Characterizing Live Fuel Moisture Content from Active and Passive Sensors in a Mediterranean Environment

Abstract

Live fuel moisture content (LFMC) influences many fire-related aspects, including flammability, ignition, and combustion. In addition, fire spread models are highly sensitive to LFMC values. Despite its importance, LFMC estimation is still elusive due to its dependence on plant species traits, local conditions, and weather patterns. Although LFMC mapping from active synthetic aperture radar has increased over the past years, their utility for LFMC estimation needs further analysis to include additional areas characterized by different vegetation species and fire regimes. This study extended the current knowledge using medium spatial resolution (20 m) time series acquired by active (Sentinel-1) and passive (Sentinel-2) sensors. Our results show that optical-based LFMC estimation may achieve acceptable accuracy (R2 = 0.55, MAE = 15.1%, RMSE = 19.7%) at moderate (20 m) spatial resolution. When ancillary information (e.g., vegetation cover) was added, LFMC estimation improved (R2 = 0.63, MAE = 13.4%). Contrary to other studies, incorporating Sentinel-1 radar data did not provide for improved LFMC estimates, while the use of SAR data alone resulted in increased estimation errors (R2 = 0.28, MAE = 19%, RMSE = 25%). For increased fire risk scenarios (LFMC < 120%), estimation errors improved (MAE = 9.1%, RMSE = 11.8%), suggesting that direct LFMC retrieval from satellite data may be achieved with high temporal and spatial detail.

Full text

Forests 2022, 13, 1846. https://doi.org/10.3390/f13111846 www.mdpi.com/journal/forests
Article
Characterizing Live Fuel Moisture Content from Active and
Passive Sensors in a Mediterranean Environment
Mihai A. Tanase 1,*, Juan Pedro Gonzalez Nova 1, Eva Marino 2, Cristina Aponte 3,*, Jose Luis Tomé 2, Lucia Yáñez 2,
Javier Madrigal 4,5, Mercedes Guijarro 4 and Carmen Hernando 4
1 Department of Geology, Geography and Environment, University of Alcala, C. Colegios 2,
28801 Alcala de Henares, Spain
2 AGRESTA S. Coop., calle Duque de Fernán Núñez 2, 28012 Madrid, Spain
3 Department of Environment and Agronomy, Centro Nacional Instituto de Investigación y Tecnología
Agraria y Alimentaria, INIA-CSIC, Ctra. Coruña Km 7.5, 28040 Madrid, Spain
4 Forestry Science Institute ICIFOR, Centro Nacional Instituto de Investigación y Tecnología Agraria y
Alimentaria, INIA-CSIC, Ctra Coruña km 7.5, 28040 Madrid, Spain
5 ETSI Montes, Forestal y del Medio Natural, Universidad Politécnica de Madrid (UPM), Ramiro de Maeztu,
28040 Madrid, Spain
* Correspondence: mihai@tma.ro (M.A.T.); cristina.aponte@inia.csic.es (C.A.)
Abstract: Live fuel moisture content (LFMC) influences many fire-related aspects, including flam-
mability, ignition, and combustion. In addition, fire spread models are highly sensitive to LFMC
values. Despite its importance, LFMC estimation is still elusive due to its dependence on plant spe-
cies traits, local conditions, and weather patterns. Although LFMC mapping from active synthetic
aperture radar has increased over the past years, their utility for LFMC estimation needs further
analysis to include additional areas characterized by different vegetation species and fire regimes.
This study extended the current knowledge using medium spatial resolution (20 m) time series ac-
quired by active (Sentinel-1) and passive (Sentinel-2) sensors. Our results show that optical-based
LFMC estimation may achieve acceptable accuracy (R2 = 0.55, MAE = 15.1%, RMSE = 19.7%) at mod-
erate (20 m) spatial resolution. When ancillary information (e.g., vegetation cover) was added,
LFMC estimation improved (R2 = 0.63, MAE = 13.4%). Contrary to other studies, incorporating Sen-
tinel-1 radar data did not provide for improved LFMC estimates, while the use of SAR data alone
resulted in increased estimation errors (R2 = 0.28, MAE = 19%, RMSE = 25%). For increased fire risk
scenarios (LFMC < 120%), estimation errors improved (MAE = 9.1%, RMSE = 11.8%), suggesting
that direct LFMC retrieval from satellite data may be achieved with high temporal and spatial detail.
Keywords: Sentinel-1; Sentinel-2; live fuel moisture content
1. Introduction
Fire danger, an important component of fire management systems, largely depends
on meteorological variables and fuel conditions, as the topography is invariable over time
[1,2]. Fuel condition is a critical parameter as it influences flammability, ignition, combus-
tion, and fire spread [3–5]. While dead fuel condition is driven by weather patterns (e.g.,
heat, dryness, wind, rain), the live fuel moisture content (LFMC) depends on plant species
traits [6,7]. LMFC spatial variation affects fire occurrence, intensity, and spread [8–11],
with inverse relationships between ignition probability and live fuel moisture content be-
ing demonstrated in semi-arid environments [3,12]. In addition, reduced vegetation mois-
ture is related to increasing large fire intensity and occurrence [13,14]. As similar atmos-
pheric conditions can result in differentiated effects due to the physiological characteris-
tics of individual trees or species-related resistance to drought, LFMC estimation based
on meteorological information alone may be affected by errors [9,15] or produce results
at the coarse spatial resolutions (e.g., 2 km) of the gridded weather data [16]. Therefore,
Citation: Tanase, M.; Nova, J.P.G.;
Marino, E.; Aponte, C.; Tomé, J.L.;
Yáñez, L.; Madrigal, J.; Guijarro, M.;
Hernando, C. Characterizing Live
Fuel Moisture Content from Active
and Passive Sensors in a
Mediterranean Environment. Forests
2022, 13, 1846. https://doi.org/
10.3390/f13111846
Academic Editors: Luis A. Ruiz
and Pablo Crespo-Peremarch
Received: 23 September 2022
Accepted: 2 November 2022
Published: 4 November 2022
Publisher’s Note: MDPI stays neu-
tral with regard to jurisdictional
claims in published maps and institu-
tional affiliations.
Copyright: © 2022 by the authors. Li-
censee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and con-
ditions of the Creative Commons At-
tribution (CC BY) license (https://cre-
ativecommons.org/licenses/by/4.0/).

Forests 2022, 13, 1846 2 of 17
remote sensing technologies [17–22] were increasingly used to understand LFMC spatio-
temporal dynamics at improved spatial resolutions [23].
Until recently, remote sensing-based LFMC retrieval relied on information acquired
within the visible and infrared spectrum (optical sensors) at different time lags, days to
weeks, and at spatial resolutions varying from meter to kilometers [17,24,25]. Such studies
used empirical relationships to link in situ measurements with surface reflectance or spec-
tral indices [17,18,21,25,26] as well as physically based radiative transfer models [17,20,27].
Such approaches take advantage of the direct effects of tissue water content on near and
shortwave infrared (NIR and SWIR) spectral reflectance absorption. However, changes in
leaf structure and pigment concentrations may obscure such relationships [28]. In addi-
tion, LFMC estimation based on optical sensors may be affected by persistent cloud cover
[29], decoupled reflectance values from dry matter content, or variations related to canopy
properties [28,30]. The performance of optical-based LFMC retrieval varies with the veg-
etation type, with lower (10%) root mean squared estimation errors being observed for
forests and shrublands when compared to grasslands (~30%) [17,20,22,25,30,31]. For more
detailed information on LFMC and its retrieval from remote sensing data, the reader is
referred to [28].
To reduce LFMC estimates uncertainties some studies focused on the use of the mi-
crowave region of the spectrum taking advantage of passive [32,33] or active microwave
[19,29,34] satellite sensors. Past studies showed that vegetation dielectric properties and
liquid water content are correlated strongly with water not bound to the vegetation ma-
terial having a significant influence on the radar signal [35]. Such influence can be used to
monitor plant water diurnal variations or water stress [36–38]. Indeed, the use of passive
microwave improved LFMC estimation error to about 20% using indirect methods based
on the time-lagged correlation of LFMC with soil moisture or vegetation optical depth
derived from passive radiometers. However, as the spatial resolution of passive radiom-
eters is coarse (e.g., 36 km for Soil Moisture Active Passive mission, SMAP), the derived
products are available with low spacings (e.g., 9 km for the SMAP L4 soil moisture prod-
uct). Therefore, the LFMC estimates derived through such approaches may not provide
adequate spatial detail for some components of the fire management systems (e.g., fire
spread simulation). Active microwaves sensors such as synthetic aperture radar (SAR)
have been used to estimate dead fuel moisture [39,40] as well as vegetation water content
[41,42], but until recently, only two studies were available on their use for LFMC retrieval
[19,29]. Although these studies suggested estimation errors ranging between 10% and 15%
in Canadian and Australian forests, the LFMC monitoring from SAR data was hindered
by the low temporal frequency of older SAR satellites or the use of airborne sensors. How-
ever, with the launch of the Sentinel-1 mission in 2014, such limitations were removed as
it provides for both the spatial resolution and the temporal frequency needed for opera-
tional LFMC monitoring. Indeed, a recent study carried out over dry shrubland-domi-
nated sites suggested that using active microwave data and semi-empirical modeling may
provide improved LFMC estimates when compared to using spectral reflectance [34]. Fur-
ther, other authors suggested that combining optical and active microwave sensors within
a deep learning framework may enhance modeling performance over the diverse ecolog-
ical conditions in the western US [23], with the LFMC root mean squared error (RMSE)
decreasing from 32% when using optical data alone, to 25% when using both sensor types.
As fire spread models are highly sensitive to LFMC [5], with over 1000% difference
in fire rate of spread being induced by only a 10% difference in LFMC estimates [43], the
utility of LFMC estimates for fire management systems ultimately depends on their accu-
racy which in turn may vary with vegetation type, sensor, and modeling approach. SAR-
based LFMC estimation is still scarce, with all available studies being limited to short time
spans (one or two fire seasons) and few forest stands spread over relatively small areas
[19,29,34]. In addition, although optical and microwave sensors may provide complemen-
tary information to estimate LFMC, the joint use of active and passive sensors is limited
to one study [23].

Forests 2022, 13, 1846 3 of 17
This study extended the work in [23] to the Mediterranean basin. More important,
we derived LFMC estimates at a spatial resolution an order of magnitude higher (20 m vs.
250 m) while also testing the unconventional “handcrafted” variables (i.e., ratios of optical
and radar data) proposed in [23] using comparable modeling approaches. We assessed
the utility of Sentinel-1 (S1) and Sentinel-2 (S2) sensors individually, as well as their joint
use, for LFMC estimation. The influence of temporal differences between satellite data
acquisition and in situ measurements was appraised together with the importance of us-
ing static variables (e.g., canopy height, vegetation fractional cover, slope, orientation) as
LFMC predictors. The study considered LFMC estimation over both the entire year as well
as under high fire risk scenarios (LFMC < 120%) associated with increased flammability
and fire spread [11]. LFMC modeling was based on artificial intelligence algorithms [44]
as such non-linear models may capture parametric relationships without assuming an a
priori analytical form. Such approaches allow for understanding the relationships be-
tween the dependent (LFMC) and the independent remote sensed based on the data [23].
This is the only second study addressing LFMC retrieval using active and passive sensors
and long-term datasets.
2. Materials and Methods
2.1. Study Area
The study was carried out in the Madrid Region, which extends over 8030 km2 in the
center of the Iberian Peninsula (Figure 1) at elevations between 400 m above sea level in
the south and 2400 m in the north. The relief changes from planes in the Tagus valley in
the south to mountains in the north which results in a significant variation in ecological
conditions. According to the data from the Spanish state meteorological agency (AEMET),
the climate is Mediterranean, with hot summers and an average annual rainfall of 450
mm, which occurs mainly in spring and autumn with an important gradient from the
Tagus valley (<500 mm year−1) towards the mountains (1500 mm year−1). The average
Figure 1. Spatial distribution of sampling locations (LID1 to LID6) within the Madrid region cen-
tered at 3.7 W and 40.5 N. Upper panels exemplify plot selection using GPS perimeters and ortho-
photos (pine—left, rockrose—right).

Forests 2022, 13, 1846 4 of 17
monthly temperature ranges between 1 °C in winter and 32 °C in summer, with a gradient
from south (hotter) to north (colder). The area covered by natural vegetation is about 0.44
M ha corresponding to 55% of the Madrid region. The main tree species are oaks (i.e.,
Quercus ilex L., Q. pyrenaica W.) and pines (i.e., Pinus halepensis M., P. sylvestris L., P. pinea,
P. pinaster A.), which account for 26% and 11.5%, respectively, of the natural vegetation
cover. An additional 12% of the natural vegetation area consists of mixed forest species.
Grasslands and riparian vegetation represent 16.5% and 2%, respectively, with the re-
maining areas (33%) being covered by shrublands (i.e., Cistus ladanifer L., Q. coccifera L.,
Retama sphaerocarpa L., Thymus sp., Erica sp. Macrochloa tenacissima L.).
2.2. Live Fuel Moisture Content Sampling
The field data consisted of LFMC samples for the main tree (oaks and pines) and
rockrose (Cistus sp.) shrub species from 16 plots grouped around six locations IDs (LIDs 1
to 6 in Figure 1). Field samples have been collected since April 2016 with a variable tem-
poral frequency: every two weeks during winter, every week during spring and autumn,
and twice a week during summer. Each sample consists of 50–100 g of fine live fuels, in-
cluding leaves and branches smaller than 6 mm. Random linear transects were established
in shrubland areas to avoid re-sampling the same plants. Samples were collected from at
least five individuals along the transect. For tree species, five random individuals were
sampled at each date to account for intra-species variability. Samples were collected
around noon, placed in individual airtight plastic containers, and transported to the la-
boratory before 3:00 PM to prevent moisture loss. Fresh samples were weighed and oven
dried at 100 °C for 24 h. Subsequently, each sample’s dry weight was measured to estimate
the moisture content as the percentage of water contained in the vegetation with respect
to the dry weight. A total of 4109 field samples were available for this study covering the
period from April 2016 to December 2020. At each location, the available samples de-
pended on the present species (Table 1). A manual screening was carried out to exclude
anomalous samples caused by incorrect handling (e.g., container incorrectly closed, water
in the container) that resulted in anomalous LFMC values or errors during field data col-
lection (e.g., samples that included flowers, fruits or woody material, samples collected
on rainy days, samples too small). In total, 2962 samples were retained for analysis after
data screening.
2.3. Remote Sensing Data
Images acquired by the Sentinel-2 Multi-Spectral Instrument (MSI) and the Sentinel-
1 C-band SAR sensors were used in this study. Six hundred and thirteen Sentinel-2 gran-
ules for the Military Grid Reference System (MGRS) tiles 30TUK, 30TVK and 30TVL were
downloaded as atmospherically corrected, topographically normalized surface reflec-
tance images from Theia Data and Services Centre, a French public center that facilitates
the use of images from spatial observations. The downloaded data covered the period
between April 2016 and December 2020. Only data with a cloud cover below 90% were
downloaded. Theia data are derived from the original L1C level imagery through the
multi-sensor atmospheric correction and cloud screening (MACSS). MACCS detects
clouds and shadows, estimates aerosol optical thickness and water vapor, and corrects for
Table 1. Field samples, by species, collected at each location.
Location ID/Name
Sampled species
Number of samples
1/Buitrago de Lozoya
Pinus pinaster, Quercus ilex, Cistus ladanifer
224, 224, 224
2/Bustarviejo
Q. ilex, C. ladanifer
224, 224
3/La Marañosa
P.halepensis, Q.coccifera, C. albidus
203, 203, 203
4/Navalcarnero
P.halepensis, P.pinaster Q. ilex, C. ladanifer
203, 203, 203, 203
5/Robledo de Chavela
P. pinaster, Q.s ilex, C. ladanifer
224, 224, 448
6/Guadarrama
P. pinaster, Q. ilex, C. ladanifer
224, 224, 224

Forests 2022, 13, 1846 5 of 17
Table 2. Spectral indices derived from Sentinel-2 data. B stands for Sentinel-2 band.
Index
Formula with Band Number
Enhanced Vegetation Index (EVI)
2.5 × (B8 − B4)/(B8 + 6 × B4 − 7.5 × B2 + 1)
Soil Adjusted Vegetation Index (SAVI)
((B8 − B4)/(B8 + B4 + 0.5)) × 1.5
Optimized Soil Adjusted Vegetation Index (OSAVI)
(B8 − B4)/(B8 + B4 + 0.16)
Normalized Difference Vegetation Index (NDVI)
(B8 − B4)/(B8 + B4)
Ratio Vegetation Index (RVI)
B8/B4
Ratio Vegetation Index 2 (RVI2)
B8/B2
Visible Atmospherically Resistant Index (VARI))
(B3 − B4)/(B3 + B4 − B2)
Normalized Difference Moisture Index (NDMI)
(B8A − B11)/(B8A + B11)
Normalized Multi-band Drought Index (NMDI)
(B8A − (B11 − B12))/(B8A + (B11 − B12))
Normalized Difference Water Index (NDWI)
(B8 − B12)/(B8 + B12)
Vegetation Index-Green (VIgreen)
(B3 − B5)/(B3 + B5)
Transformed Chlorophyll Absorption Index (TCARI)
3 × ((B5 − B4) − 0.2 × (B5 − B3) × (B5/B4))
TCARI/OSAVI (ratio)
TCARI/OSAVI
Specific leaf area vegetation index (SLAVI)
(B8)/(B4 + B12)
Normalized difference infrared index (NDII)
(B8 − B11)/(B8 + B11)
Global Vegetation Moisture Index (GVMI)
((B8A + 0.1) − (B12 + 0.02))/((B8a + 0.1) + (B12 + 0.02))
Moisture stress index (MSIn)
MSI = B11/B8
Atmospherically resistant vegetation index (ARVI)
B8 − (B4 − (B2 − B4)/B8 + (B4 − (B2 − B4)
TC_Brightness (TC_B)
0.3037 × B2 + 0.2793 × B3 + 0.4743 × B4 + 0.5585 × B8 +
0.5082 ×B11 + 0.1863 ×B12
TC_Wetness (TC_W)
0.1509 × B2 + 0.1973 × B3 + 0.3279 × B4 + 0.3406 × B8 −
0.7112 × B11 − 0.4572 × B12
TC_Greeness (TC_G)
− 0.2848 × B2 − 0.2435 × B3 − 0.5436 × B4 + 0.7243 × B8 +
0.084 × B11 − 0.18 × B12
atmospheric effects. MACCS is currently integrated within the MACCS-ATCOR joint al-
gorithm (MAJA), open-source software that takes advantage of multi-temporal methods
for optical image corrections [45–47]. Surface reflectance data from the 13 visible and in-
frared bands of the MSI was used to compute 21 spectral indices (SI, Table 2) sensitive to
vegetation properties and live fuel moisture content [17,20,21,23–25,30].
In total, 3550 ground range detected (GRD) images (10 m pixel spacing) acquired by
the C-band Sentinel-1 satellites in interferometric wide (IW) mode were downloaded from
the Copernicus Open Access Hub repository and processed through the s1tiling tool that
uses SAR specific algorithms included with the Orfeo ToolBox (OTB), an open-source soft-
ware developed and maintained by the Centre National D’Etudes Spatiales (CNES),
France [48]. Sentinel-1 processing involves thermal noise removal, radiometric calibration
to gamma naught, orthorectification, and temporal filtering [49]. Thermal noise removal
and radiometric calibration were based on product metadata, while the orthorectification
process, carried out at 20 m pixel spacing to match the Sentinel-2 data, used the Shuttle
Radar Topographic Mission (SRTM) digital elevation model (DEM) at 30 m spacing. The
resulting images, providing the backscatter coefficient for the VV or VH polarizations, are
subsequently tiled and registered to the Sentinel-2 images as the same MGRS geographic
reference is used for orthorectification. Further, the radar SPAN (VV + VH) and polariza-
tion ratio (VV/VH) were computed. Out of all Sentinel-1 relative orbits intersecting the
three tiles of interest (30TUK, 30TVK, and 30TVL), only the relative orbits 1 ascending
(1A) and 81 descending (81D) were used as these orbits covered the entire area of interest.
2.4. Matching LFMC Samples with Ground Areas
As LFMC field sampling was not designed for use with satellite imagery, the field
crew collected samples close to specific xy coordinates (i.e., location ID) without spatially

Forests 2022, 13, 1846 6 of 17
explicit identification of each sampled tree or shrubland transect. Therefore, the LFMC
samples needed to be matched to ground areas (plots) covered by the sampled species.
Such matching was achieved based on fieldwork and high-resolution (0.5 m) orthophoto-
graphs of the Spanish National Program of Aerial Photogrammetry (PNOA). Plots of ho-
mogeneous vegetation conditions (pure or mixed species) of at least 800 m2 (two satellite
pixels) were delineated on the ground with a handheld GPS in close vicinity (< 250 m) to
the location ID coordinates and maintained elevation, slope, and aspect. LFMC samples
were subsequently matched with the plots identified at each location. In addition, for
mixed vegetation plots, the corresponding LFMC values were averaged using as weight
the proportion of fractional cover for each species as estimated during the fieldwork.
Matching the field samples with ground areas of homogeneous vegetation resulted in 16
plots totaling 2962 LFMC estimates (Table 3). To avoid border effects and location uncer-
tainties, the 20 m Sentinel-2 satellite imagery grid was used to select full pixels well within
the GPS-delineated homogeneous area (Figure 1). The plots were subsequently used to
extract information on the surface reflectance and backscatter coefficient from the optical
and the radar images, respectively, acquired at the closest date for a given maximum time
difference when compared to the field sampling date.
2.5. Exploratory Analysis and LFMC Estimation
An exploratory analysis was carried out to identify LFMC annual trends for each sampled
species, evaluate LFMC variability between locations and years, and assess the relation-
ships between satellite data and field sampled LFMC in different configurations (i.e., plot
level, species level, yearly samples vs. all samples). The exploratory analysis also com-
pared the use of SI against relative SI values, as well as the influence of the allowed time
difference between field sampling and satellite data acquisition. The relative values were
computed as in Equations (1) and (2). The allowed temporal range was ±3 days and ±6
days. The range allows for one clouded image over a 5-day period for the Sentinel-2 sat-
ellite and one missed acquisition over a 6-day period for the Sentinel-1 satellite.
Table 3. Number of unique sampling dates matched with satellite images (2016–2020 period) for
each location ID (LID) and plotted together with the static variables (CH—canopy height, FCC—
forest fractional cover, elevation, slope, aspect) computed from PNOA Aerial Laser Scanning data.
LID
Plot ID and
Dominant Species
Area
(m2)
Unique
Dates
Static Variables
CH (m)
FCC (%)
Elevation
(m)
Slope
(deg.)
Aspect
(deg.)
1
C-1; Pinus sp.
1600
197
16.2
80.8
1037
5
327
C-2; Quercus sp.
1200
197
5.9
34.0
1045
7
302
C-3; Cistus sp.
1200
198
2.8
6.3
1037
2
121
C-4; Quercus sp. and Cistus sp.
1600
198
6.9
32.8
1040
3
133
2
2B; Cistus sp.
1600
199
0.4
0.0
1248
7
140
3
C-1; Pinus sp.
1600
160
14.7
68.3
654
8
214
C-2; Pinus sp. and Quercus sp.
1600
160
12.0
57.4
658
11
259
C-3; Quercus sp.
1600
158
2.4
2.7
647
10
265
C-4; Quercus sp. and Cistus sp.
2400
159
1.0
0.4
647
12
292
4
B; Cistus sp.
800
175
8.2
12.1
611
8
158
C; Pinus sp.
1600
178
15.6
47.5
596
10
155
5
A; Quercus sp. and Cistus sp.
1200
198
7.3
43.9
881
13
111
B; Cistus sp.
2400
192
1.2
0.0
1055
22
89
6
A-1; Pinus sp.
1600
197
15.3
81.5
1097
9
69
A-2; Cistus sp.
800
197
7.7
21.8
1103
7
78
A-3; Quercus sp. and Cistus sp.
1200
199
5.8
46.8
1111
6
67

Forests 2022, 13, 1846 7 of 17
It should be noted that while haze, clouds, and cloud shadows may often obscure the
landscape in the case of optical data, missed data acquisitions are uncommon for Sentinel-
1 and may arise due to changes in the basic operation scenario during emergencies (e.g.,
natural disasters), orbiting maneuvers, or sensor malfunction. Within the exploratory
analysis, we also evaluated the influence of Sentinel-1 processing (temporally filtered vs.
unfiltered data), the satellite-looking geometry (ascending vs. descending passes), and the
use of the so-called handcrafted variables, “ratios of microwave and optical data”, sug-
gested in [23].
rSI = (SI − SImin)/(SImax − SImin)
(1)
rLFMC = (LFMC − LFMCmin)/(LFMCmax − LFMCmin)
(2)
where r is relative, SI is the spectral index, LFMC is the live fuel moisture content, and
min and max values are relative to the plot/species combination.
LFMC estimation was carried out considering the results of the exploratory analysis
using random forests (RF) regression, a non-parametric modeling approach [44], found to
minimize LFMC estimation errors compared to other machine learning approaches [50].
As non-parametric models have no assumptions regarding the statistical properties of the
data and offer the opportunity to include non-linearly related variables, they are often
used when enough samples are available for model calibration. RFs use ensemble learning
to improve the predictive power by aggregating predictions from constituent sub-models
(i.e., trees). Each tree is built using a deterministic algorithm by selecting a random set of
variables and a random sample from the training dataset [44].
The LMFC estimation was based on both remote sensing (i.e., optic, radar) and static
(St) variables. The use of additional static variables, i.e., vegetation height and fractional
cover, species, elevation, slope, eastness (i.e., sine of aspect), and northness (i.e., cosine of
aspect), was also evaluated to ascertain the opportunity for LFMC retrieval improve-
ments. Each sensor was individually tested to generate a reference baseline and allow for
cross-sensor evaluation. For parsimony, the models were trained using a subset of predic-
tor variables as evaluated within the exploratory analysis. The predictor variables were
selected by clustering highly correlated variables (|r| > 0.7) using both correlograms and
principal component analysis (PCA) and retaining only the most important variable in
each cluster. Variable importance was evaluated through the increase in the mean squared
error (MSE) when the variable was removed from the predictor pool during RF modeling.
To better understand predictor variables’ role for relevant fire risk scenarios, lower values
(LFMC < 120%) were modeled independently. LFMC below 120% is associated with in-
creased fire occurrence in Mediterranean vegetation [9,11].
Following the preliminary analysis (see Section 3.1), LMFC estimation (model fitting)
was based on the absolute LFMC values (i.e., as opposed to using relative LMFC values)
with a maximum difference of ±6 days. In the case of the Sentinel-1 SAR data, the tempo-
rally filtered backscatter coefficient was used. All models were fitted with a minimum
common dataset (1486 samples: 516 oak, 520 rockrose, 450 pine) using 1250 decision trees.
The proportion of samples used for training was 60%, with the remaining samples being
used for out-of-bag (OOB) validation. The tested models were split into seven sets de-
pending on the included predictors: all groups of variables, only remote sensing variables
(S1, S2, and S1 + S2), only static variables, and combinations of remote sensing and static
variables (i.e., S1 + St, S2 + St). Models were fitted to three different datasets, namely the
entire set, by vegetation type (i.e., oak, pine, and rockrose) subsets as well as to a subset
corresponding to potentially higher fire risk (LFMC < 120%).
Model assessment (goodness of fit) was based on metrics computed based on the
OOB set and included the correlation between actual and predicted values (r), the ex-
plained variance, the mean absolute error (MAE, 3), and the bias (4). In addition to the
OOB-based validation, an ex-situ cross-validation was carried out by training the models
using data from all but one location and using the remaining location for validation.

Forests 2022, 13, 1846 8 of 17
  



(3)
 


(4)
where P is the predicted values, O is the in situ observed values, and n is the number of samples,
and 
 and 
 are the mean values.
3. Results
3.1. Exploratory Analysis
There were statistically significant differences between the LFMC values depending
on the sampled species and location, with mean LFMC values being highest for pines and
lowest for oaks (Figure 2). Across locations, mean LFMC values varied for the same veg-
etation types, with samples from location ID (LID) 1 and 6 showing higher mean values
when compared to the remaining ones. Over time, rockrose (Cistus sp.) showed the high-
est variability, although oak trees showed a similar temporal pattern (highs and lows)
albeit with a lower amplitude, particularly during dry years (Figure 3). Pine trees showed
higher LFMC in summer, while the temporal trend was the opposite (peaks instead of
lows) when compared to rockrose and oaks. Pine samples from lower elevations (LID 3)
showed smaller LFMC values, while samples from pines located on northern aspects and
higher elevations (LIDs 1, 6) were characterized by higher LFMC values. Notice that such
patterns were not observed for oak and rockrose samples. Lastly, the forest structural pa-
rameters influenced LFMC (pines and oak stands), with increased values being related to
increased canopy height and cover.
Sentinel-2 SIs were highly correlated (r > 0.7) with the main group associated with
the first PCA axis and a remaining few SIs (e.g., VIgreen, TCARI, ARVI, TC_B, RVI, MSI)
orthogonal to the rest (see Figure S1 in the Supplementary Materials). Similarly, the Sen-
tinel-1 polarizations (VV, VH), the SPAN (VV + VH), and the polarization`s ratio (VV/VH)
were also correlated and covaried in the PCA analysis, suggesting three main groups re-
lated to (1) the ascending (asc) satellite pass, (2) the descending (des) satellite pass, and (3)
the polarization ratio for each pass (Figure S2). Filtered and unfiltered Sentinel-1 data
were highly correlated (r > 0.75), with mean values and dispersion being similar. A com-
parison of RF models based on Sentinel-2 SIs showed very similar results regardless
.
Figure 2. Live fuel moisture content (LFMC) variability across species and sampled locations (2016–
2021). Boxplots show median values, and dots are outliers.

Forests 2022, 13, 1846 9 of 17
Figure 3. Temporal trends for live fuel moisture content (LFMC) by species at locations 1 and 6 (all
species present). Lines show general additive models with integrated smoothness estimation. Grey
shades show confidence (0.95) interval around the smooth.
of the time difference between in situ and remote sensing data. Marginal improvements
in the explained variance (62% vs. 61%) were observed when using the longer (6-days)
time difference. One should note that data pairs with time differences above six days were
limited (<2%) and mostly occurred for Sentinel-2 during winter/spring, due to increasing
clouds, when the accuracy of LFMC prediction is less relevant. Removing such data from
the analysis would have little impact on the modeling outcomes.
The use of relative LFMC as a dependent variable rendered lower prediction accura-
cies when compared to using absolute values (62% vs. 45%). Using either absolute or rel-
ative satellite variables as predictors did not render differences in model fits for Sentinel-
1 (16% vs. 16%) nor Sentinel-2 (54% vs. 56%) data but including both the relative and the
absolute satellite variables increased the explained variance by approximately 10% (R2 =
28% Sentinel-1, R2 = 63% Sentinel-2). Therefore, all subsequent analyses used a time dif-
ference of six days, absolute LFMC values, temporally filtered SAR data, and both abso-
lute and relative satellite variables as predictors.
3.2. Predictor Variables
The evaluation of the static predictor variables showed that FCC and CH were highly
correlated (r = 0.99), with the latter being retained as it had a higher importance in RF
modeling (Figure S3). The most important radar variables for LFMC estimation were the
VH polarization and the SPAN acquired during both ascending and descending passes
(Figure S4). After removing the highly correlated radar variables, nine were kept for mod-
eling (Table 4). Among the Sentinel-2 SIs, the most important variable was the VARI. In
general, the relative version of the SIs had lower importance except for the relative version
of the EVI and VARI (Figure S5). A total of 11 variables were retained for RF modeling
after removing the highly correlated ones in each PCA cluster.
Table 4. Variables selected for LFMC estimation. Variable names as per Table 2. Suffix _r indicates
the relative version (on separate lines) of the variable.
Static Variables
Sentinel 1 Variables
Sentinel 2 Variables
CH, Elevation, Slope,
Northness, Eastness
VHasc, VHdes, VVasc, VV/VHasc,
SPANasc_r, VV/VHdes_r, VHdes_r,
VV/VHasc_r, VVdes_r
VARI, ARVI, RVI2, TCARI/OSAVI
EVI_r, RVI_r, TCARI_r, NMDI_r, TC_G_r,
RVI2_r, TC_B_r

Forests 2022, 13, 1846 10 of 17
3.3. LFMC Estimation
The RF model based on predictor variables from all groups (i.e., optic, radar, static)
and using all data (full model) performed similarly (Table 5) when compared to the model
containing only the optic and static (S2 + St) variables (r = 0.79 and MAE = 13.3%). By type
of predictor variables, the most accurate LFMC estimates were obtained when using the
optic variables (S2 model), with the r values decreasing slightly (0.74) and MAE values
slightly increasing (15%) when compared to the Full model. Radar-derived LFMC (S1
model) showed higher errors, with MAE reaching 19.0% and r decreasing to 0.53. For the
St model, canopy height was the most important variable when the entire dataset (all lo-
cations, all samples) was considered. The most important optic predictor variable (S2
model) was, by a significant margin, the VARI followed by the relative version of the EVI
and the RVI2 (Figure 4a). The VH polarization and the SPAN from the ascending pass
were the most important radar variables when radar metrics alone were used for LFMC
Table 5. LFMC retrieval error as a function of the predictor variables. S1 stand for Sentinel-1 varia-
bles, S2 for Sentinel-2 variables, and St for static variables. In bold are the most accurate estimates.
Data Set
Metric
S1
S2
St
S1 + S2
S1 + St
S2 + St
Full
All data
R2
0.28
0.55
0.26
0.56
0.38
0.62
0.62
Explained variance
27.3
53.8
26
54.2
38.4
62.3
61.6
Bias
−0.12
−0.31
−0.01
−0.31
−0.12
−0.23
−0.32
MAE
19.0
15.1
18.7
15.0
16.9
13.3
13.4
RMSE
24.8
19.7
25.0
19.7
22.8
17.8
18.0
LMFC <
120%
R2
0.23
0.48
0.21
0.48
0.31
0.54
0.54
Explained variance
22.5
47.1
20.6
46.8
30.7
53.7
52.6
Bias
0.38
0.12
−0.02
0.16
0.30
0.12
0.18
MAE
12.6
10.1
12.3
10.2
11.5
9.20
9.3
RMSE
15.3
12.7
15.5
12.7
14.5
11.8
12.0
(a)
(b)
Figure 4. Predictor variables’ importance (horizontal axis) by model when all data are used for mod-
eling (a) and when low (<120%) LFMC values (high fire risk scenario) are modeled (b).
0 500 1000 1500
CH
Slope
Eastness
VARI
ARVI
RVI_r
NMDI_r
TC_B_r
VHdes
VHasc
VVdes_r
VV/VHasc_r
SPANasc_r
Variables
0 200 400
CH
Slope
Eastness
VARI
ARVI
RVI_r
NMDI_r
TC_B_r
VHdes
VHasc
VVdes_r
VV/VHasc_r
SPANasc_r S1
S2
St
S1+St
S2+St
S1+S2

Forests 2022, 13, 1846 11 of 17
estimation (S1 model). In general, the full model showed that LFMC was overestimated
for values below 75% and underestimated for values above 150%, with the highest varia-
bility of residual errors being observed for the rockrose. The residual error was homoge-
neous throughout the year for all species (Figure S6).
When models were trained using data from all but one location and tested on the
excluded location (ex situ), the average performance decreased (Figure 5) for all locations
except LID1. For the full model, r decreased by 0.24 points (from 0.79 to a minimum of
0.49), while MAE increased by 12% (to a maximum of 25%). On average, the use of radar
(S1 model) or optical (S2 model) data made little difference in the LFMC retrieval error
across locations (2.7%) except for LID5, where the use of SIs resulted in almost 10% MAE
improvement. The LFMC difference between the most and the least accurate models at
each location was, on average, 6.1%, with the highest difference (~9%) being observed for
LID5 and the lowest (~3%) for LID 2 and LID3. The variables’ importance varied slightly
across locations, although the most important ones (VARI, EVI_r, and CH) remained un-
changed (Figure 6b).
Generating individual models for each vegetation type resulted in similar trends
across species (10% < MAE < 27%), with the most accurate models (10% < MAE < 12%)
being observed for oak and pine. Compared to the model based on the entire dataset
(MAE = 13.4%), by species MAE increased for rockrose (16.6%) and decreased for oak
(11.2%) and pine (11.9%). For species-specific models, the most important variables for
LFMC estimation differed, with the relative EVI being the most important one for rock-
rose, the VARI for oak, and the static ones (CH, elevation, northness) for pines (Figure 6a).
Figure 5. LFMC mean absolute error for each ex situ Location ID (LID 1 to 6).
(a)
(b)
Figure 6. Predictor variables’ importance for LFMC estimation by species (a) and by location ID (b).
0
10
20
30
S1 S2 St S1+St S2+St S1+S2 All
Mean absolute error
(%)
LID 1 LID 2 LID 3 LID 4 LID 5 LID 6
0%
20%
40%
60%
80%
100%
Rockrose Oak Pine
Variable importance
0%
20%
40%
60%
80%
100%
123456

Forests 2022, 13, 1846 12 of 17
When only lower (<120%) LFMC values were modeled, the MAE varied between
9.5% and 12.5% depending on model configuration, with the highest errors being ob-
served when using SAR metrics (S1 model). The importance of the variables did not
change (Figure 4b) when compared to modeling the entire dataset.
4. Discussion
Overall, when static variables were included, the performance of the radar-based
LFMC model (S1 + St model, MAE = 17.0%) was largely similar when compared to the
optical-based model (S2 + St model, MAE = 13.4%). The lowest LFMC errors were ob-
served at LID1 (15%), with all the remaining locations showing higher MAE (19%–25%).
The ordination of locations by environmental conditions (data not shown) indicated
largely different conditions for locations 3 and 4 which were drier areas dominated by
vegetation with lower canopy height and cover. As vegetation types varied from low
shrubs to medium-height oaks and taller pine forests, the static variables had an important
impact on LFMC estimation, with canopy height being the second most important varia-
ble after VARI despite the reduced variability of climatic and edaphic conditions and spe-
cies compositions. Removing the static variables degraded model performance by about
2% when using radar (MAE = 19.0%) and, respectively, optical (MAE = 15.1%) data. How-
ever, the large differences in the explained variance (27 vs. 54%) suggested that SIs reflec-
tance provided additional information not available in the backscatter coefficient.
The full model estimated LFMC with a cross-validated accuracy of 0.62 (R2), a mean
absolute error (MAE) of 13.4%, and an RMSE of 18.0% when jointly using optic, radar, and
static variables. Contrary to [23], the addition of Sentinel-1 backscatter data did not im-
prove LFMC estimation, i.e., model performance did not decrease when removing the
radar-derived variables. The relatively high importance of the VV polarization and the
open Mediterranean vegetation suggests that a large part of the SAR backscatter may orig-
inate at the soil surface, thus explaining the reduced importance of the radar metrics for
LFMC estimation as soil surface moisture may not linearly relate to vegetation water con-
tent, particularly for forest tree species [15]. Further, the inclusion of the so-called “hand-
crafted inputs”, ratios of microwave and spectral reflectance (or indices), did not provide
any improvements for LFMC retrieval as suggested in [23]. Such disagreement may be
related to the different vegetation types and spatial extents and thus increased variability,
spatial resolution (20 m vs 250 m), or to the inadequacy of such unconventional indices
hardly used in the remote sensing literature. Further research should be carried out to
validate the utility of such indices.
The importance of individual SI contributions for pooled models (all sites) confirmed
previous results [21,25] as three of the SIs selected here (EVI, NMDI, TCARI/OSAVI) were
also identified as the best-performing indices, either in their original or relative form, by
these previous studies suggesting similar spectral sensitivity to variation in LFMC over a
range of species and fire regimes. Although [25] suggested using relative SI and LFMC
values to increase model performance (i.e., they observed a decrease in the RMSE from
33% to 19%), such improvements were not observed with our data when using relative
indices alone. The discrepancy may be related to the estimation of LFMC over both shrub
and forest vegetation as opposed to only shrublands. Indeed, a more detailed analysis of
the variables’ importance by species revealed that relative indices were the most im-
portant for rockrose but not for oaks and pines, for which the original indices and struc-
tural variables, respectively, were more important. Overall, using both original and rela-
tive SIs resulted in the most accurate estimates, although the improvement of the RMSE
was rather marginal (from 17.9% to 13.4%, data not shown).
Individual models for each vegetation type performed slightly worse for rockrose
(MAE of 16.6%) but not for oak (10.4%) and pine (12.0%). Such differences were attributed
to the smaller dataset available for training at the species level (~500 observations). Indeed,
the use of random subsample (500 observations) from all sites and vegetation types for

Forests 2022, 13, 1846 13 of 17
LFMC estimation resulted in an MAE of 16.1% for the full model, a value similar to or
worse when compared to LFMC estimation at species level.
The use of predictor variables with high seasonality (e.g., day of the year), such as
those used in [22], may increase overall model precision but with negligible effect on
LFMC retrieval errors during higher fire risk periods. Indeed, estimation errors observed
when using static variables alone (St model) decreased when the day of the year (DOY)
was included (25% vs. 20%), while the explained variance doubled, reaching 53.6% (data
not shown). However, the inclusion of DOY in the full model improved performance only
marginally (e.g., MAE decreased by 0.6%), suggesting that remote sensing data can char-
acterize the LFMC variability along the year. Further, when estimating only low LMFC
values (<120%), the inclusion of DOY did not improve model performance, with MAE
changing by only 0.1%. Such differences suggest that seasonality is important for LFMC
prediction only when remote sensing information is not available and that LFMC esti-
mates based on static variables alone may be affected by larger errors.
Different patterns related to the predictor variables’ importance were observed for
rockrose and pine. For rockrose, the most important variables were the relative indices
which may be related to the higher interannual variability in shrublands and the potential
compensatory effect relativization has on such variations. In contrast, for pines, the static
variables were by far the most important ones, which may be related to more stable LFMC
values at the species level regardless of variation in environmental conditions [51].
Our results are slightly more accurate than those of [23,25] (RMSE of 20%–33%) and
similar to those from [22] (RMSE 16%–20% and MAE 13%–15%). When compared to [21]
(RMSE = 12.5%), our errors were higher, but the explained variance was significantly
lower (53.8% vs. 29%) despite the use of a much longer data series. One should note,
though, that like-for-like comparisons are difficult due to the different areal extents (con-
tinental vs. subcontinental vs. regional), the spatial resolution of estimates (250 m vs. 500
m vs. 20 m), and differences in predictor variables (e.g., day of year, land surface temper-
ature) and modelling approaches (e.g., random forest, recurrent neural network). Each of
these factors may influence the precision of the estimates. For example, LFMC estimation
over larger areas may result in an increased error as models need to cater to more species
and environmental conditions. Conversely, estimation errors at lower spatial resolution
(i.e., 250–500 m) improve when compared to higher (20 m) spatial resolution products due
to the increased correlation among spatial phenomena as areal unit size increases (i.e., the
modifiable areal unit problem), partly related to decreased variability caused by averag-
ing over many pixels [52].
Our study was limited by several factors, including the number of locations used for
in situ data collection. Although their spatial density was an order of magnitude higher
when compared to other studies [23,25], a limited number of locations may fail to com-
prehensively characterize spatial and temporal LFMC patterns within the entire region of
interest, which could affect any wall-to-wall mapping product. The relatively high pro-
portion (~30%) of in situ samples excluded from the analysis, coupled with the need to
establish a common dataset for the Sentinel-1 and Sentinel-2 imagery, has reduced the
number of samples available for model training and validation, which may have posi-
tively influenced the overall accuracy estimates as a drop in r and MAE was observed for
the ex situ analysis (Table 5 and Figure 5). Lastly, extrapolating LFMC from a few sampled
individuals to all trees within one pixel may increase uncertainties of the modeled rela-
tionships when sampled trees fail to represent the surrounding conditions accurately.
Such limitations may be partially addressed by using only the Sentinel-2 data or the Har-
monized Landsat Sentinel-2 (HLS) dataset [53] and thus increasing the number of samples
available for model training and validation. In addition, the operational LFMC content
predictions may be recalibrated on-the-fly with in situ data collected within the target
year.

Forests 2022, 13, 1846 14 of 17
5. Conclusions
This study is only the second assessing the utility of medium-resolution Sentinel-1
and Sentinel-2 sensors for LFMC retrieval in a typical Mediterranean environment, as well
as their synergy. We showed that machine learning techniques, particularly random for-
ests, may be used to estimate LFMC on a weekly basis in Mediterranean vegetation from
remote sensing data with acceptable model performance when input satellite variables
(optical) and ancillary information related to vegetation structure and site (static varia-
bles) are available. Model performance decreased when static variables were excluded,
but optical-based models were able to largely compensate for such shortcomings. The use
of static variables alone to estimate LFMC resulted in increased error estimates, particu-
larly over the lower LFMC ranges, with only 20.6% of variance explained when compared
to over 47% of the optical-based model and 54% of the combined optical-static-based
model. Such results suggest that spectral indices add unique information that cannot be
obtained from the static or the radar variables alone. In contrast with previous studies, the
information provided by the radar variables was not relevant nor unique in our region, as
removing such variables did not result in decreased model performance. However, our
results also showed that in areas of persistent cloud cover, radar-based models might be
used to estimate LFMC, albeit with lower accuracies.
Supplementary Materials: The following supporting information can be downloaded at:
https://www.mdpi.com/article/10.3390/f13111846/s1, Figure S1: Relationships between Sentinel 2
spectral indices (SI). Left panel shows the contribution of each absolute SI to the PCA axes. Right
panel shows the correlogram of absolute SIs; Figure S2: Figure S2 Relationships between Sentinel 1
polarizations acquired within the ascending (asc) and descending (des) satellite passes. The left
panel shows the contributions to the first PCA axes. The right panel shows the correlogram (VV—
vertical transmit vertical receive, VH—vertical transmit horizontal receive); Figure S3: Correlogram
of static variables (left panel) together with the utility as predictor variables (increase in mean
squared error, node purity) for LFMC estimation using random forests. Eastness and northness
were computed as the sine and cosine, respectively, of the aspect angle; Figure S4: Correlogram of
Sentinel-1 variables (left panel) together with their importance as predictor variables estimated as
mean decrease accuracy (i.e., increment in MSE when variable is excluded from the model) for
LFMC estimation using random forests; Figure S5: Correlogram of Sentinel-2 spectral indices (left
panel) together with their utility as predictor variables (increase in mean squared error, node purity)
for LFMC estimation using random forests. Only values above 0.7 are shown in the correlogram;
Figure S6: Distribution of residual errors from the full model (S1, S2, St) in relation to LFMC (left
panel) and day of year (right). Errors are colored by vegetation type.
Author Contributions: Conceptualization, E.M. and M.A.T.; methodology, E.M., M.A.T., J.P.G.N.,
and C.A.; formal analysis, J.P.G.N. and C.A.; data curation M.A.T., E.M., J.L.T., L.Y., J.M., C.H., and
M.G.; writing—original draft preparation, M.A.T., J.P.G.N., and E.M.; writing—review and editing,
C.A., J.M.; supervision, E.M. and M.T.; funding acquisition, E.M., M.A.T., and C.H. All authors have
read and agreed to the published version of the manuscript.
Funding: This work was funded by the Madrid regional government (grant CM/JIN/2021–024) and
the Spanish Ministry for Science and Innovation (grants PID2020–114062RA-I00, RYC-2017–22555
and RYC2018–024614-I, projects RTA2017–00042-C05–01, PID2020–116494RR-C41).
Informed Consent Statement: Not applicable.
Data Availability Statement: Sentinel-1 and Sentinel-2 data are freely available. The in situ data
was obtained from the Madrid regional government under a specific research agreement.
Acknowledgments: We are grateful to the firefighting service from Comunidad de Madrid for
providing the field data. We are also grateful to Carmen Díez (ICIFOR) for coordinating laboratory
works (sample managing, oven-drying and preliminary dataset).
Conflicts of Interest: The authors declare no conflict of interest. The funders had no role in the
design of the study; in the collection, analyses, or interpretation of data; in the writing of the manu-
script; or in the decision to publish the results.

Forests 2022, 13, 1846 15 of 17
Abbreviations
ARVI
atmospherically resistant vegetation index
ATCOR
atmospheric effects
CH
canopy height
CNES
Centre National D’Etudes Spatiales
DEM
digital elevation model
DOY
day of year
EVI
enhanced vegetation Index
FCC
forest fractional cover
GRD
ground range detected
GVMI
global vegetation moisture index
IW
interferometric wide
LFMC
live fuel moisture content
LID
location ID
MACCS
multi-sensor atmospheric correction and cloud screening
MAE
mean absolute error
MAJA
MACCS-ATCOR joint algorithm
max
maximum
MGRS
military grid reference system
min
minimum
MSE
mean squared error
MSI
multi-spectral instrument
MSIn
moisture stress index
NDII
normalized difference infrared index
NDMI
normalized difference moisture index
NDVI
normalized difference vegetation index
NDWI
normalized difference water index
NIR
near infrared
NMDI
normalized multi-band drought index
OOB
out of bag
OSAVI
optimized soil adjusted vegetation Index
OTB
Orfeo toolbox
PCA
principal component analysis
PNOA
national program of serial photogrammetry
RF
random forests
RMSE
root mean squared error
RVI
ratio vegetation index
RVI2
ratio vegetation index 2
S1
Sentinel-1
S2
Sentinel-2
SAR
synthetic aperture radar
SAVI
soil adjusted vegetation index
SI
spectral index
SLAVI
specific leaf area vegetation index
SMAP
soil moisture active passive mission
SRTM
shuttle radar topographic mission
St
static variables
SWIR
short wave infra red
TC_B
tasseled cap brightness
TC_G
tasseled cap greenness
TC_W
tasseled cap wetness
TCARI
transformed chlorophyll absorption index
VARI
visible atmospherically resistant index
VH
vertical horizontal
VIgreen
vegetation index-green
VV
vertical vertical

Forests 2022, 13, 1846 16 of 17
References
1. Carlson, J.D.; Burgan, R.E. Review of users' needs in operational fire danger estimation: The Oklahoma example. Int. J. Remote
Sens. 2003, 24, 1601–1620. https://doi.org/10.1080/01431160210144651.
2. Chowdhury, E.H.; Hassan, Q.K. Operational perspective of remote sensing-based forest fire danger forecasting systems. ISPRS
J. Photogramm. Remote Sens. 2014, in press. https://doi.org/10.1016/j.isprsjprs.2014.03.011.
3. Plucinski, M.P.; Anderson, W.R.; Bradstock, R.A.; Gill, A.M. The initiation of fire spread in shrubland fuels recreated in the
laboratory. Int. J. Wildland Fire 2010, 19, 512–520.
4. Anderson, S.A.J.; Anderson, W.R. Ignition and fire spread thresholds in gorse (Ulex europaeus). Int. J. Wildland Fire 2010, 19,
589–598.
5. Pimont, F.; Ruffault, J.; Martin-StPaul, N.K.; Dupuy, J.L. Why is the effect of live fuel moisture content on fire rate of spread
underestimated in field experiments in shrublands? Int. J. Wildland Fire 2019, 28, 127–137.
6. Jolly, W.M.; Johnson, D.M. Pyro-Ecophysiology: Shifting the Paradigm of Live Wildland Fuel Research. Fire 2018, 1, 8.
https://doi.org/10.3390/fire1010008.
7. Pivovaroff, A.L.; Emery, N.; Sharifi, M.R.; Witter, M.; Keeley, J.E.; Rundel, P.W. The Effect of Ecophysiological Traits on Live
Fuel Moisture Content. Fire 2019, 2, 28. https://doi.org/10.3390/fire2020028.
8. Chuvieco, E.; Riano, D.; Aguadi, I.; Cocero, D. Estimation of fuel moisture content from multitemporal analysis of Landsat
Thematic Mapper reflectance data: Applicatios in fire danger assessement. Int. J. Remote Sens. 2002, 23, 2145–2162.
9. Chuvieco, E.; González, I.; Verdú, F.; Aguado, I.; Yebra, M. Prediction of fire occurrence from live fuel moisture content
measurements in a Mediterranean ecosystem. Int. J. Wildland Fire 2009, 18, 430–441.
10. Caccamo, G.; Chisholm, L.A.; Bradstock, R.A.; Puotinen, M.L. Using remotely-sensed fuel connectivity patterns as a tool for fire
danger monitoring. Geophys. Res. Lett. 2012, 39, 1–5.
11. Ruffault, J.; Curt, T.; Martin-StPaul, N.K.; Moron, V.; Trigo, R.M. Extreme wildfire events are linked to global-change-type
droughts in the northern Mediterranean. Nat. Hazards Earth Syst. Sci. 2018, 18, 847–856. https://doi.org/10.5194/nhess-18-847-
2018.
12. Dimitrakopoulos, A.P.; Papaioannou, K.K. Flammability Assessment of Mediterranean Forest Fuels. Fire Technol. 2001, 37, 143–
152.
13. Pompea, A.; Vinesa, R.G. The influence of moisture on the combustion of leaves. Aust. For. 1966, 30, 231–241.
14. Park, I.; Fauss, K.; Moritz, M.A. Forecasting Live Fuel Moisture of Adenostema Fasciculatum and its Relationship to Regional
Wildfire Dynamics across Southern California Shrublands. Fire 2022, 5, 110.
15. Nolan, R.H.; Hedo, J.; Arteaga, C.; Sugai, T.; Resco de Dios, V. Physiological drought responses improve predictions of live fuel
moisture dynamics in a Mediterranean forest. Agric. For. Meteorol. 2018, 263, 417–427,
https://doi.org/10.1016/j.agrformet.2018.09.011.
16. Capps, S.B.; Zhuang, W.; Liu, R.; Rolinski, T.; Qu, X. Modelling chamise fuel moisture content across California: A machine
learning approach. Int. J. Wildland Fire 2022, 31, 136–148.
17. Yebra, M.; Chuvieco, E.; Riaño, D. Estimation of live fuel moisture content from MODIS images for fire risk assessment. Agric.
For. Meteorol. 2008, 148, 523–536, https://doi.org/10.1016/j.agrformet.2007.12.005.
18. Caccamo, G.; Chisholm, L.A.; Bradstock, R.A.; Puotinen, M.L.; Pippen, B.G. Monitoring live fuel moisture content of heathland,
shrubland and sclerophyll forest in south-eastern Australia using MODIS data. Int. J. Wildland Fire 2012, 21, 257–269.
19. Tanase, M.A.; Panciera, R.; Lowell, K.; Aponte, C. Monitoring live fuel moisture in semi-arid environments using L-band radar
data. Int. J. Wildland Fire 2015, 24, 560–572.
20. Marino, E.; Yebra, M.; Guillén-Climent, M.; Algeet, N.; Tomé, J.L.; Madrigal, J.; Guijarro, M.; Hernando, C. Investigating Live
Fuel Moisture Content Estimation in Fire-Prone Shrubland from Remote Sensing Using Empirical Modelling and RTM
Simulations. Remote Sens. 2020, 12, 2251. https://doi.org/10.3390/rs12142251.
21. Costa-Saura, J.M.; Balaguer-Beser, Á.; Ruiz, L.A.; Pardo-Pascual, J.E.; Soriano-Sancho, J.L. Empirical Models for Spatio-
Temporal Live Fuel Moisture Content Estimation in Mixed Mediterranean Vegetation Areas Using Sentinel-2 Indices and
Meteorological Data. Remote Sens. 2021, 13, 3726. https://doi.org/10.3390/rs13183726.
22. Cunill Camprubí, À.; González-Moreno, P.; Resco de Dios, V. Live Fuel Moisture Content Mapping in the Mediterranean Basin
Using Random Forests and Combining MODIS Spectral and Thermal Data. Remote Sens. 2022, 14, 3162.
https://doi.org/10.3390/rs14133162.
23. Rao, K.; Williams, A.P.; Flefil, J.F.; Konings, A.G. SAR-enhanced mapping of live fuel moisture content. Remote Sens. Environ.
2020, 245, 111797. https://doi.org/10.1016/j.rse.2020.111797.
24. Chuvieco, E.; Cocero, D.; Riaño, D.; Martin, P.; Martınez-Vega, J.; de la Riva, J.; Pérez, F. Combining NDVI and surface
temperature for the estimation of live fuel moisture content in forest fire danger rating. Remote Sens. Environ. 2004, 92, 322–331,
https://doi.org/10.1016/j.rse.2004.01.019.
25. García, M.; Riaño, D.; Yebra, M.; Salas, J.; Cardil, A.; Monedero, S.; Ramirez, J.; Martín, M.P.; Vilar, L.; Gajardo, J.; et al. A Live
Fuel Moisture Content Product from Landsat TM Satellite Time Series for Implementation in Fire Behavior Models. Remote Sens.
2020, 12, 1714. https://doi.org/10.3390/rs12111714.
26. Verbesselt, J.; Somers, B.; Lhermitte, S.; Jonckheere, I.; van Aardt, J.; Coppin, P. Monitoring herbaceous fuel moisture content
with SPOT VEGETATION time-series for fire risk prediction in savanna ecosystems. Remote Sens. Environ. 2007, 108, 357–368.

Forests 2022, 13, 1846 17 of 17
27. Yebra, M.; Chuvieco, E. Linking ecological information and radiative transfer models to estimate fuel moisture content in the
Mediterranean region of Spain: Solving the ill-posed inverse problem. Remote Sens. Environ. 2009, 113, 2403–2411.
https://doi.org/10.1016/j.rse.2009.07.001.
28. Yebra, M.; Dennison, P.E.; Chuvieco, E.; Riaño, D.; Zylstra, P.; Jr., E.R.H.; Danson, F.M.; Qi, Y.; Jurdao, S. A global review of
remote sensing of live fuel moisture content for fire danger assessment: Moving towards operational products. Remote Sens.
Environ. 2013, 136, 455–468.
29. Leblon, B. Monitoring forest fire danger with remote sensing. Nat. Hazards 2005, 35, 343–359. https://doi.org/10.1007/s11069-004-
1796-3.
30. Danson, F.M.; Bowyer, P. Estimating live fuel moisture content from remotely sensed reflectance. Remote Sens. Environ. 2004,
92, 309–321. https://doi.org/10.1016/j.rse.2004.03.017.
31. García, M.; Chuvieco, E.; Nieto, H.; Aguado, I. Combining AVHRR and meteorological data for estimating live fuel moisture
content. Remote Sens. Environ. 2008, 112, 3618–3627.
32. Fan, L.; Wigneron, J.P.; Xiao, Q.; Al-Yaari, A.; Wen, J.; Martin-StPaul, N.; Dupuy, J.L.; Pimont, F.; Al Bitar, A.; Fernandez-Moran,
R.; et al. Evaluation of microwave remote sensing for monitoring live fuel moisture content in the Mediterranean region. Remote
Sens. Environ. 2018, 205, 210–223. https://doi.org/10.1016/j.rse.2017.11.020.
33. Lu, Y.; Wei, C. Evaluation of microwave soil moisture data for monitoring live fuel moisture content (LFMC) over the
coterminous United States. Sci. Total Environ. 2021, 771, 145410. https://doi.org/10.1016/j.scitotenv.2021.145410.
34. Wang, L.; Quan, X.; He, B.; Yebra, M.; Xing, M.; Liu, X. Assessment of the Dual Polarimetric Sentinel-1A Data for Forest Fuel
Moisture Content Estimation. Remote Sens. 2019, 11, 1568. https://doi.org/10.3390/rs11131568.
35. Salas, W.A.; Ranson, J.K.; Rock, B.N.; Smith, K.T. Temporal and Spatial Variations in Dielectric Constant and Water Status of
Dominant Forest Species from New England. Remote Sens. Environ. 1994, 47, 109–119.
36. Frolking, S.; Milliman, T.; Palace, M.; Wisser, D.; Lammers, R.; Fahnestock, M. Tropical forest backscatter anomaly evident in
SeaWinds scatterometer morning overpass data during 2005 drought in Amazonia. Remote Sens. Environ. 2011, 115, 897–907.
37. Friesen, J.; Steele-Dunne, S.C.; Giesen, N.v.d. Diurnal Differences in Global ERS Scatterometer Backscatter Observations of the
Land Surface. IEEE Trans. Geosci. Remote Sens. 2012, 50, 2595–2602.
38. Saatchi, S.; Asefi-Najafabady, S.; Malhi, Y.; Aragão, L.E.O.C.; Anderson, L.O.; Myneni, R.B.; Nemani, R. Persistent effects of a
severe drought on Amazonian forest canopy. Environ. Sci. 2013, 110, 565–570.
39. Abbott, K.N.; Leblon, B.; Staples, G.C.; Maclean, D.A.; Alexander, M.E. Fire danger monitoring using RADARSAT-1 over
northern boreal forests. Int. J. Remote Sens. 2007, 28, 1317–1338. https://doi.org/10.1080/01431160600904956.
40. Bourgeau-Chavez, L.L.; Leblon, B.; Charbonneau, F.; Buckley, J.R. Evaluation of polarimetric Radarsat-2 SAR data for
development of soil moisture retrieval algorithms over a chronosequence of black spruce boreal forests. Remote Sens. Environ.
2013, 132, 71–85. https://doi.org/10.1016/j.rse.2013.01.006.
41. Saatchi, S.S.; van Zyl, J.J. Estimation of canopy water content in Konza Prairie grasslands using synthetic aperture radar
measurements during FIFE. J. Geophys. Res. Atmos. 1995, 100, 25481–25496.
42. Hill, M.J.; Donald, G.E.; Vickery, P.J. Relating Radar Backscatter to Biophysical Properties of Temperate Perennial Grassland.
Remote Sens. Environ. 1999, 67, 15–31.
43. Jolly, W.M. Sensitivity of a surface fire spread model and associated fire behaviour fuel models to changes in live fuel moisture.
Int. J. Wildland Fire 2007, 16, 503–509.
44. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32.
45. Baetens, L.; Desjardins, C.; Hagolle, O. Validation of Copernicus Sentinel-2 Cloud Masks Obtained from MAJA, Sen2Cor, and
FMask Processors Using Reference Cloud Masks Generated with a Supervised Active Learning Procedure. Remote Sens. 2019,
11, 433. https://doi.org/10.3390/rs11040433.
46. Hagolle, O.; Huc, M.; Villa Pascual, D.; Dedieu, G. A Multi-Temporal and Multi-Spectral Method to Estimate Aerosol Optical
Thickness over Land, for the Atmospheric Correction of FormoSat-2, LandSat, VENμS and Sentinel-2 Images. Remote Sens. 2015,
7, 2668–2691.
47. Hagolle, O.; Huc, M.; Pascual, D.V.; Dedieu, G. A multi-temporal method for cloud detection, applied to FORMOSAT-2, VENµS,
LANDSAT and SENTINEL-2 images. Remote Sens. Environ. 2010, 114, 1747–1755.
48. Inglada, J.; Christophe, E. The Orfeo Toolbox remote sensing image processing software. In Proceedings of the 2009 IEEE
International Geoscience and Remote Sensing Symposium, IGARSS, Cape Town, South Africa, 12–17 July 2009; p. IV-733.
49. Quegan, S.; Le Toan, T.; Yu, J.J.; Ribbes, F.; Floury, N. Multitemporal ERS SAR analysis applied to forest mapping. ITGRS 2000,
38, 741–753.
50. McCandless, T.C.; Kosovic, B.; Petzke, W. Enhancing wildfire spread modelling by building a gridded fuel moisture content
product with machine learning. Mach. Learn. Sci. Technol. 2020, 1, 035010. https://doi.org/10.1088/2632-2153/aba480.
51. Balaguer-Romano, R.; Díaz-Sierra, R.; De Cáceres, M.; Cunill-Camprubí, À.; Nolan, R.H.; Boer, M.M.; Voltas, J.; Resco de Dios,
V. A semi-mechanistic model for predicting daily variations in species-level live fuel moisture content. Agric. For. Meteorol. 2022,
323, 109022, https://doi.org/10.1016/j.agrformet.2022.109022.
52. Openshaw, S. The modifiable areal unit problem. Concepts Tech. Mod. Geogr. 1984, 38, 41.
53. Claverie, M.; Ju, J.; Masek, J.G.; Dungan, J.L.; Vermote, E.F.; Roger, J.-C.; Skakun, S.V.; Justice, C. The Harmonized Landsat and
Sentinel-2 surface reflectance data set. Remote Sens. Environ. 2018, 219, 145–161. https://doi.org/10.1016/j.rse.2018.09.002.