.2% to 18.6%) than for the cover mean (9 to with similar characteristics (Masek et al., 2020).
11.7%). Limitations from optical data to characterise vertical structure The possibility to validate models both temporally and spatially is
are well-known in relation to the decrease of the sensor sensitivity in crucial to obtain robust estimates (Matasci et al., 2018). Model perfor­
closed canopy conditions (Goetz and Dubayah, 2011; Pflugmacher et al., mance could improve with the availability of more LiDAR datasets at
2014). Even so, the models yielded a MAE below 1.5 m for mean height different dates throughout the time series. This methodology is appro­
estimates in both spatial and temporal validation (RMSE of 1.67 and priate for regional and national analysis, using LiDAR coverages
1.93 m, respectively). Errors found here are in the same order of collected at low point densities. The annual maps of forest structure
magnitude as those reported by Zald et al. (2016) (RMSE of 1.84 m in variables could be considered a starting point for the analysis of post-fire
mean height) across taiga and boreal forests in Central Canada. forest dynamics at larger scales. The recent accessibility to LiDAR ac­
Compared to other research, our SVR model outperformed linear quisitions from the near-global mission GEDI (Dubayah et al., 2020)
regression approaches according to the results obtained by Pascual et al. makes possible the integration with Landsat imagery to generate infor­
(2010) in central Spain (RMSE of 1.9–2.3 m) from a unique date of mation on forest attributes for larger area applications (Potapov et al.,
LiDAR data collected at higher point density (5pts/m2). In the same line, 2021).
Ahmed et al. (2015) reported varying errors in canopy height estimates
according to forest maturity (RMSE ranging from 2.4 and 3.5 m) using 5.2. Characterization of forest structure recovery dynamics
Random Forests algorithm in boreal forests from West Canada. In
accordance to our findings, the coefficient of variation of heights is the We examined the evolution of post-fire recovery attending to the
variable estimated with the lowest precision in both Zald et al. (2016) LiDAR-derived forest cover, height and heterogeneity attributes to get a
and Matasci et al. (2018) due to the limitations in capturing vertical broad overview of forest structure dynamics after fire. Vegetation cover
structure variability from the spectral signal. Similarly, Senf et al. (2019) and mean height displayed a different rhythm of recovery. Vegetation
in a study carried out in temperate forest from Central Europe found cover reached pre-fire values within 26-years on average and tree cover
better results for tree cover than for stand height estimations with a remains close to pre-fire values, which can be attributed to tree regrowth
RMSE in tree cover of 15% and a RMSE in stand height estimates up to 3 and a higher density of stems per hectare (Bolton et al., 2015; Kane et al.,
m when comparing to an independent sample from the same LiDAR 2014; Matasci et al., 2018). Although mean height showed a slower
acquisition. recovery, it reached values close to the pre-fire situation. Nonetheless, it
Limitations of Landsat indices related to saturation in closed can­ should be mentioned that the recovery of forest height might also be
opies and the influence of soil and understory reflectance in open forests underestimated, since small footprint LiDAR tend to underestimate the
may explain the slight underestimations of forest attributes observed in canopy height in coniferous forests, particularly when point density is
the higher mean height ranges in contrast to the overestimations low, because of the low probability of pulses hitting the tree tops
observed at the lower values (García et al., 2018; Potapov et al., 2021). A (Magnussen and Boudewyn, 1998).
further potential source of error can be attributed to the dynamics of Studies addressing temporal extrapolation of LiDAR-derived forest
vegetation succession itself, implying that the spectral signal of vege­ attributes for recovery quantification are limited. Comparing our find­
tation at 2 m in height (tree cover) within a pixel is a mixture of early- ings with those obtained by Matasci et al. (2018) in boreal forests, both
growing trees and shrubs in the initial stages of the recovery (Martín- agree on the marked differences in the evolution of cover with respect to
Alcón et al., 2015). Although uncertainties may also arise due to the height estimates, where tree cover reached pre-fire values earlier but
nature of the low-density LiDAR data itself, we can assume that LiDAR- canopy height did not recovered before 25 years after fire. Recovery
derived cover and mean vegetation height display similar trends to the estimates obtained from Landsat time-series in previous studies, were
natural vegetation dynamics (Falkowski et al., 2009; Kane et al., 2010). clearly overestimated, since estimations from spectral indices suggested
Another crucial factor is that the model accuracy relies on relation­ that vegetation reached the pre-fire state in the short and mid-term in
ships between response variables and predictor variables. This means to Mediterranean ecosystems (Hislop et al., 2018; Morresi et al., 2019;
assume that the dataset adequately contains the variables of interest and Viana-Soto et al., 2017). Although the results obtained here show that
represents the range of variation in these forest variables. The models after 26 years we have a forest in an advanced state of recovery in terms
here obtained show high values regarding the goodness of fit spatially of structure, our results evidenced that the variables used to characterize
achieving an R2 from 0.63 to 0.78. Previous studies that also attempted forest structure had not reached the pre-fire state. A temporal window of
to integrate LiDAR and Landsat to derive forest structure attributes 26 years can give a clear picture of whether there is a recovery process
confirmed that indices including SWIR bands (NBR, NDWI) but also TCT ongoing even if a longer period is needed to reach a mature forest
as TCA and TCW are closely related to LiDAR-derived structural vari­ (González-De Vega et al., 2016).
ables in different ecosystems (Ahmed et al., 2015; García et al., 2018; Many factors influence post-fire successional pathways (McLauchlan
Gómez et al., 2011; Pflugmacher et al., 2014). et al., 2020), depending not only on fire severity but also on the adap­
Temporally, the models also showed to be consistent when cali­ tations, surviving trees, canopy seed banks, changes in dominant species
brating the model in a different year (2016), although a wider variability associated with inter-species competition processes (Calvo et al., 2008;
was observed for the 2009 estimates, with model R2 ranging between Pausas and Keeley, 2014) and post-fire climatic conditions (Mazza and
0.51 and 0.74, and an increase in the errors (relative RMSE from 30 to Sarris, 2021). In these pine forests, the minimum interval commonly
44%). Errors reported here are in the same order of magnitude as those used for considering a forest as having recovered is above 15–20 years
found by Matasci et al. (2018) in the temporal assessment, with relative (Eugenio et al., 2006). This is connected to the fact that the juvenile
RMSE ranging from 38.3% to 46.6 % in mean height and 23% to 32% in reproductive phase is not reached until 20 years after fire in seed-species

10
A. Viana-Soto et al. International Journal of Applied Earth Observation and Geoinformation 108 (2022) 102754

(Thanos and Daskalakou, 2000), meaning that pine trees take several CRediT authorship contribution statement
years to establish from seed. Therefore, we expected a lower rhythm of
recovery in mean height in the short-term that slightly increases after Alba Viana-Soto: Conceptualization, Methodology, Formal analysis,
the first decade on average (Fig. 5). Writing – original draft. Mariano García: Conceptualization, Method­
Using the Recovery Indicator (RI) for the four structure variables, we ology, Writing – review & editing. Inmaculada Aguado: Conceptuali­
managed to monitor how the recovery has evolved at successive stages zation, Writing – review & editing. Javier Salas: Conceptualization,
(Fig. 8). Vegetation cover increased rapidly following fire indicating a Writing – review & editing.
quicker recovery of the canopy-filling gaps with the colonization of
pioneer herbs and shrubs and an initial reestablishment of remnant trees Declaration of Competing Interest
(Bolton et al., 2015; González-De Vega et al., 2016). On the contrary,
mean height remained further from reaching pre-fire values in the first The authors declare that they have no known competing financial
decade, but values gradually increased according to the progress from a interests or personal relationships that could have appeared to influence
stage of stand initiation to the stem exclusion, in which shrub dominance the work reported in this paper.
is replaced by greater tree dominance (Martín-Alcón et al., 2015).
The recovery scenarios evidenced that percentage of pixels that Acknowledgments
completely recovered the pre-fire values of cover and height does not
fulfil 50% within 26 years. Nevertheless, considering a recovery A.V.S. is supported by the Spanish Ministry of Science, Innovation
threshold of 80% of the pre-fire value (White et al., 2018), we observe and Universities through a FPU doctoral fellowship (FPU17/03260).
higher percentages of recovered pixels (70% in vegetation cover and Department support: Grant for Excellence in Teaching Staff of the
44% in mean height). With these results, we can consider that resilience Community of Madrid (EPU-DPTO/2020/008).
to fire of these Mediterranean pine forests is moderate. The challenge
here is, as suggested by Moya et al. (2018) and Taboada et al. (2018), the
References
risk involved if the ecosystem remains in an immature state in the face of
a likely recurrence of fire in shorter intervals than these pine species Ahmed, O.S., Franklin, S.E., Wulder, M.A., 2014. Integration of lidar and Landsat data to
need to develop a productive seed bank. Combining the results of the estimate forest canopy cover in coastal British Columbia. Photogramm. Eng. Remote
Sens. 80, 953–962. https://doi.org/10.14358/PERS.80.10.953.
time series (Fig. 5) and the Recovery Indicator (Fig. 8), these suggest the
Ahmed, O.S., Franklin, S.E., Wulder, M.A., White, J.C., 2015. Characterizing stand-level
presence of patches that remain in ongoing recovery although they have forest canopy cover and height using Landsat time series, samples of airborne LiDAR,
not reached the pre-fire benchmark. In this way, a new fire may limit and the Random Forest algorithm. ISPRS J. Photogramm. Remote Sens. 101, 89–101.
recovery by producing ecosystem degradation involving changes in https://doi.org/10.1016/j.isprsjprs.2014.11.007.
Awad, M., Khanna, R., 2015. Support Vector Regression. In: Efficient Learning Machines.
ecosystem composition (Baudena et al., 2019). Analysing the forest Apress, Berkeley, CA, pp. 67–80. https://doi.org/10.1007/978-1-4302-5990-9_4.
structure values obtained for the first 3 years after the 2017 fire in the Baig, M.H.A., Zhang, L., Shuai, T., Tong, Q., 2014. Derivation of a tasselled cap
area with fire recurrence (Fig. 5), signals of difficulty in initiating re­ transformation based on Landsat 8 at-satellite reflectance. Remote Sens. Lett. 5,
423–431. https://doi.org/10.1080/2150704X.2014.915434.
covery after fire recurrence can be gathered from the time-series of cover Bartels, S.F., Chen, H.Y.H., Wulder, M.A., White, J.C., 2016. Trends in post-disturbance
and height estimates, whereas this initial recovery was observed in the recovery rates of Canada’s forests following wildfire and harvest. For. Ecol. Manage.
1994 fire. 361, 194–207. https://doi.org/10.1016/j.foreco.2015.11.015.
Baudena, M., Santana, V.M., Baeza, M.J., Bautista, S., Eppinga, M.B., Hemerik, L., Garcia
Some methodological considerations should be considered when Mayor, A., Rodriguez, F., Valdecantos, A., Vallejo, V.R., Vasques, A., Rietkerk, M.,
explaining ecological processes from the findings obtained here. Our 2020. Increased aridity drives post-fire recovery of Mediterranean forests towards
approach provides an overview of the evolution of the forest structure open shrublands. New Phytologist 225 (4), 1500–1515. https://doi.org/10.1111/
nph.16252.
after a large forest fire in concordance with the potential and limitations
Bolton, D.K., Coops, N.C., Wulder, M.A., 2015. Characterizing residual structure and
of the available low-resolution LiDAR data (Martín-Alcón et al., 2015). forest recovery following high-severity fire in the western boreal of Canada using
The use of Landsat data provides a unique opportunity to analyse the Landsat time-series and airborne lidar data. Remote Sens. Environ. 163, 48–60.
https://doi.org/10.1016/j.rse.2015.03.004.
evolution through decades, but the 30-m pixel size may conceal a larger
Bottalico, F., Chirici, G., Giannini, R., Mele, S., Mura, M., Puxeddu, M., McRoberts, R.E.,
structural variability within. Even so, cover and height are important Valbuena, R., Travaglini, D., 2017. Modeling Mediterranean forest structure using
indicators of forest recovery that can be derived to larger areas, airborne laser scanning data. Int. J. Appl. Earth Obs. Geoinf. 57, 145–153. https://
providing useful information to support post-fire restoration activities doi.org/10.1016/j.jag.2016.12.013.
Bowman, D.M.J.S., Williamson, G.J., Abatzoglou, J.T., Kolden, C.A., Cochrane, M.A.,
(de las Heras et al., 2012). Smith, A.M.S., 2017. Human exposure and sensitivity to globally extreme wildfire
events. Nat. Ecol. Evol. 1, 1–6. https://doi.org/10.1038/s41559-016-0058.
6. Conclusion Bright, B.C., Hudak, A.T., Kennedy, R.E., Braaten, J.D., Henareh Khalyani, A., 2019.
Examining post-fire vegetation recovery with Landsat time series analysis in three
western North American forest types. Fire Ecol. 15 https://doi.org/10.1186/s42408-
The feasibility of extrapolating LiDAR-derived structural variables 018-0021-9.
using Landsat images over time has been demonstrated. The stability of Calvo, L., Santalla, S., Valbuena, L., Marcos, E., Tárrega, R., Luis-Calabuig, E., 2008. Post-
fire natural regeneration of a Pinus pinaster forest in NW Spain. Plant Ecol. 197,
the estimations both temporal and spatially enabled to analyse differ­ 81–90. https://doi.org/10.1007/s11258-007-9362-1.
ences in post-fire forest structure recovery. Noticeable differences were Chance, C.M., Hermosilla, T., Coops, N.C., Wulder, M.A., White, J.C., 2016. Effect of
found in terms of vegetation cover and mean height, as the cover values topographic correction on forest change detection using spectral trend analysis of
Landsat pixel-based composites. Int. J. Appl. Earth Obs. Geoinf. 44, 186–194.
recovered to a pre-fire state but mean height did not reach this bench­
https://doi.org/10.1016/j.jag.2015.09.003.
mark along the temporal window analysed. With less than 50% of Chuvieco, E., Aguado, I., Salas, J., García, M., Yebra, M., Oliva, P., 2020. Satellite Remote
burned pixels completely recovered to a pre-fire structure 26 years after Sensing Contributions to Wildland Fire Science and Management. Curr. For. Reports
6, 81–96. https://doi.org/10.1007/s40725-020-00116-5.
fire, we may conclude that the resilience to fire of these Mediterranean
Crist, E.P., 1985. A TM Tasseled Cap equivalent transformation for reflectance factor
pine forests is moderate. Providing data on the dynamics of forest data. Remote Sens. Environ. 17, 301–306. https://doi.org/10.1016/0034-4257(85)
structure response is a valuable source of information to stakeholders 90102-6.
and decision makers concerned with the assessment of post-fire forest De las Heras, J., Moya, D., Vega, J.A., Daskalakou, E., Vallejo, R., Grigoriadis, N.,
Tsitsoni, T., Baeza, J., Valdecantos, A., Fernández, C., Espelta, J., Fernandes, P.,
recovery. Our work provides a first step in investigating the temporal 2012. Post-Fire Management of Serotinous Pine Forests, in: Moreira, F., Arianoutsou,
evolution of the Mediterranean pine forests structure over three de­ M., Corona, P., de las Heras, J. (Eds.), Post-Fire Management and Restoration of
cades, which can support the design of mitigation and adaptation stra­ Southern European Forests. Springer, pp. 121–149. https://doi.org/10.1007/978-
94-007-2208-8_6.
tegies that enhance fire resilience. Duane, M.V., Cohen, W.B., Campbell, J.L., Hudiburg, T., Turner, D.P., Weyermann, D.L.,
2010. Implications of alternative field-sampling designs on landsat-based mapping of

11
A. Viana-Soto et al. International Journal of Applied Earth Observation and Geoinformation 108 (2022) 102754

stand age and carbon stocks in Oregon forests. For. Sci. 56, 405–416. https://doi. national park. Remote Sens. Environ. 151, 89–101. https://doi.org/10.1016/j.
org/10.1093/forestscience/56.4.405. rse.2013.07.041.
Dubayah, R., Blair, J.B., Goetz, S., Fatoyinbo, L., Hansen, M., Healey, S., Hofton, M., Keeley, J.E., Bond, W.J., Bradstock, R.A., Pausas, J.G., Rundel, P.W., 2012. Fire in
Hurtt, G., Kellner, J., Luthcke, S., Armston, J., Tang, H., Duncanson, L., Hancock, S., mediterranean ecosystems: Ecology, evolution and management. Cambridge
Jantz, P., Marselis, S., Patterson, P.L., Qi, W., Silva, C., 2020. The Global Ecosystem University Press, Cambridge, 10.1017/CBO9781139033091.
Dynamics Investigation: High-resolution laser ranging of the Earth’s forests and Kennedy, R.E., Andréfouët, S., Cohen, W.B., Gómez, C., Griffiths, P., Hais, M., Healey, S.
topography. Sci. Remote Sens. 1, 100002 https://doi.org/10.1016/j. P., Helmer, E.H., Hostert, P., Lyons, M.B., Meigs, G.W., Pflugmacher, D., Phinn, S.R.,
srs.2020.100002. Powell, S.L., Scarth, P., Sen, S., Schroeder, T.A., Schneider, A., Sonnenschein, R.,
Eugenio, M., Verkaik, I., Lloret, F., Espelta, J.M., 2006. Recruitment and growth decline Vogelmann, J.E., Wulder, M.A., Zhu, Z., 2014. Bringing an ecological view of change
in Pinus halepensis populations after recurrent wildfires in Catalonia (NE Iberian to landsat-based remote sensing. Front. Ecol. Environ. 12, 339–346. https://doi.org/
Peninsula). For. Ecol. Manage. 231, 47–54. https://doi.org/10.1016/j. 10.1890/130066.
foreco.2006.05.007. Kennedy, R.E., Yang, Z., Cohen, W.B., Pfaff, E., Braaten, J., Nelson, P., 2012. Spatial and
Falkowski, M.J., Evans, J.S., Martinuzzi, S., Gessler, P.E., Hudak, A.T., 2009. temporal patterns of forest disturbance and regrowth within the area of the
Characterizing forest succession with lidar data: An evaluation for the Inland Northwest Forest Plan. Remote Sens. Environ. 122, 117–133. https://doi.org/
Northwest, USA. Remote Sens. Environ. 113, 946–956. https://doi.org/10.1016/j. 10.1016/j.rse.2011.09.024.
rse.2009.01.003. Key, C.H., Benson, N.C., 2006. Landscape assessment: Remote sensing of severity, the
Flood, N., 2013. Seasonal composite landsat TM/ETM+ Images using the medoid (a Normalized Burn Ratio. FIREMON Fire Eff Monit. Invent. Syst. Gen. Tech. Report
multi-dimensional median). Remote Sens. 5, 6481–6500. https://doi.org/10.3390/ RMRS-GTR-164-CD 305–325. https://doi.org/10.1002/app.1994.070541203.
rs5126481. Lary, D.J., Alavi, A.H., Gandomi, A.H., Walker, A.L., 2016. Machine learning in
Frazier, R.J., Coops, N.C., Wulder, M.A., 2015. Boreal Shield forest disturbance and geosciences and remote sensing. Geosci. Front. 7, 3–10. https://doi.org/10.1016/j.
recovery trends using Landsat time series. Remote Sens. Environ. 170, 317–327. gsf.2015.07.003.
https://doi.org/10.1016/j.rse.2015.09.015. Lefsky, M.A., Cohen, W.B., Parker, G.G., Harding, D.J., 2002. Lidar remote sensing for
Gao, B.-C., 1996. NDWI A Normalized Difference Water Index for Remote Sensing of ecosystem studies. Bioscience 52, 19–30. https://doi.org/10.1641/0006-3568(2002)
Vegetation Liquid Water From Space. Remote Sens. Environ. 58, 257–266, 10.1016/ 052[0019:LRSFES]2.0.CO;2.
S0034-4257(96)00067-3. Magnussen, S., Boudewyn, P., 1998. Derivations of stand heights from airborne laser
García, M., North, P., Viana-Soto, A., Stavros, N.E., Rosette, J., Martín, M.P., scanner data with canopy-based quantile estimators. Can. J. For. Res. 28,
Franquesa, M., González-Cascón, R., Riaño, D., Becerra, J., Zhao, K., 2020. 1016–1031. https://doi.org/10.1139/x98-078.
Evaluating the potential of LiDAR data for fire damage assessment: A radiative Martín-Alcón, S., Coll, L., De Cáceres, M., Guitart, L., Cabré, M., Just, A., González-
transfer model approach. Remote Sens. Environ. 247, 111893 https://doi.org/ Olabarría, J.R., 2015. Combining aerial LiDAR and multispectral imagery to assess
10.1016/j.rse.2020.111893. postfire regeneration types in a Mediterranean forest. Can. J. For. Res. 45, 856–866.
García, M., Riaño, D., Chuvieco, E., Salas, J., Danson, F.M., 2011. Multispectral and https://doi.org/10.1139/cjfr-2014-0430.
LiDAR data fusion for fuel type mapping using Support Vector Machine and decision Masek, J.G., Wulder, M.A., Markham, B., McCorkel, J., Crawford, C.J., Storey, J.,
rules. Remote Sens. Environ. 115, 1369–1379. https://doi.org/10.1016/j. Jenstrom, D.T., 2020. Landsat 9: Empowering open science and applications through
rse.2011.01.017. continuity. Remote Sens. Environ. 248, 111968 https://doi.org/10.1016/j.
García, M., Saatchi, S., Casas, A., Koltunov, A., Ustin, S.L., Ramirez, C., Balzter, H., 2017. rse.2020.111968.
Extrapolating forest canopy fuel properties in the California Rim fire by combining Matasci, G., Hermosilla, T., Wulder, M.A., White, J.C., Coops, N.C., Hobart, G.W.,
airborne LiDAR and landsat OLI data. Remote Sens. 9, 1–18. https://doi.org/ Bolton, D.K., Tompalski, P., Bater, C.W., 2018. Three decades of forest structural
10.3390/rs9040394. dynamics over Canada’s forested ecosystems using Landsat time-series and lidar
García, M., Saatchi, S., Ustin, S., Balzter, H., 2018. Modelling forest canopy height by plots. Remote Sens. Environ. 216, 697–714. https://doi.org/10.1016/j.
integrating airborne LiDAR samples with satellite Radar and multispectral imagery. rse.2018.07.024.
Int. J. Appl. Earth Obs. Geoinf. 66, 159–173. https://doi.org/10.1016/j. Mazza, G., Sarris, D., 2021. Identifying the full spectrum of climatic signals controlling a
jag.2017.11.017. tree species’ growth and adaptation to climate change. Ecol. Indic. 130, 108109
Gelabert, P.J., Montealegre, A.L., Lamelas, M.T., Domingo, D., 2020. Forest structural https://doi.org/10.1016/j.ecolind.2021.108109.
diversity characterization in Mediterranean landscapes affected by fires using McCarley, T.R., Kolden, C.A., Vaillant, N.M., Hudak, A.T., Smith, A.M.S., Wing, B.M.,
Airborne Laser Scanning data. GIScience Remote Sens. 57, 497–509. https://doi.org/ Kellogg, B.S., Kreitler, J., 2017. Multi-temporal LiDAR and Landsat quantification of
10.1080/15481603.2020.1738060. fire-induced changes to forest structure. Remote Sens. Environ. 191, 419–432.
Goetz, S., Dubayah, R., 2011. Advances in remote sensing technology and implications https://doi.org/10.1016/j.rse.2016.12.022.
for measuring and monitoring forest carbon stocks and change. Carbon Manag. 2, McGaughey, R.J., 2018. FUSION/LDV: Software for LIDAR Data Analysis and
231–244. https://doi.org/10.4155/cmt.11.18. Visualization - V3.80. U.S. Department of Agriculture, Forest Service, Pacific
Gómez, C., White, J.C., Wulder, M.A., 2011. Characterizing the state and processes of Northwest Research Station, Seattle.
change in a dynamic forest environment using hierarchical spatio-temporal McLauchlan, K.K., Higuera, P.E., Miesel, J., Rogers, B.M., Schweitzer, J., Shuman, J.K.,
segmentation. Remote Sens. Environ. 115, 1665–1679. https://doi.org/10.1016/j. Tepley, A.J., Varner, J.M., Veblen, T.T., Adalsteinsson, S.A., Balch, J.K., Baker, P.,
rse.2011.02.025. Batllori, E., Bigio, E., Brando, P., Cattau, M., Chipman, M.L., Coen, J., Crandall, R.,
González-De Vega, S., De las Heras, J., Moya, D., 2016. Resilience of Mediterranean Daniels, L., Enright, N., Gross, W.S., Harvey, B.J., Hatten, J.A., Hermann, S.,
terrestrial ecosystems and fire severity in semiarid areas: Responses of Aleppo pine Hewitt, R.E., Kobziar, L.N., Landesmann, J.B., Loranty, M.M., Maezumi, S.Y.,
forests in the short, mid and long term. Sci. Total Environ. 573, 1171–1177. https:// Mearns, L., Moritz, M., Myers, J.A., Pausas, J.G., Pellegrini, A.F.A., Platt, W.J.,
doi.org/10.1016/j.scitotenv.2016.03.115. Roozeboom, J., Safford, H., Santos, F., Scheller, R.M., Sherriff, R.L., Smith, K.G.,
Gordon, C.E., Price, O.F., Tasker, E.M., 2017. Mapping and exploring variation in post- Smith, M.D., Watts, A.C., Durigan, G., 2020. Fire as a fundamental ecological
fire vegetation recovery following mixed severity wildfire using airborne LiDAR. process: Research advances and frontiers. J. Ecol. 108 (5), 2047–2069. https://doi.
Ecol. Appl. 27, 1618–1632. https://doi.org/10.1002/eap.1555. org/10.1111/1365-2745.13403.
Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., Moore, R., 2017. Montealegre, A.L., Lamelas, M.T., Tanase, M.A., De la Riva, J., 2014. Forest fire severity
Google Earth Engine: planetary-scale geospatial analysis for everyone. Remote Sens. assessment using ALS data in a mediterranean environment. Remote Sens. 6,
Environ. 202, 18–27. https://doi.org/10.1016/j.rse.2017.06.031. 4240–4265. https://doi.org/10.3390/rs6054240.
Gu, D., Gillespie, A., 1998. Topographic normalization of Landsat TM images of forest Morresi, D., Vitali, A., Urbinati, C., Garbarino, M., 2019. Forest spectral recovery and
based on subpixel Sun-canopy-sensor geometry. Remote Sens. Environ. 64, 166–175. regeneration dynamics in stand-replacing wildfires of central Apennines derived
https://doi.org/10.1016/S0034-4257(97)00177-6. from Landsat time series. Remote Sens. 11 (3), 308. https://doi.org/10.3390/
Hislop, S., Jones, S., Soto-Berelov, M., Skidmore, A., Haywood, A., Nguyen, T., 2018. rs11030308.
Using Landsat spectral indices in time-series to assess wildfire disturbance and Morsdorf, F., Kötz, B., Meier, E., Itten, K.I., Allgöwer, B., 2006. Estimation of LAI and
recovery. Remote Sens. 10 (3), 460. https://doi.org/10.3390/rs10030460. fractional cover from small footprint airborne laser scanning data based on gap
Huete, A., Didan, K., Miura, T., Rodriguez, E.P., Gao, X., Ferreira, L.G., 2002. Overview fraction. Remote Sens. Environ. 104, 50–61. https://doi.org/10.1016/j.
of the radiometric and biophysical performance of the MODIS vegetation indices. rse.2006.04.019.
Remote Sens. Environ. 83, 195–213. https://doi.org/10.1016/S0034-4257(02) Moya, D., González-De Vega, S., García-Orenes, F., Morugán-Coronado, A., Arcenegui,
00096-2. V., Mataix-Solera, J., Lucas-Borja, M.E., De las Heras, J., 2018. Temporal
Ibáñez, I., Acharya, K., Juno, E., Karounos, C., Lee, B.R., Mccollum, C., Schaffer- characterisation of soil-plant natural recovery related to fire severity in burned Pinus
Morrison, S., Tourville, J., 2019. Forest resilience under global environmental halepensis Mill. forests. Sci. Total Environ. 640–641, 42–51. https://doi.org/
change : Do we have the information we need ? A systematic review. PLoS One 14 10.1016/j.scitotenv.2018.05.212.
(9), 1–17. https://doi.org/10.1371/journal.pone.0222207. Nguyen, T.H., Jones, S.D., Soto-Berelov, M., Haywood, A., Hislop, S., 2018. A spatial and
Jennings, S.B., Brown, N.D., Sheil, D., 1999. Assessing forest canopies and understorey temporal analysis of forest dynamics using Landsat time-series. Remote Sens.
illumination: Canopy closure, canopy cover and other measures. Forestry 72, 59–73. Environ. 217, 461–475. https://doi.org/10.1016/j.rse.2018.08.028.
https://doi.org/10.1093/forestry/72.1.59. Nikinmaa, L., Lindner, M., Cantarello, E., Jump, A.S., Seidl, R., Winkel, G., Muys, B.,
Kane, V.R., McGaughey, R.J., Bakker, J.D., Gersonde, R.F., Lutz, J.A., Franklin, J.F., 2020. Reviewing the Use of Resilience Concepts in Forest Sciences. Curr. For. Rep. 6,
2010. Comparisons between field- and LiDAR-based measures of stand structural 61–80. https://doi.org/10.1007/s40725-020-00110-x.
complexity. Can. J. For. Res. 40, 761–773. https://doi.org/10.1139/X10-024. Ninyerola, M., Fernàndez, X.P., i Nolla, J.M.R., 2005. Atlas climático digital de la
Kane, V.R., North, M.P., Lutz, J.A., Churchill, D.J., Roberts, S.L., Smith, D.F., Península Ibérica: metodología y aplicaciones en bioclimatología y geobotánica.
McGaughey, R.J., Kane, J.T., Brooks, M.L., 2014. Assessing fire effects on forest Universitat Autònoma de Barcelona.
spatial structure using a fusion of landsat and airborne LiDAR data in Yosemite

12
A. Viana-Soto et al. International Journal of Applied Earth Observation and Geoinformation 108 (2022) 102754

Pascual, C., García-Abril, A., Cohen, W.B., Martín-Fernández, S., 2010. Relationship Smola, A.J., Schölkopf, B., 2004. A tutorial on support vector regression. Stat. Comput.
between LiDAR-derived forest canopy height and Landsat images. Int. J. Remote 14 (3), 199–222. https://doi.org/10.1023/B:STCO.0000035301.49549.88.
Sens. 31, 1261–1280. https://doi.org/10.1080/01431160903380656. Soenen, S.A., Peddle, D.R., Coburn, C.A., 2005. SCS+C: A modified sun-canopy-sensor
Pausas, J.G., Bladé, C., Valdecantos, A., Seva, J.P., Fuentes, D., Alloza, J.A., Vilagrosa, A., topographic correction in forested terrain. IEEE Trans. Geosci. Remote Sens. 43,
Bautista, S., Cortina, J., Vallejo, R., 2004. Pines and oaks in the restoration of 2148–2159. https://doi.org/10.1109/TGRS.2005.852480.
Mediterranean landscapes of Spain: New perspectives for an old practice - A review. Taboada, A., Fernández-García, V., Marcos, E., Calvo, L., 2018. Interactions between
Plant Ecol. 171, 209–220. https://doi.org/10.1023/B:VEGE.0000029381.63336.20. large high-severity fires and salvage logging on a short return interval reduce the
Pausas, J.G., Keeley, J.E., 2014. Evolutionary ecology of resprouting and seeding in fire- regrowth of fire-prone serotinous forests. For. Ecol. Manage. 414, 54–63. https://doi.
prone ecosystems. New Phytol. 204, 55–65. https://doi.org/10.1111/nph.12921. org/10.1016/j.foreco.2018.02.013.
Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Thanos, C.A., Daskalakou, E.N., 2000. Reproduction in Pinus halepensis and P. brutia, in:
Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Ecology, Biogeography and Management of Pinus Halepensis and P. Brutia. Forest
Cournapeu, D., 2011. Scikit-learn. J. Machine Learn. Res. 12, 2825–2830. https:// Ecosystems in the Mediterranean Basin. Backhuys Publishers, Leiden, The
doi.org/10.1145/2786984.2786995. Netherlands, pp. 79–90.
Pérez-Cabello, F., Montorio, R., Alves, D.B., 2021. Remote Sensing Techniques to assess Tijerín, J., Moreno-Fernández, D., Zavala, M.A., Astigarraga, J., García, M., 2022.
Post-Fire Vegetation Recovery. Curr. Opin. Environ. Sci. Heal. 21, 100251. https:// Identifying Forest Structural Types along an Aridity Gradient in Peninsular Spain:
doi.org/10.1016/j.coesh.2021.100251. Integrating Low-Density LiDAR, Forest Inventory, and Aridity Index. Remote Sens.
Pflugmacher, D., Cohen, W.B., Kennedy, R.E., Yang, Z., 2014. Using Landsat-derived 14, 235. https://doi.org/10.3390/rs14010235.
disturbance and recovery history and lidar to map forest biomass dynamics. Remote Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring
Sens. Environ. 151, 124–137. https://doi.org/10.1016/j.rse.2013.05.033. vegetation. Remote Sens. Environ. 8, 127–150. https://doi.org/10.1016/0034-4257
Pickell, P.D., Hermosilla, T., Frazier, R.J., Coops, N.C., Wulder, M.A., 2016. Forest (79)90013-0.
recovery trends derived from Landsat time series for North American boreal forests. Turco, M., Rosa-Cánovas, J.J., Bedia, J., Jerez, S., Montávez, J.P., Llasat, M.C.,
Int. J. Remote Sens. 37, 138–149. https://doi.org/10.1080/ Provenzale, A., 2018. Exacerbated fires in Mediterranean Europe due to
2150704X.2015.1126375. anthropogenic warming projected with non-stationary climate-fire models. Nat.
Potapov, P., Li, X., Hernandez-Serna, A., Tyukavina, A., Hansen, M.C., Kommareddy, A., Commun. 9, 1–9. https://doi.org/10.1038/s41467-018-06358-z.
Pickens, A., Turubanova, S., Tang, H., Silva, C.E., Armston, J., Dubayah, R., Blair, J. Viana-Soto, A., Aguado, I., Martínez, S., 2017. Assessment of Post-Fire Vegetation
B., Hofton, M., 2021. Mapping global forest canopy height through integration of Recovery Using Fire Severity and Geographical Data in the Mediterranean Region
GEDI and Landsat data. Remote Sens. Environ. 253, 112165 https://doi.org/ (Spain). Environments 4, 90. https://doi.org/10.3390/environments4040090.
10.1016/j.rse.2020.112165. Viana-Soto, A., Aguado, I., Salas, J., García, M., 2020. Identifying post-fire recovery
Pourshamsi, M., Xia, J., Yokoya, N., Garcia, M., Lavalle, M., Pottier, E., Balzter, H., 2021. trajectories and driving factors using Landsat time series in fire-prone Mediterranean
Tropical forest canopy height estimation from combined polarimetric SAR and pine forests. Remote Sens. 12 (9), 1499. https://doi.org/10.3390/rs12091499.
LiDAR using machine-learning. ISPRS J. Photogramm. Remote Sens. 172, 79–94. Vogeler, J.C., Braaten, J.D., Slesak, R.A., Falkowski, M.J., 2018. Extracting the full value
https://doi.org/10.1016/j.isprsjprs.2020.11.008. of the Landsat archive: Inter-sensor harmonization for the mapping of Minnesota
Powell, S.L., Cohen, W.B., Healey, S.P., Kennedy, R.E., Moisen, G.G., Pierce, K.B., forest canopy cover (1973–2015). Remote Sens. Environ. 209, 363–374. https://doi.
Ohmann, J.L., 2010. Quantification of live aboveground forest biomass dynamics org/10.1016/j.rse.2018.02.046.
with Landsat time-series and field inventory data: A comparison of empirical White, J.C., Saarinen, N., Kankare, V., Wulder, M.A., Hermosilla, T., Coops, N.C.,
modeling approaches. Remote Sens. Environ. 114, 1053–1068. https://doi.org/ Pickell, P.D., Holopainen, M., Hyyppä, J., Vastaranta, M., 2018. Confirmation of
10.1016/j.rse.2009.12.018. post-harvest spectral recovery from Landsat time series using measures of forest
Rigo, D. De, Libertà, G., Durrant, T.H., Artés, T., San-Miguel-, J., 2017. Forest fire danger cover and height derived from airborne laser scanning data. Remote Sens. Environ.
extremes in Europe under climate change: variability and uncertainty. Publications 216, 262–275. https://doi.org/10.1016/j.rse.2018.07.004.
Office of the European Union, Luxembourg. https://doi.org/10.2760/13180. White, J.C., Wulder, M.A., Hermosilla, T., Coops, N.C., Hobart, G.W., 2017. A nationwide
Roberts, D.W., Cooper, S. V, 1989. Concepts and techniques of vegetation mapping. annual characterization of 25 years of forest disturbance and recovery for Canada
Pages 90–96 BT - compilers. Proceedings of a symposium—land classifications based using Landsat time series. Remote Sens. Environ. 194, 303–321. https://doi.org/
on vegetation: applications for resource management. USDA Forest Service General 10.1016/j.rse.2017.03.035.
Technical Report INT-257, in: Ferguson, D.E., Morgan, P., Johnson, F.D. (Eds.), . Xu, Y., Goodacre, R., 2018. On Splitting Training and Validation Set: A Comparative
Intermountain Research Station, Ogden. Study of Cross-Validation, Bootstrap and Systematic Sampling for Estimating the
Roy, D.P., Kovalskyy, V., Zhang, H.K., Vermote, E.F., Yan, L., Kumar, S.S., Egorov, A., Generalization Performance of Supervised Learning. J. Anal. Test. 2, 249–262.
2016. Characterization of Landsat-7 to Landsat-8 reflective wavelength and https://doi.org/10.1007/s41664-018-0068-2.
normalized difference vegetation index continuity. Remote Sens. Environ. 185, Zald, H.S.J., Wulder, M.A., White, J.C., Hilker, T., Hermosilla, T., Hobart, G.W.,
57–70. https://doi.org/10.1016/j.rse.2015.12.024. Coops, N.C., 2016. Integrating Landsat pixel composites and change metrics with
Seidl, R., Thom, D., Kautz, M., Martin-Benito, D., Peltoniemi, M., Vacchiano, G., Wild, J., lidar plots to predictively map forest structure and aboveground biomass in
Ascoli, D., Petr, M., Honkaniemi, J., Lexer, M.J., Trotsiuk, V., Mairota, P., Saskatchewan. Canada. Remote Sens. Environ. 176, 188–201. https://doi.org/
Svoboda, M., Fabrika, M., Nagel, T.A., Reyer, C.P.O., 2017. Forest disturbances 10.1016/j.rse.2016.01.015.
under climate change. Nat. Clim. Chang. 7, 395–402 https://doi.org/10.1038/ Zhao, K., Popescu, S., Meng, X., Pang, Y., Agca, M., 2011. Characterizing forest canopy
nclimate3303. structure with lidar composite metrics and machine learning. Remote Sens. Environ.
Senf, C., Müller, J., Seidl, R., 2019. Post-disturbance recovery of forest cover and tree 115, 1978–1996. https://doi.org/10.1016/j.rse.2011.04.001.
height differ with management in Central Europe. Landsc. Ecol. 34, 2837–2850. Zhu, Z., 2017. Change detection using Landsat time series: A review of frequencies,
https://doi.org/10.1007/s10980-019-00921-9. preprocessing, algorithms, and applications. ISPRS J. Photogramm. Remote Sens.
Shi, Y., Wang, T., Skidmore, A.K., Heurich, M., 2018. Important LiDAR metrics for 130, 370–384. https://doi.org/10.1016/j.isprsjprs.2017.06.013.
discriminating forest tree species in Central Europe. ISPRS J. Photogramm. Remote
Sens. 137, 163–174. https://doi.org/10.1016/j.isprsjprs.2018.02.002.

13