remote sensing

Article
Remote Sensing and Meteorological Data Fusion in Predicting
Bushfire Severity: A Case Study from Victoria, Australia
Saroj Kumar Sharma 1,2, *, Jagannath Aryal 1,2 and Abbas Rajabifard 1,2

1 Centre for Spatial Data Infrastructures and Land Administration, Department of Infrastructure Engineering,
Faculty of Engineering and IT, The University of Melbourne, Parkville, VIC 3010, Australia;
[email protected] (J.A.); [email protected] (A.R.)
2 Centre for Disaster Management and Public Safety (CDMPS), Department of Infrastructure Engineering,
Faculty of Engineering and IT, The University of Melbourne, Parkville, VIC 3010, Australia
* Correspondence: [email protected]; Tel.: +61-406821848

Abstract: The extent and severity of bushfires in a landscape are largely governed by meteorological
conditions. An accurate understanding of the interactions of meteorological variables and fire
behaviour in the landscape is very complex, yet possible. In exploring such understanding, we used
2693 high-confidence active fire points recorded by a Moderate Resolution Imaging Spectroradiometer
(MODIS) sensor for nine different bushfires that occurred in Victoria between 1 January 2009 and
31 March 2009. These fires include the Black Saturday Bushfires of 7 February 2009, one of the worst
bushfires in Australian history. For each fire point, 62 different meteorological parameters of bushfire
time were extracted from Bureau of Meteorology Atmospheric high-resolution Regional Reanalysis
for Australia (BARRA) data. These remote sensing and meteorological datasets were fused and
further processed in assessing their relative importance using four different tree-based ensemble
machine learning models, namely, Random Forest (RF), Fuzzy Forest (FF), Boosted Regression Tree



(BRT), and Extreme Gradient Boosting (XGBoost). Google Earth Engine (GEE) and Landsat images


were used in deriving the response variable–Relative Difference Normalised Burn Ratio (RdNBR),
Citation: Sharma, S.K.; Aryal, J.;
which was selected by comparing its performance against Difference Normalised Burn Ratio (dNBR).
Rajabifard, A. Remote Sensing and
Meteorological Data Fusion in
Our findings demonstrate that the FF algorithm utilising the Weighted Gene Coexpression Network
Predicting Bushfire Severity: A Case Analysis (WGCNA) method has the best predictive performance of 96.50%, assessed against 10-fold
Study from Victoria, Australia. cross-validation. The result shows that the relative influence of the variables on bushfire severity is
Remote Sens. 2022, 14, 1645. https:// in the following order: (1) soil moisture, (2) soil temperature, (3) air pressure, (4) air temperature,
doi.org/10.3390/rs14071645 (5) vertical wind, and (6) relative humidity. This highlights the importance of soil meteorology in
Academic Editors: Carmen Quintano
bushfire severity analysis, often excluded in bushfire severity research. Further, this study provides
and Quazi K. Hassan a scientific basis for choosing a subset of meteorological variables for bushfire severity prediction
depending on their relative importance. The optimal subset of high-ranked variables is extremely
Received: 18 January 2022
useful in constructing simplified and computationally efficient surrogate models, which can be
Accepted: 25 March 2022
particularly useful for the rapid assessment of bushfire severity for operational bushfire management
Published: 29 March 2022
and effective mitigation efforts.
Publisher’s Note: MDPI stays neutral
with regard to jurisdictional claims in Keywords: dimensionality reduction; dNBR; ensemble machine learning; bushfire severity; Google
published maps and institutional affil- Earth Engine; meteorological drivers; RdNBR; remote sensing; variable selection
iations.

1. Introduction
Copyright: © 2022 by the authors.
Licensee MDPI, Basel, Switzerland. Bushfires are frequently occurring natural phenomena experienced in various parts
This article is an open access article of the world including Australia, Mediterranean regions in Europe, and the United States
distributed under the terms and and play a key role in shaping the landscape and ecological dynamics [1–4]. Growing
conditions of the Creative Commons scientific evidence suggests that climate change is causing the increment in the scale, fre-
Attribution (CC BY) license (https:// quency, and severity of bushfires, posing a catastrophic threat to fire-prone areas including
creativecommons.org/licenses/by/ Australia [5–11]. This increment in disaster events has placed huge economic, social, psy-
4.0/). chological, emotional, and environmental costs on Australian people and society [8,12,13].

Remote Sens. 2022, 14, 1645. https://doi.org/10.3390/rs14071645 https://www.mdpi.com/journal/remotesensing

Remote Sens. 2022, 14, 1645 2 of 27

Given the frequency, severity, and impact of bushfires related to extreme climate events,
there is a societal need to investigate scientifically their cause and improve our understand-
ing to prevent and mitigate their effects as they unfold [14–16]. Nevertheless, the drivers of
bushfires and their influence during the fire are not yet fully understood, which has posed
challenges in implementing bushfire management and mitigation policies [16].
Bushfires depend on a range of biophysical and meteorological conditions of the
earth [17–20]. Of these, weather largely influences fire behaviour and governs the size,
intensity, speed, and predictability of bushfires [19,21]. Extreme weather conditions such
as severe drought, high temperature, low relative humidity, strong winds, etc. promote
bushfire-favourable conditions by increasing the rate of fuel production and the flammabil-
ity condition of live and dead fuel [20,22].
Most of the previous bushfire studies include only key meteorological variables, in-
cluding temperature, relative humidity, air pressure, and wind in bushfire analysis [23–26].
More importantly, the incorporated meteorological data in most of the existing research are
of coarse spatial and temporal resolutions, constraining the accuracy in assessing variable
influence. This has further limited the ability to capture dynamic spatiotemporal variability
in fire weather during bushfires. It is agreed that most of the literature have included the
majority of highly influencing variables including wind, temperature, humidity, and rain-
fall; however, they often lack a robust scientific basis for making these choices [18,25,27–30].
For example, Jenkins et al. [25] included temperature, precipitation, and solar radiation
as meteorological variables to model bushfire hazards. Oldenborgh et al. [18] mentioned
temperature, precipitation, humidity, and wind (speed and direction) as key meteorological
variables influencing bushfires. Similarly, Blanchi et al. [28] and Nolan et al. [29] focused
on maximum temperature, relative humidity, wind speed, and drought, considering them
as key contributors to bushfire events. Similarly, the widely used McArthur Forest Fire
Danger Index (FFDI) is based on temperature, wind speed, humidity, and drought factor
calculated by using antecedent precipitation and temperature [30].
Apart from the key meteorological variables, several other potential parameters that
could influence or improve the understanding of bushfire behaviour have been identi-
fied but were rarely used in previous studies. For example, a handful of studies have
demonstrated that different meteorological parameters including soil temperature and
moisture [31,32], surface flux [33,34], vertical wind [19], humidity and temperature [34],
measured at different vertical pressure isobar levels can improve understanding of bush-
fire behaviour; however, these variables are not often considered in bushfire studies. For
example, the effect of buoyancy can cause vertical displacement of air generating intense
winds influencing fire propagation [19]. Similarly, atmospheric stability—a measure of
buoyancy of the parcel of air determined by vertical air motion—is a crucial factor affect-
ing fire behaviour [35]. Even so, not all such parameters have been included in a single
comprehensive study, especially for assessing their relative influence on bushfire severity.
Such assessment and analysis using meteorological drivers measured at different vertical
levels would allow for a better understanding of fire behaviour. Hence, there is a need for
a comprehensive assessment of all potential meteorological parameters and to rank their
importance based on their influence on bushfire severity.
The overarching aim of this research is to investigate all potential meteorological
parameters and assess their relative importance in bushfire severity prediction. Recent
advancements in Earth Observation (EO) and in situ sensors provide an opportunity to
collect high-resolution space-time meteorological data [36], creating an environment for
scientific investigations. Similarly, environmental variables including a comprehensive list
of meteorological datasets for bushfire characterisation are often characterised by complex,
multicollinear, and high-dimensional data nature. To assess the relative importance of
variables with such complexity, Machine Learning (ML)-based variable selection models
are increasingly applied [37]. ML models are known for their robustness and high gener-
alisation capability and typically outperform traditional models (e.g., generalised linear
models) [38]; hence, they are widely applied for modelling high dimensional, complex
Remote Sens. 2022, 14, 1645 3 of 27

environmental problems [39,40]. More recently, ensemble machine learning (EML) tech-
niques, also called multiple classifier systems, are becoming popular due to their proven
effectiveness and versatility in a broad spectrum of real-world problems [39,41,42]. These
models produce accurate estimates by averaging rough predictions of weak learners rather
than finding a single high-accuracy predictor [41,43]. These relatively new approaches have
been successfully implemented in addressing a variety of problems including prediction
and variable selection. Of these EML techniques, tree-based models have been widely
used in variable selection because these models, unlike most of the well-established ML
models, are simple yet powerful for classification and regression problems [44]. Traditional
tree-based models utilise a single tree to make predictions, but the prediction outcome
based on a single tree is prone to different inaccuracies. However, relatively new tree-based
models, including Random Forest and Boosted Regression Trees, combine many simple
trees to form a powerful model and have optimised predictive performance [45,46].
In this study, we examined the relative influence of 62 different meteorological pa-
rameters (Table 1) as bushfires drivers and assessed them over their relative importance
in bushfire severity prediction. We implemented four tree-based different variable se-
lection models and identify the subset of the top 15 highest ranked parameters that
could potentially influence bushfire severity. In this research, we used the word vari-
able (i.e., meteorological variable) for denoting different weather factors, namely, wind,
temperature, precipitation, humidity, pressure, and moisture. Further, we used different
variations of these variables—for example, measurements at different vertical pressure
levels—during the analysis (Table 1), and we call them, collectively, parameters (most often
meteorological parameters).
Here, the analysis focuses on the nine bushfires in Victoria, Australia that occurred
between 1 January 2009 and 31 March 2009. Among them are the fires of 7 February 2009,
also known as Black Saturday Bushfires, one of the worst bushfire disasters in Australian
history [47]. We chose this bushfire in our study because of its huge extent, destructiveness,
and catastrophic loss incurred by that fire in terms of lives and properties, raising a
fundamental question about community bushfire safety in Australia [48]. There are a
few previous studies on various aspects of the Black Saturday bushfires [47,49,50]. For
example, Cruz et al. (2012) reported weather conditions, fuels, and fire propagation of
Black Saturday Bushfires of Kilmore east region [47]. Similarly, Cai et al. (2009) considered
the Black Saturday bushfires among 21 bushfires and attempted to develop an insight on
the effect of positive Indian Ocean Dipole events on Australian bushfires [49]. Likewise,
Kala et al. (2015) studied the influence of antecedent soil moisture focusing on the heatwave
event preceding the Black Saturday Bushfires [50]. However, unlike other studies, this
study not only focused on Black Saturday bushfires of 7 February 2009 but also considered
other fires in Victoria, Australia that occurred between 1 January 2009 to 31 March 2009.
There have been several studies implementing dimensionality reduction and variable
selection methods in different environmental phenomena including bushfires [23,25]. How-
ever, no study yet has assessed and ranked the most influencing variables against bushfire
severity as a target variable. Further, unlike existing studies, the ranking of bushfire severity
drivers in this research utilises a thorough list of meteorological variables recorded in high
spatial and temporal resolutions and in multiple vertical hierarchies, which is the novelty
of this research. Considering the 2009 bushfires of Victoria as a case study, in this research,
we ask the following questions and attempt to answer them:
1. What are the most important meteorological variables and their relative influence on
bushfire severity prediction?
2. What is the predictive performance capability of the different ensemble machine
learning models?
3. What management and policy recommendations can be synthesised from the research
outcomes and transformed to community wellbeing?
More specifically, the main contributions of this study are as follows:
Remote Sens. 2022, 14, 1645 4 of 27

1. This is the first work to our knowledge that has performed a thorough analysis of
62 meteorological parameters (including humidity, temperature, and wind in multiple
vertical isobar levels) of high spatial resolution and temporal frequency to quantify
the relative influence of variables in bushfire severity prediction.
2. A comparative assessment of predictive performances of widely used machine learning
models on handling complex, high-dimensional, multicollinear meteorological data.
3. Improve understanding of bushfire-severity-influencing variables that help formulate
better bushfire management and suppression strategies.
The rest of the paper is organised as follows: In Section 2, we explain the data and
methods including meteorological data, variables, and bushfire locations. Section 3 presents
the results of the variable selection analysis and performance outcome of different models.
Section 4 discusses the investigative results obtained from our experiment. Finally, Section 5
concludes the paper with potential future research directions.

2. Materials and Methods

2.1. Study Area
This study area consists of the regions of Victoria, Australia (Figure 1) that experienced
bushfires between 1 January 2009 and 31 March 2009, where the first ignition occurred on
29 January 2009 (Table 2). The study region includes 9 different fires including the fires
in the Beechworth area, Bendigo area, Kinglake–Marysville area, Dargo area, Churchill
Complex area, Bunyip state park area, and Wilson Promontory area of Victoria. The study
area also includes fires of the Kilmore–Murrindindi region, Victoria, also colloquially known
as the Black Saturday bushfires. The Black Saturday bushfires were among Australia’s
all-time worst bushfire disasters with the highest fatalities of 173 and Kilmore East fire
was the most significant of these fires, accounting for 70% of total fatalities on the day and
burning 100,000 hectares in less than 12 h [47]. The study area encompasses a wide range
of vegetation types, from grassland and woodland (lower elevation), tall wet-sclerophyll
eucalypt forest (upper elevation), to temperate rainforest (localised gullies) [47]. The area
has a cool and temperate climate with some microclimatic variations due to the diverse
vegetation and topography [51]. Cruz et al. (2012) described in detail the climatic variability
and fire weather of the area experienced during the 2009 Black Saturday bushfire period [47].

2.2. Fire Data

We obtained the spatial extent of bushfires prepared by the Victoria government’s
Department of Environment, Land, Water and Planning (DELWP), which has delineated the
bushfire boundary primarily on the public land. The data show that 9 different fires with
different ignition dates (Table 2) occurred in the study period—i.e., between 1 January 2009
and 31 March 2009—within our study area. Hence, for the same fire period, we obtained
the active fire point data recorded by the Moderate Resolution Imaging Spectroradiometer
(MODIS) sensor. MODIS sensors can detect fire pixels with unparalleled accuracies and
have been demonstrated in different previous studies [52–54]. An alternative approach for
delineating bushfire boundary is to aggregate the MODIS active fire clusters or burned area
events as demonstrated by Briones-Herrera et al. (2020) and Laurent et al. (2018) [55,56].
This approach can be very helpful to extract fire perimeter information in near-real-time or
to verify historical burnt extent. However, in this study, we utilised fire boundary maps
prepared by DELWP. These maps were prepared by interpreting satellite images and aerial
photos. Yet, we visually inspected and verified that the active fire points by MODIS are
within the fire boundary delineated by DELWP. This further ensures a high degree of
certainty that the utilised active MODIS fire points accurately detected the fires. This gives
a total of 2693 active bushfires points recorded by MODIS on different days within the
study period. The meteorological readings of all these fire points corresponding to the
recorded time by MODIS sensors were extracted from Bureau of Meteorology Atmospheric
High-resolution Regional Reanalysis for Australia (BARRA) data. These 2693 active fire
data with corresponding meteorological readings were utilised for further analyses.
Remote Sens. 2022, 14, 1645 5 of 27

Figure 1. Topographic map of the study area (Victoria) showing bushfire points of 9 different fires
used in the analysis. These points are active bushfire points recorded by Moderate Resolution Imaging
Spectroradiometer (MODIS) sensors.

2.3. Meteorological Data

This research uses the Bureau of Meteorology Atmospheric High-resolution Regional
Reanalysis for Australia (BARRA) data for meteorological variables. This dataset was
released completely in July 2019, specifically developed for Australia, and is the first
of its kind for the Australian region [57]. The data offer higher resolution in space and
time compared with existing global reanalysis products [58]. The data of our study area,
BARRA-R, have a 12-km spatial resolution and varying temporal resolutions, hourly
to 6-hourly, depending on the meteorological parameter. We converted all the input
variables to a 6-hourly frequency for the analysis to maintain temporal consistency in all
input parameters.
The analysis includes 62 different meteorological parameters extracted from BAARA_R
data (Table 1) for each active fire point collected from the MODIS sensor. These meteorolog-
ical parameters can be broadly categorised into seven different categories of meteorological
variables, i.e., precipitation, temperature, pressure, wind, humidity, surface flux, and soil
moisture. Precipitation includes the sum of large-scale and convective rainfall and snowfall
at the surface. Upward air velocity, relative humidity, and temperature are included for
11 different vertical levels determined by pressure isobars, i.e., 1000 hPa, 975 hPa, 950 hPa,
925 hPa, 900 hPa, 850 hPa, 800 hPa, 750 hPa, 700 hPa, 600 hPa, and 500 hPa. Similarly, soil
moisture and soil temperature are included for 4 different depth levels: 0.05 m, 0.225 m,
0.675 m, and 2 m. Soil moistures include frozen and unfrozen contents available for land
points. A detailed list of the data with a brief description of the parameters is presented
in Table 1.
Remote Sens. 2022, 14, 1645 6 of 27

Table 1. List of meteorological variables and parameters included in the analysis. The level column
represents the number of vertical levels data collected. Single-level (S) data are measured at 1.5 m
(unless specified otherwise). Multilevel data (M) represent 11 vertical levels determined by pressure
isobars (1000 hPa, 975 hPa, 950 hPa, 925 hPa, 900 hPa, 850 hPa, 800 hPa, 750 hPA, 700 hPa, 600 hPa,
500 hPa) and multidepth data (D) include measurements from 4 different depth levels (0.05 m,
0.225 m, 0.675 m, and 2 m).

Variable Parameter Data name Level Unit

6-hourly average of hourly mean atmospheric specific humidity AvQsair S kg kg−1
6-hourly average of instantaneous 10 min specific humidity QsairScrn S kg kg−1
Humidity 6-hourly average of hourly instantaneous relative humidity RelHumPrs M %
6-hourly average of instantaneous specific humidity at the end of the
SpecHum S Kg kg−1
hourly timestep
Moisture 6-hourly average of hourly instantaneous soil moisture SoilMois D Kg m−2
Precipitation Total precipitation amount at the surface AccPrcp S Kg m−2
6-hourly average of hourly mean air pressure AvMslp S Pa
6-hourly average of instantaneous 10 min air pressure MslpSpec S Pa
Pressure
6-hourly average of instantaneous 10 min air pressure at 20 m height PresSpec S Pa
Surface pressure at the end of the timestep SfcPres S Pa
6-hourly average of hourly instantaneous air temperature (multi-levels) AirTemp M Kelvin
6-hourly average of hourly maximum temperature MaxTemp S Kelvin
6-hourly average of hourly minimum temperature MinTemp S Kelvin
Temperature 6-hourly average of instantaneous 10 min surface temperature (topsoil
SfcTemp S Kelvin
layer) at the end of the timestep
6-hourly average of hourly instantaneous soil temperature (multidepth) SoilTemp D Kelvin
6-hourly average of instantaneous 10 min temperature at 1.5 m TempScrn S Kelvin
6-hourly average of hourly mean surface upward latent heat flux AvLatHflx S Wm−2
6-hourly average of hourly mean surface upward sensible heat flux AvSenHflx S Wm−2
Surface Flux
6-hourly average of hourly instantaneous surface upward total
AvSfc S kg m−2 s−1
moisture flux
6-hourly average of hourly mean of wind U component at 10 m AvUwnd10 m S ms−1
6-hourly average of hourly mean of wind V component at 10 m AvVwnd10 m S ms−1
6-hourly average of hourly mean of wind speed of gust at 10 m WndGst10 m S ms−1
Wind 6-hourly average of instantaneous 10 min U component of wind at 10 m Uwnd10 m S ms−1
Upward air velocity (multilevel) VwndPrs M ms−1
6-hourly average of instantaneous 10 min V component of wind at 10 m Vwnd10 m S ms−1
6-hourly average of instantaneous 10 min wind speed of gust at 10 m Wgust10 m S ms−1

2.3.1. Meteorological Data Extraction

The meteorological data (BARRA_R) are available in the multiple Network Com-
mon Data Form (NetCDF), which is a data abstraction format for storing and retrieving
multidimensional data [59]. We first determined the fire recorded time in the MODIS
active bushfire points and extracted the corresponding meteorological information for all
62 parameters.

2.3.2. Temporal Frequency

We divided days into four equal 6-hourly intervals, 00:00 h, 06:00 h, 12:00 h, and
18:00 h and rounded the data acquisition time of active fire points to these nearest 6-hourly
timestamps. Most of the variables in the meteorological BARRA data are available in
the defined 6 h timestamp, except precipitation and temperature, which are available for
every hour. Depending on the structure of the dataset of the variable, we either sum or
average the hourly parameters and convert them into 6 h frequency. For the accumulation
of precipitation, as we are measuring total precipitation over the 6 h interval, we sum the
amount of rain measured within the 6 h window. For example, accumulation precipitation
for 18:00 h is the sum of precipitation recorded at 13:00 h, 14:00 h, 15:00 h, 16:00 h, 17:00 h,
Remote Sens. 2022, 14, 1645 7 of 27

and 18:00 h. For the temperature, we computed 6-hourly readings by averaging the
hourly measurements.

2.4. Target Variable–Bushfire Burn Severity

We used bushfire burn severity index as the target variable for the analysis. The burn
severity indices have been utilised as a target variable in different previous studies [2,60].
Traditionally, the burn severity assessment was performed by field-based observations
including estimation of tree mortality, fuel consumption, soil and rock cover changes,
char height, and vegetation burn or scorch at different strata (substrates, subcanopy, and
upper canopy) [61]. However, in recent decades, remote-sensing-based multitemporal
analysis methods are widely utilised in estimating burn severity indices [62–65]. Among
the different severity indices, Difference Normalised Burn Indices (dNBR) and Relative
Difference Normalised Burn Indices (RdNBR) are widely utilised and have demonstrated
better accuracies in several studies [62,66,67]. Further, several alternatives or variations
of dNBR are proposed and many of them have demonstrated improved accuracies. For
example, Park et al. (2014) proposed a Landsat-based metric—Relativised Burn Ratio
(RBR)—as an alternative to dNBR and RdNBR [65]. Further, Park et al. (2014) demonstrated
that RBR better corresponded to ground truth data and generated higher classification
accuracy compared with dNBR and RdNBR in the study of mixed broadleaf–coniferous
forests in western USA. We still limit this study to the most widely used indices, dNBR and
RdNBR; however, we also recommend and consider other severity metrics including RBR
when examining them in future studies. In this study, we first performed a comparative
performance assessment of dNBR and RdNBR indices against field-collected validation
data. Those indices producing better accuracy were utilised as a target variable for further
assessment. The processing was performed with the Google Earth Engine (GEE) platform
(Figure 2).

Figure 2. Methodological flowchart for generating bushfire burn severity target variable. The dNBR
and RdNBR indices are produced by computing and applying phenology correction offset following
Parks et al. (2018). The severity index giving higher accuracy against ground truth data is further
prepared and utilised as the target variable.
Remote Sens. 2022, 14, 1645 8 of 27

The burn severity index is generated by multitemporal change analysis through a

difference in Normalised Burn Ratio (NBR) calculated by using shortwave-infrared (SWIR)
and near-infrared (NIR) bands of prefire and postfire imageries [66,68,69].
The NBR and the difference (dNBR) are calculated as follows:

NIR − SWIR
NBR = (1)
NIR + SWIR
dNBR = NBRprefire − NBRpostfire (2)
In Landsat TM, the NIR band has a wavelength of 0.77–0.90µm and is represented by
band 4. Similarly, the SWIR band represented in band 7 has a wavelength of 2.09–2.35 µm.
Similarly, the RdNBR is calculated by generating the Normalised Burn Ratio of prefire
and postfire as below [70]:

NBRprefire − NBRpostfire
RdNBR = (3)
√
NBRprefire

The details of the image tiles and the acquisition dates utilised for calculating dNBR
and RdNBR severity indices in this study are presented in Table 2. Before extracting
the corresponding severity value for each bushfire point, the generated burn severity
rasters, dNBR and RdNBR, were smoothed by using a 3 × 3 pixels majority filter. This
method helps to eliminate major speckles or sporous data and enhance features otherwise
not visibly apparent to the data by assigning the majority value within a 3 × 3 window
to the central pixel. However, the smoothed data were only utilised for assessing the
relative performance of the indices against ground truth data. Further, aggregation of the
severity index matching meteorological data resolution was performed before using it as a
target variable. The aggregation of severity index with matching meteorological resolution
ensures a better representation of severity data for analysing the relative influence of
meteorological variables.

2.4.1. Approach to Multiscale Data Integration

Scale discrepancy is a known problem in Remote Sensing and prevails especially
during the integration of multiresolution data. In this research, the burn severity index
prepared using Landsat data of 30 m does not match the spatial resolution of 12 km me-
teorological pixel (Figure 3). Although we smoothed the outcome by applying a filter of
size 90 × 90 m for validating accuracy against ground truth data, we only utilised this
information to choose the burn severity index with better accuracy (RdNBR) for further
analyses. To address the scale issue in assessing the relative importance of meteorological
variables, we aggregated the severity index (RdNBR) to match the spatial resolution of the
explicative meteorological variables (12 × 12 km). For this, we calculated the percentile of
the index (RdNBR) by each climatic pixel and used the information to classify the severity
class of that pixel. Percentile approach has been used in extracting the pixels in fusing
multisensor imagery including Landsat and MODIS. For example, Holsinger et al. (2021)
implemented a percentile-based approach to extract snow-free images from MODIS by
assigning percentile values to snowmelt days [71]. A similar approach was used to up-
scale phenology data and demonstrated superior performance while addressing upscaling
problems in remote sensing [72,73]. Among different percentile values, Zhang et al. (2017)
mentioned that at least the 30th percentile value in high-resolution data was required in
upscaling for capturing phenological information. They further demonstrated that the 30th
percentile value resulted in the best accuracy in generating coarser resolution data in a
heterogeneous area [72], which is in line with the heterogeneity of the study area of this
work. Hence, we adopted the 30th percentile value and used it as the cutoff threshold.
Furthermore, we averaged the severity value of the remaining (70%) pixels to assign the
new severity value of the aggregated pixels. The severity of the pixel is determined by
Remote Sens. 2022, 14, 1645 9 of 27

averaging the top 70% of the pixel severity and, thus, preventing dilution of the pixel values
by eliminating low or unburnt patches with a relatively smaller area. We observed that
these small patches are due to topographical or other factors rather than meteorological
influence. Thus, the obtained burn severity index in the aggregated map is used as a target
variable for assessing relative meteorological influence in bushfire severity to ascertain the
representative samples and their utilisation in further analyses.

Figure 3. An example of aggregated data. (a) RdNBR index calculated using 30-m Landsat TM data;
(b) RdNBR pixels aggregated to match spatial resolutions of meteorological pixels.

2.4.2. Remote Sensing Data

This study used multitemporal Landsat 5 TM data of the study area to generate
severity indices. Although this research utilised MODIS active fire points for extracting
corresponding measurements of meteorological variables, we used Landsat images over
MODIS for the bushfire severity assessment due to their higher spatial resolution. We
consider Landsat images with a higher spatial resolution (30 × 30 m) than MODIS images
(1 × 1 km) to be more suitable to characterise the severity variations of the bushfires in
a heterogonous hilly landscape. Again, these data were only utilised to assess relative
performance between dNBR and RdNBR indices against field-collected ground truth data.
Further, upscaling of the best performing index is performed to match climatic resolution for
using it as a target variable, as described in Section 2.4.1. Further, we carefully considered
the date while selecting the tiles for prefire and postfire timestamps. We attempted to
keep them closer to the bushfire dates with minimal seasonal variations to minimise the
phenological effect. However, due to cloud cover, it was not always possible to obtain
images of the desired date throughout the study area.
Hence, to tackle the phenological effect and the effect of precipitation between prefire
and postfire imagery, we applied correction by calculating dNBRoffset. Different studies
have demonstrated significant improvement in bushfire severity results upon applying
the offset [74,75]. We adopted the method implemented by Parks et al. (2018), where the
offset value is calculated by averaging the dNBR value across all pixels located 180 m
outside of the delineated bushfire boundary. The RdNBR index was produced by using
phenology-corrected dNBR imagery.
Further, the prefire and postfire images are within a year of the date of burning, as
suggested by Key and Benson (2006) [61]. The only image with more than a year difference
is the postfire image of fire 2, which is within 2 years after the fire and still meets the
recommendation by Key and Bension (2006). This was because the cloud-free image of the
Remote Sens. 2022, 14, 1645 10 of 27

area for the preferred date was not always available. The details of the Landsat TM tiles
used for bushfire severity prediction of different fires are presented in Table 2.

Table 2. Prefire and postfire Landsat 5 TM image tiles used for bushfire severity prediction of different
fire events with different ignition dates under study.

Fire Event Fire Ignition Date Acquisition Date (Prefire) Acquisition Date (Postfire)
1 7/02/2009 4/10/2008 23/10/2009
2 7/02/2009 3/11/2008 8/10/2010
3 6/02/2009 4/10/2008 23/10/2009
4 8/02/2009 4/10/2008 8/11/2009
5 4/02/2009 13/01/2008 1/12/2009
6 29/01/2009 4/10/2008 8/11/2009
7 7/02/2009 13/01/2008 1/12/2009
8 7/02/2009 13/01/2008 1/12/2009
9 7/02/2009 27/10/2008 30/10/2009

2.4.3. Bushfire Severity Classification

The generated burn severity raster layers after phenology correction, represented by
dNBR and RdNBR, are classified into (1) unburned, (2) low severity with no crown scorch,
(3) moderate severity with understory burn and light crown scorch, (4) high severity with
crown scorch, and (5) very high severity with crown burn. Different studies adopted different
threshold values for defining severity categories; we adopted the one used by Cai and
Wang (2020) [76]. The classification threshold defined for each class is defined in Table 3. The
severity value of dNBR is scaled to 1000, as recommended by Key and Benson (2006) [61].

Table 3. Burn severity category and classification threshold for dNBR and RdNBR raster layers. The class
value of dNBR is scaled to 1000. The classification threshold was adopted from Cai and Wang (2020).

Severity Category Class Value (dNBR) Class Value (RdNBR)

Unburnt <100 <0.1
Low 100–150 0.1–1.5
Moderate 150–250 1.5–2.5
High 250–350 2.5–3.5
Very high ≥350 ≥3.5

2.4.4. Bushfire Severity Accuracy Assessment

Accuracy assessment was performed using a total of 620 ground truth validation data
points. The data were collected in the field for 2009 Victorian bushfire severity mapping by
the Victoria government’s Department of Environment Land, Water and Planning (DELWP).
The severity assessment of field data is prepared by estimating the proportion of areas
burnt or scorched due to bushfires over different plots [77]. Further, the measurement also
estimated changes in soil cover, bark, charcoal height, etc. for each plot. The validation
data are categorised into five different severity classes: (1) very high severity (crown burn),
(2) high severity (crown scorch), (3) moderate severity (moderate crown scorch), (4) low
severity (light or no crown scorch), and (5) no crown scorch or understorey burnt [77]. The
accuracy assessment was performed for both severity indices, dNBR and RdNBR, against
all five severity classes. A metric-based performance assessment providing overall accuracy,
kappa coefficient, and user’s and producer’s accuracy was used.

2.4.5. Processing Platform—Google Earth Engine

The severity index was prepared based on spectral indices of multispectral satellite
imagery utilising Landsat 5 Thematic Mapper images in the Google Earth Engine (GEE)
platform. GEE is a cloud-based geospatial analysis platform with very high computational
capabilities designed for storing and processing very large geospatial datasets [78,79]. GEE
Remote Sens. 2022, 14, 1645 11 of 27

datasets consist of Earth Observing remote sensing images including the entire Landsat
archive, Sentinel-1, and Sentinel-2, which were already preprocessed for fast and efficient
access. The bushfire burn severity analysis was implemented through the scripts written in
Earth Engine Application Programming Interface (API) and processed in Code Editor, a
web-based IDE for the Earth Engine.

2.5. Methods for Variable Selection

2.5.1. Multicollinearity Test
The choice of variable selection model depends on the nature of data, which can be
tested by multicollinearity analysis. It is performed by calculating the Variance Inflation
Factor (VIF) and tolerance (TOL) indices. VIF measures the inflation of the variance of
the estimated regression coefficient due to the correlation between the predictors in the
model [80,81]. By taking the square root of VIF, it can be determined how big the standard
error is in comparison with what it would be if that variable is not correlated with the other
explanatory variables [81]. VIF is calculated as follows:

1 1
VIF = 2
= (4)
1−R TOL
Similarly, the tolerance is calculated as follows:

TOL = 1 − R2 (5)

where R2 is the coefficient of determination for the regression of predictor variable of

the model [80]. Although there is no formal cutoff threshold for tolerance indicating
multicollinearity, the value close to 0 suggests that there exists strong multicollinearity,
whereas there exists little or no multicollinearity if the value is close to 1. Similarly, the
VIF > 10 is often considered as indicating multicollinearity.

2.5.2. Variable Selection Models

In this study, four tree-based ensemble machine learning models—namely, Random
Forest (RF), Boosted Regression Tree (BRT), Extreme Gradient Boosting (XGBoost), and
Fuzzy Forest (FF)—were implemented to select important meteorological variables of
bushfire severity (Figure 4). All four models were utilised for modelling nonlinear rela-
tionships in complex environmental phenomena including bushfires [21,42,82–87]. All
these models can be used for both classification and regression data and, in most cases,
they produce a robust outcome. Among them, RF and BRT [45] are widely popular and
have been extensively used in environmental applications over the last decades [42,88,89],
whereas XGBoost [90] and FF [91] are relatively new models. However, XGBoost has been
increasingly used in recent years especially due to its accuracy, scalability, and operative
efficiency. For example, unlike many variable selection models, XGBoost is capable of
utilising CPU for multithreaded calculations [92]. Similarly, FF, although not yet widely used,
is specially designed for dealing with complex, high-dimensional, correlated features and has
successfully demonstrated its capability in dealing with complex environmental data [86].
Random Forest, first introduced by Breiman in 2001 [93], is a bagged decision tree
model that has received increasing attention due to its higher-accuracy results and process-
ing speed compared with other single tree models. It uses a randomly selected subset of
variables and training samples and produces multiple decision trees; results from each tree
are aggregated for the prediction outcome [94]. We tuned the parameters by setting up
different values during the implementation of this method. The value for the number of
forest-grown was set to 25 for the interpretation and prediction step, whereas value 50 was
used for the thresholding step. For each forest, the number of trees grown was set to 3000.
Remote Sens. 2022, 14, 1645 12 of 27

Figure 4. Tree-based ensemble machine learning models implemented in this study for selecting
meteorological variables influencing bushfire severity.

The BRT model combines the strength of two algorithms, regression trees and boosting,
and can automatically handle the interaction between predictors. This model provides
the variable ranking based on relative influence (or contribution) of predictor variables
with their sum adding to 100 [88]—the higher the rank, the stronger the influence. The
ranking process depends on how many times a variable is selected for splitting, weighted
by squared improvement to the model due to each split and averaged over all trees [45,95].
In this research, we used tree complexity of 5, learning rate 0.004, bag fraction 0.75, maxi-
mum trees 10,000, ntrees 100 while fitting the model. The outcome was tested using the
10-fold cross-validation method. For further explanation of the parameters, please refer to
Elith et al. (2008) [45].
XGBoost is a decision-tree-based gradient boosting model designed for high scalabil-
ity [90]. This model, similar to other gradient boosting models, minimises the loss function
and builds an additive expansion of the objective function. Further, to identify the best split,
XGBoost proposes a sparsity-aware algorithm and automatically removes zero and/or
missing values from the computations of gain for split candidates [96]. This algorithm
centres on reducing the complexity in computation to identify the best split and implements
different models to improve the training speed of decision trees [90,96]. We tuned different
XGBoost parameters while running the model. The maximum depth of each tree was set
to 5, and the training step for each iteration to 0.1.
FF algorithm is an extension of random forests specially designed to minimise the bi-
ased feature ranking in the presence of highly correlated features in high-dimensional data.
Hence, it utilises the strengths of RF as well as accounts for the problem of correlation [91].
Although FF is specially designed for high-dimensional data, it has also demonstrated
performance in low-dimensional data. In FF, features are clustered into a different group
called modules such that correlations between the modules are low but, at the same time,
correlation within the modules is high. It uses Recursive Feature Elimination–Random
Forest model within the modules and retains the top prespecified fraction of features from
each module [91,97]. Surviving features are combined in one dataset and the process is re-
peated to select the features. Within Fuzzy Forest, we utilised Weighted Gene Coexpression
Network Analysis (WGCNA) to determine the cluster of correlated features. This model
constructs a network of predictors, defines similarity measures, and uses a hierarchical
clustering approach to identify the modules. While the WGCNA can be applied to the
most high-dimensional dataset, it is widely used in system biology [98] but, to our knowl-
edge, it has not been applied in assessing the environmental variables yet. The variable
importance was generated using different optimised control parameters in R studio. For
WGCNA control, we used power value 6 with a minimum module size of 10. Similarly,
we set up select control to drop fraction = 0.25, mtry factor 1, the number selected 20, keep
Remote Sens. 2022, 14, 1645 13 of 27

fraction 0.25, minimum ntree 5000, and ntree factors 5. A detailed explanation of the Fuzzy
Forest algorithm and used parameters can be found in Conn et al. (2019) [91].

2.5.3. Accuracy Assessment Methods

The accuracy assessment of the methods is performed by comparing the response
variable against the classified burn severity index, calculated by aggregating RdNBR pixels
matching meteorological data resolution as described in Section 2.4.1. For the model
performance assessment, the response variable is classified in binary classes and coded
with values of 1 for high and very high burn severity and 0 for low burn severity [2].
We used a training–testing split of 70:30% and performed an accuracy assessment of the
models by validating the predictive outcomes using the widely adopted k-fold cross-
validation method [99]. A typical choice of k is between 5 and 10 [99]. We used 10-fold
cross-validations in all the models. The training set represents a randomly sampled fraction
of the data on which our models are trained. The performance is then assessed based on
predictive accuracy on a separate validation set.
The cross-validation Mean Squared Prediction Error (MSPE) was computed by using
the following equation [100]:
∗
∑n (Yi − Y 0 i ) 2
MSPE = i=1 (6)
n∗
where Yi is the observed response for item i and Y 0 is the predicted response of the model
based on training data; the sum is calculated from validation data over n∗ observations.

2.6. Method Implementation

The implementation of this research utilised different software and programmatic
approaches including ArcGIS, R language in RStudio, and Python in Jupyter Lab. The
meteorological data obtained from the Bureau of Meteorology (BOM) are stored in the
network Common Data Form (netCDF) format. Considering the computational complexity
of netCDF data, we created and implemented python-3.7-based netCDF data extraction
scripts in a High-Performance Computing (HPC) facility, SPARTAN [101], running under
the Linux operating system. The netCDF data extraction script used different python pack-
ages including Xarray for handling multidimensional data. Similarly, the fire location data
are obtained from MODIS. The Fire Severity index is generated using Google Earth Engine
and the product is analysed and mapped in ArcGIS 10.8. The explorative visualisation
of the extracted data is carried out in Jupyter Lab v.2.2.6 using R v.4.0.4 programming
language. Variable selection models, Random Forest, Fuzzy Forest, BRT, and XGBoost are
implemented in R v.4.0.4 using RStudio v.1.1.456.

3. Results
3.1. Multicollinearity Analysis
We carried out a multicollinearity analysis to understand the relationships between
the variables. The calculated variance inflation factor (VIF) and tolerance (TOL) value show
that 51 of 62 variables have VIF greater than 10 and Tolerance less than 0.1, indicating strong
multicollinearity between the variables [81]. Variables with the highest multicollinearity
include air pressure and temperature at different vertical pressure levels, latent heat flux,
surface upward moisture flux, and relative humidity near the surface. The variables with
the least multicollinearity (VIF < 10 and Tol < 0.1) include precipitation, soil moisture,
vertical wind, soil temperature, surface upward sensible heat flux, and relative humidity
above 850 hPA pressure level. Although the results observed multicollinearity among the
variables, we still included all variables in the variable selection analysis. This is because
the models we implemented can deal with the multicollinear nature of the data or at
least provide a useful basis for interpretation of the results serving our purpose [45,91].
Further, the information of VIF can provide readers with additional insights on the asso-
Remote Sens. 2022, 14, 1645 14 of 27

ciated multicollinearity between the variables that helps better understand and interpret
the results.

3.2. Target Variable

We performed the bushfire severity assessment of the area by calculating dNBR and
RdNBR indices in the Google Earth Engine platform. The map shows a good correspon-
dence between these indices as they both produce close results in burn variability for each
severity class across the study area (Figure 5).

Figure 5. Bushfire severity maps of Kilmore-east–Murrindindi region Victoria. The burn severity is
calculated by using dNBR and RdNBR index. The severity classes are categorised into very high,
high, moderate, and low severity classes as per the classification threshold presented in Table 3.

The classification accuracy assessment was performed by comparing the classified

data with the 620 ground truth data. The confusion matrix is generated for all classes for
phenology-corrected dNBR and RdNBR indices.
The results show that the overall accuracy of RdNBR and dNBR are 90.0% and 86.6%
and the kappa statistics of 0.86 and 0.81, respectively (Table 4). Further, we consistently
observed that the accuracy of the RdNBR is higher for very high severity classes, mostly in
those areas with heterogenous vegetation distribution (Figure 6). For most of the classes,
RdNBR has better user and producer accuracy than dNBR. Hence, we utilised burn severity
values generated from RdNBR as our target variable for variable selection and analysis in
the next stages. However, the upscaling of RdNBR index was performed to match the scale
of meteorological data as described in Section 2.4.1.
Remote Sens. 2022, 14, 1645 15 of 27

Table 4. Error matrices of bushfire severity classification for dNBR and RdNBR methods.

Class (dNBR) Very High High Moderate Low Total User Accuracy Kappa
Very High 135 26 0 0 161 0.839
High 14 208 2 4 228 0.912
Moderate 0 1 35 22 58 0.603
Low 0 1 13 159 173 0.919
Total 149 236 50 185 620 0.000
Producer Accuracy 0.906 0.881 0.700 0.859 0.000 0.866
Kappa 0.811
Class (RdNBR) Very High High Moderate Low Total User Accuracy Kappa
Very High 132 14 0 0 146 0.904
High 17 220 2 3 242 0.909
Moderate 0 1 44 19 64 0.688
Low 0 1 4 163 168 0.970
Total 149 236 50 185 620 0.000
Producer Accuracy 0.886 0.932 0.880 0.881 0.000 0.902
Kappa 0.861

Figure 6. An example of variation in (a) dNBR and (b) RdNBR indices in a heterogenous landscape.
Unlike dNBR, RdNBR shows very high severity in the highlighted area (black rectangle in map) (c).
The ground truth plot data in the region support the results obtained from RdNBR method.

3.3. Predictive Assessment of Feature-Selection Models

Our results indicate that Fuzzy Forest (FF) is the best representative model to deter-
mine the most influencing bushfire severity drivers. This conclusion is drawn from 10-fold
cross-validation generated for RF, BRT, XGBoost, and FF models. The accuracy of these
models runs from 89.8% to 96.5%. The model with the highest accuracy is FF, where the
prediction accuracy is 96.5%. Second to FF is BRT, which gives predictive accuracy of 92.4%.
The accuracy of RF(VSURF) is 89.8%, which is the lowest prediction accuracy among the
models. Similarly, the accuracy of the prediction using the XGBoost model is 91.6%, which
Remote Sens. 2022, 14, 1645 16 of 27

is comparable to the BRT model. The accuracy of all implemented models is presented in
Figure 7.

Figure 7. Predictive accuracy of implemented models. Accuracy assessment is performed using

10-fold cross-validation metrics.

The accuracy assessment result indicates that any of these models can be implemented
for variable selection. However, depending on the nature of data, especially on dealing with
highly correlated features, the implementation of the Fuzzy Forest algorithm utilising the
WGCNA package for partitioning covariates into distinct clusters can give better accuracy.

3.4. Variable Ranking

Based on the four different variable selection models used, we listed 15 of the 62 most
important variable parameters (Figure 8; Table 5). Among them, seven covariates, i.e.,
soil moisture and soil temperature at 3 different depth levels (2.0 m, 0.225 m, 0.675 m)
and air pressure at 20 m from the surface (PresSpec), are deemed important in all four
models. Similarly, the surface pressure (SfcPres), vertical direction wind at 1000 hPa
(Vwnd1000), and soil temperature at 0.05 m (STempP05) are identified among the top
15 variables by at least three variable selection models. There exist some variations in the
listed parameters and in the outcome order among the implemented models; however,
all the tested models identified that soil moisture and soil temperature at different depths
are in the topmost rank. Three of the four models—BRT, XGBoost, and FF—listed soil
moisture at 0.225 m as the most influencing bushfire severity variable. Similarly, three
of the four models—RF(VSURF), BRT, and XGBoost—ranked vertical direction wind at
different pressure levels in the top list, whereas BRT, XGBoost, and FF identified relative
humidity in the top 15 ranks. Among them, relative humidity at 500 hPa is ranked as
12th and 14th by two of the implemented models, XGBoost and BRT, respectively. Air
temperature at different vertical levels, from 700 hPa to 1000 hPa, are identified among the
top severity-influencing parameters by RF(VSURF) and FF models; however, neither of the
air temperature parameters are listed by the other two models. Similarly, mean specific
atmospheric humidity (AvQsair) and average surface upward latent heat flux (AvLatHflx)
are identified as the 13th and 14th most important variable by XGBoost and FF, respectively,
but have not been ranked elsewhere.
Remote Sens. 2022, 14, 1645 17 of 27

Figure 8. Variable ranking based on different implemented variable selection models. The ranking
is performed based on relative influence percentage in BRT, feature score in XGBoost, and feature
importance score in Fuzzy Forest. The RF(VSURF) model only provided rankings of the variables
without their corresponding relative importance, please refer to Table 5 for the ranking.

In a broader sense, the importance of these variables can be ranked in the order of
(1) soil moisture, (2) soil temperature, (3) air pressure, (4) air temperature, (5) vertical
wind, and (6) relative humidity, as evidenced by the variable importance scores of different
models (Table 5). This is because multiple parameters of these variables have made the
top 15 ranks of different models indicating the order (Table 5). However, the XGBoost and
BRT models show some variations in the ranking, which resulted in different vertical wind
parameters among the top variables. A detailed list of the variables and the ranking of the
tested model is given below (Table 5).

Table 5. Bushfire-influencing variables and their ranking obtained from different variable selection
models. A brief explanation of the variables is presented in Table 1.

RF (VSURF) Boosted Regression Trees Fuzzy Forest XGBoost

Rank
Relative
Variable Variable Variable Importance Variable F Score
Influence
1 STempP225 STempP225 32.37 SMoisP225 0.108 SMoisP225 58
2 SMoisP225 SMoisP225 25.13 STempP225 0.106 STemp2P0 41
3 STempP675 STemp2P0 8.33 STempP675 0.071 SMoisP05 38
4 SMois2P0 STempP675 6.77 ATemp750 0.053 PresSpec 30
5 STemp2P0 PresSpec 2.60 PresSpec 0.051 STempP225 26
6 ATemp700 SMois2P0 2.18 SfcPres 0.050 SMoisP0675 25
7 SfcPres SMoisP05 1.98 STemp2P0 0.049 STempP675 22
8 PresSpec SMoisP0675 1.68 SMoisP0675 0.042 Vwnd500 18
9 SMoisP0675 Vwnd1000 1.53 STempP05 0.033 ATemp700 17
10 ATemp750 ATemp950 1.24 SMoisP05 0.027 SMois2P0 17
11 ATemp950 STempP05 0.87 ATemp925 0.025 Vwnd1000 14
12 STempP05 Vwnd975 0.84 ATemp950 0.021 RH500 12
13 SMoisP05 AvVwnd10m 0.79 ATemp1000 0.018 AvLatHflx 11
14 RH700 RH500 0.73 RH600 0.013 SfcPres 10
15 Vwnd1000 RH750 0.69 RH950 0.010 Vwnd700 10
Remote Sens. 2022, 14, 1645 18 of 27

4. Discussion
The present study performed a comparative assessment of meteorological variables
influencing bushfire severity, considering nine fires that occurred between 29 January 2009
and 31 March 2009 in Victoria, Australia. Our discussion is framed into three key elements
that emerged from this study: variable selection models and their performances; bushfire
severity driving variables and their influence; management and policy outlook that can be
synthesised from this study.

4.1. Performance of Variable Selection Models

The implemented four different models of variable selection listed a similar subset
of ranked variables. In terms of predictive performance against 10-fold cross-validation,
the Fuzzy Forest (FF) model performed the best (96.5%) among all models. Two of the
four models, BRT and XGBoost, have comparable predictive performances of 92.4% and
91.6%, respectively. The RF(VSURF) model had the lowest performance (89.8%) among
all models. The higher performance of FF can be justified in relation to the nature of our
input datasets. The analysis using VIF and Tolerance values indicates that there exists
strong multicollinearity between several input parameters [102,103]. Perfectly correlated
variables may be redundant in the sense that no additional information can be gained
by adding them; however, if the data contain noise, adding correlated variables can help
to achieve better class separations [104] and can also improve performance when taken
with other variables. This issue of multicollinearity in our datasets can be of concern,
especially when we are assessing the relative influence of the variables [24]. FF can tackle
this issue during the analysis because this algorithm is specially designed for handling
correlated features [91,97]. Additionally, the WGCNA algorithm implemented in FF settings
provides a rigorous framework for finding highly correlated clusters (modules) suitable
for analysing high-dimensional multicorrelated datasets [105], which could justify the
reason behind the highest accuracy of FF among the models. Similarly, the BRT model can
also reliably select the predictors without being affected by multicollinearity and have the
second-highest predictive accuracy among the implemented models [45]. However, the
predictive performance is not a primary focus of this research; we are more concerned with
the relative influence of the variables over bushfire severity. During the experiment, while
fine-tuning the hyperparameters for optimum accuracy, we also observed that a small
variation in the accuracy of prediction does not necessarily make a significant difference in
the subset and rank of influencing variables. Hence, our effort is concentrated on making
a subset of choice, rather than focusing on accuracy in predictions, which we consider
an intermediate product of our experimental design. Even from the accuracy aspect,
our experimental models worked very well because all the models resulted prediction
accuracies greater than 89.90%. This demonstrates the reliability of the chosen models in
terms of predictive performance.

4.2. Relative Influence of Meteorological Variables on Bushfire Severity

The results indicate that soil moisture, soil temperature, air pressure, air temperature,
vertical wind, and relative humidity are the top six drivers of bushfire severity. All the
variable selection methods that we implemented supported this finding, although they
show some variations in the ranking order. The agreement in the outcome between all used
methods ensures the reliability of the findings.
Several important insights emerge from our analysis. Firstly, soil moisture is the
topmost meteorological variable that emerged from all four variable selection methods.
There are several studies investigating relationships between soil moisture at different depth
levels (near-surface, subsurface, and deep layers) and their influence on vegetation water
content, which can be linked to bushfire risks [106,107]. However, only a handful of studies
identified soil moisture as an important parameter for bushfire predictions [18,23,108–110].
Further, no bushfire severity research has highlighted soil moisture influencing the severity
of the burn and considered it among the most influencing severity drivers; this is often
Remote Sens. 2022, 14, 1645 19 of 27

because they are not widely considered in bushfire severity studies. However, some of the
fire prediction models, including the McArthur Forest Fire Danger Index [30], include soil
moisture deficit information from Keetch-Byram Drought Index or Mount’s Soil Dryness
Index [109]. Although the relationship between soil moisture and extreme meteorological
conditions that could potentially influence the bushfires, including heat waves, are well
documented [18], the comparative analysis of meteorological variables including soil
moisture in bushfire severity studies is unexplored, at least in the Australian context.
However, our finding is concordant with the research that explored the influence of soil
moisture conditions on the heatwave event that preceded two weeks to Black Saturday,
where Kala et al. (2015) considered antecedent soil moisture conditions as a critical factor
for heatwaves that preceded the Black Saturday bushfires [50]. This finding is quite
unsurprising because Australia was facing a drought condition during the Black Saturday
bushfire, when moisture was strongly constrained, intensifying the effect of heat waves on
bushfires, and making soil moisture one of the most influencing drivers of bushfire severity.
Further, soil moisture at two metres depth, which is the deepest vertical level considered in
this study, is ranked as the topmost influencing parameter of bushfire severity in all models.
This indicates that soil moisture is related to the live fuel water content and above ground
combustibles [23,110], which influence the severity of burning.
Similarly, the soil temperature is the second most influencing variable, as resulted
in all four models; however, to our knowledge, no study has identified soil temperature
as one of the influencing drivers to bushfire severity. This is mostly because it has not
been explored in relation to bushfire severity analysis. There are some studies studying
impacts on soil temperature during and following bushfires [31,111–114]. For example,
Bradstock et al. [31] studied the effect of experimental woodland fires on soil temperature
at the surface and at different depth levels to determine the influence of fire intensity.
Unlike our study, which focuses on fire severity, they are more focused on fire intensity.
However, the research identified soil temperature as an important driver of bushfire inten-
sity, which, in part, appears to support our research findings. Similarly, Hirschi et al. [112]
identified that the extreme heat is deep-rooted in dry soils [113], which indicates that soil
moisture deficits and warmer soil temperature can lead to more frequent and severe hot
temperature, leading to heatwaves, which are related to bushfires, as evidenced by different
studies [114]. This experiment provides observational evidence to consider soil tempera-
ture as a highly influencing meteorological driver to bushfire severity. Hence, this study
ranked soil temperature and soil moisture among the most influencing meteorological
drivers, often overlooked in the existing research, adding novel insights for researchers and
decision-makers.
Further, air pressure, air temperature, and vertical wind are identified as the 3rd,
4th, and 5th most influencing variables of bushfire severity in this study. Many studies
suggest that different climatic conditions including dynamic channelling of the wind
depend on air pressure, which influences fire behaviour [115]. This finding is supported
by different literature including Davies et al. (2016), who identified that the atmospheric
pressure jump was evident during the Black Saturday bushfire [116], ultimately affecting
the burn severity. Similarly, temperature and wind have long been identified to have major
effects on bushfires [18,28,117–119]. Different bushfire prediction and simulation models
including Spark [120] and Phoenix [121] consider both temperature and wind as important
factors in bushfire analysis. Generally, bushfires are more prone to high temperature
and strong gusty winds, which is evidenced in our data analysis. Our result shows that
the vertical component of the wind up to 700 hPa has a higher influence on the bushfire
severity. Although several studies are exploring the relationship between temperature and
bushfires, little is known regarding how the severity of bushfire is influenced by the vertical
component of wind. In accordance with our studies, Kochanski et al. (2013) provided
a proof of concept demonstrating that the surface wind impacts the vertical component
of wind influencing bushfire propagation and ultimately affects bushfire severity [122].
Further, a previous study suggested that the antecedent weather condition of the study
Remote Sens. 2022, 14, 1645 20 of 27

area showed significant above-average temperature (12 consecutive warmer-than-average

years prior to 2009) [47]. Two heatwaves occurred in the study area, on 27–31 January and
6–7 February, exceeding the previous maximum temperature records. Previous studies
describing antecedent weather conditions supported our findings, listing air temperature
and wind as bushfire-severity-influencing variables.
Similarly, this study ranked Relative Humidity (RH) as the sixth-highest influencing
driver of bushfire severity. RH has been long identified and considered as one of the major
drivers of bushfires [23,123–127]. As an example, RH has been identified as the second most
influencing meteorological variable for Fire Weather Index (FWI) and Forest Fire Danger
Index (FFDI), which also supports our findings [126]. Similarly, our finding is concordant
with a detailed assessment on identifying forest fire driving factors in China [23], which
also identified RH as a major factor for bushfires and is negatively correlated with the
occurrence of forest fires probability because high relative humidity reduces the probability
of fire by increasing the moisture content of combustible materials. Further, if we consider
the measurement reading, according to Stephenson et al. [127], the probability of bushfires
is higher for the region with RH < 65%, which is the case in our study where the average of
the relative humidity is 45%, which further supports our findings.

4.3. Management and Policy Outlook

The improved understanding of bushfire severity in terms of meteorological variables
provides a range of options for land managers and policy-makers. The existing fire danger
ratings do not adequately reflect the potential destructiveness and intensity of bushfires
and, hence, do not provide an objective warning to the people and community regarding
the potential loss of lives and properties [128]. This is because the severity and intensity of
bushfires are still not well-quantified. Even those studies quantifying bushfires severity
are mostly based on the limited number of variables, which often lack a scientific basis for
choosing the severity-driving variables. However, this study showcases a scientific basis for
choosing the variables and variable selection models with assessed performance outcomes.
This helps researchers select the most important variables for quantifying bushfire severity,
ultimately assisting land managers and policy-makers in making informed decisions.
The social impact of the bushfires makes them disastrous; hence, many recent bush-
fire studies have focused on understanding social and policy elements to mitigate bush-
fire impact [8,129,130]. Further, different research incorporated socioeconomic and ma-
terial vulnerability drivers of bushfires to determine the extent of lives and properties
affected [131–134]. Ghorbanzadeh et al. (2019) derived social vulnerability indicators and
developed an index to assess the social/infrastructural vulnerability [132]. However, the
knowledge on the meteorological influence of bushfire severity developed in this study
can add further insights on assessing its impact and on drafting and implementing the
mitigation policies in protecting the vulnerable infrastructure. For example, the fire author-
ities of Australian states and territories have adopted a ‘Prepare, Stay and defend or Leave
Early’ policy underpinned by strong evidence that late evacuation can lead to a dangerous
situation and houses can be successfully defended if well-prepared [135]. Our research
helps decision-makers in assessing the potential severity of bushfires. Different research
suggests that highly severe bushfires have a greater possibility of destroying homes [136],
in which case choosing to stay and defend the house can be extremely dangerous. With
changing weather dynamics, the risks of fire can vary in a small-time window, requiring a
frequent update on the estimated risks. However, running a complete model incorporating
a series of driving variables can be costly concerning time and computational efficiency.
Hence, the list of highly influencing variables that we generated in this research can help
create surrogate models of fire severity for rapid assessment [137]. This way, the authorised
agencies can make frequent updates to the public with the changing bushfire severity
conditions and associated risks, or to the firefighters involved in combating fires, thus
saving lives and properties.
Remote Sens. 2022, 14, 1645 21 of 27

4.4. Limitations
Unlike previous bushfire studies, which have usually been based on a limited number
of bushfire drivers, this study has considered extensive meteorological variables and pa-
rameters for the bushfire severity study including the measurements at different vertical
pressure levels, obtained at a high temporal frequency (6-hourly interval). This study at-
tempted to explore the influence of parameters beyond near-surface levels—i.e., at different
vertical levels—but a more in-depth analysis on how the measurement at different vertical
levels influences the fire severity is required in future. The VIF analysis in this study shows
that there exists multicollinearity between the meteorological variables, which might affect
the interpretation of the results. However, we tackled this by using models that can handle
multicollinearity. Further, the spatial resolution of the meteorological data is 12 km, which
might have limited variability in meteorological readings for smaller areas. This study en-
tirely focuses on the meteorological conditions and have not considered nonmeteorological
factors including vegetation fuels, anthropogenic drivers, and landscape topographies such
as slope and aspect during the analysis, although they have been identified as important
drivers in previous studies [21,26,138]. This is because the research aims to go deeper into
the meteorological influences, irrespective of other drivers in bushfire severity. Similarly,
bushfires depend on meteorological factors as well as fuel availability in an area including
humidity, precipitation, and temperature in the weeks, months, and sometimes even years
before the actual bushfire event. This research focuses only on the concurrent fire weathers;
hence, the relationship between the fire and antecedent meteorological conditions is out of
the scope of this research.

5. Conclusions
A comparative study and quantitative assessment of meteorological variables could
lead to improved bushfire severity models that can help land managers and decision-
makers to make informed decisions. In this study, we demonstrated the predictive perfor-
mance of different decision-tree-based machine learning models for variable selection and
implemented in a complex multicollinear meteorological dataset, taking the bushfires in
Victoria in the first quarter of 2009 as a case study. Further, we identified the top 15 most
influencing bushfire severity variables and ranked them based on their feature importance.
The feature importance outcome provides a scientific basis for choosing variables for the
bushfire severity analysis. Our analysis shows that Fuzzy Forest, an extension of random
forest algorithm designed for handling multicorrelated data, has the best performance
(96.5%) among the four implemented methods. Further, the model outcome shows that
the bushfire-severity-influencing variables—(1) soil moisture, (2) soil temperature, (3) air
pressure, (4) air temperature, (5) vertical wind, and (6) relative humidity—can be ranked
in order of their importance to bushfire severity. We recommend and aim to pursue three
avenues for future research: First, the landscape, anthropogenic drivers, and their analysis
in the integrated models. Second, a comparative analysis with additional performance
assessment measures. Third, an extended time window of analysis for useability and
transferability potential to other areas.

Author Contributions: Conceptualisation, methodology, writing–review & editing, S.K.S. and J.A.,
software, validation, formal analysis, data curation, writing–original draft, visualization, S.K.S.; su-
pervision, J.A. and A.R. All authors have read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Data Availability Statement: Not applicable.
Remote Sens. 2022, 14, 1645 22 of 27

Acknowledgments: The authors would like to acknowledge The University of Melbourne and the
Australian Commonwealth Government for providing the research platform and environment includ-
ing Research Training Program Scholarships to undertake PhD research of Saroj K Sharma. We thank
the Bureau of Meteorology for Atmospheric high-resolution regional reanalysis for Australia (BARRA)
data, Victorian Government, Department of Environment Land Water and Planning (DELWP) for
2009 bushfire severity validation point data, U.S. Geological Survey for Landsat images, NASA for
MODIS active fire products, and Google for Google Earth Engine platform.
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 manuscript, or
in the decision to publish the results.

References
1. Sharples, J.J.; Cary, G.J.; Fox-hughes, P.; Mooney, S.; Evans, J.P.; Fletcher, M.; Fromm, M.; Grierson, P.F.; Mcrae, R.; Baker, P. Natural
Hazards in Australia: Extreme Bushfire. Clim. Change 2016, 139, 85–99. [CrossRef]
2. Bowman, D.M.J.S.; Williamson, G.J.; Gibson, R.K.; Bradstock, R.A.; Keenan, R.J. The Severity and Extent of the Australia 2019–20
Eucalyptus Forest Fires Are Not the Legacy of Forest Management. Nat. Ecol. Evol. 2019, 5, 1003–1010. [CrossRef] [PubMed]
3. Turco, M.; Rosa-Cánovas, J.J.; Bedia, J.; Jerez, S.; Montávez, J.P.; Llasat, M.C.; Provenzale, A. Exacerbated Fires in Mediterranean
Europe Due to Anthropogenic Warming Projected with Non-Stationary Climate-Fire Models. Nat. Commun. 2018, 9, 1–9.
[CrossRef] [PubMed]
4. Burke, M.; Driscoll, A.; Heft-Neal, S.; Xue, J.; Burney, J.; Wara, M. The Changing Risk and Burden of Wildfire in the United States.
Proc. Natl. Acad. Sci. USA 2021, 118, e2011048118. [CrossRef] [PubMed]
5. Bowman, D.M.J.S.; Balch, J.K.; Artaxo, P.; Bond, W.J.; Carlson, J.M.; Cochrane, M.A.; Antonio, C.M.D.; Defries, R.S.; Doyle, J.C.;
Harrison, S.P.; et al. Fire in the Earth System. Science 2009, 324, 481–485. [CrossRef] [PubMed]
6. Dutta, R.; Das, A.; Aryal, J. Big Data Integration Shows Australian Bush-Fire Frequency Is Increasing Significantly. R. Soc. Open
Sci. 2016, 3, 150241. [CrossRef] [PubMed]
7. Dutta, R.; Aryal, J.; Das, A.; Kirkpatrick, J.B. Deep Cognitive Imaging Systems Enable Estimation of Continental-Scale Fire
Incidence from Climate Data. Sci. Rep. 2013, 3, 03188. [CrossRef]
8. Prior, T.; Eriksen, C. Wildfire Preparedness, Community Cohesion and Social-Ecological Systems. Glob. Environ. Change 2013, 23,
1575–1586. [CrossRef]
9. Linnenluecke, M.K.; Stathakis, A.; Griffiths, A. Firm Relocation as Adaptive Response to Climate Change and Weather Extremes.
Glob. Environ. Change 2011, 21, 123–133. [CrossRef]
10. Halgamuge, M.N.; Daminda, E.; Nirmalathas, A. Best Optimizer Selection for Predicting Bushfire Occurrences Using Deep
Learning. Nat. Hazards 2020, 103, 845–860. [CrossRef]
11. Shi, G.; Yan, H.; Zhang, W.; Dodson, J.; Heijnis, H.; Burrows, M. Rapid Warming Has Resulted in More Wildfires in Northeastern
Australia. Sci. Total Environ. 2021, 771, 144888. [CrossRef]
12. Beale, J.; Jones, W. Preventing and Reducing Bushfire Arson in Australia: A Review of What Is Known. Fire Technol. 2011, 47,
507–518. [CrossRef]
13. Mcaneney, J.; Chen, K.; Pitman, A. 100-Years of Australian Bushfire Property Losses: Is the Risk Significant and Is It Increasing?
J. Environ. Manag. 2009, 90, 2819–2822. [CrossRef]
14. Godfree, R.C.; Knerr, N.; Encinas-viso, F.; Albrecht, D.; Bush, D.; Cargill, D.C.; Clements, M.; Gueidan, C.; Guja, L.K.; Harwood,
T.; et al. Implications of the 2019–2020 Megafires for the Biogeography and Conservation of Australian Vegetation. Nat. Commun.
2020, 12, 1–13. [CrossRef]
15. Tosic, I.; Mladjan, D.; Gavrilov, M.B.; Zivanović, S.; Radaković, M.G.; Putniković, S.; Petrović, P.; Mistridzelović, I.K.; Marković,
S.B. Potential Influence of Meteorological Variables on Forest Fire Risk in Serbia during the Period 2000–2017. Open Geosci. 2019,
11, 414–425. [CrossRef]
16. Santika, T.; Budiharta, S.; Law, E.A.; Dennis, R.A.; Dohong, A.; Struebig, M.J.; Medrilzam; Gunawan, H.; Meijaard, E.; Wilson,
K.A. Interannual Climate Variation, Land Type and Village Livelihood Effects on Fires in Kalimantan, Indonesia. Glob. Environ.
Change 2020, 64, 102129. [CrossRef]
17. Bradstock, R.A. A Biogeographic Model of Fire Regimes in Australia: Current and Future Implications. Glob. Ecol. Biogeogr. 2010,
19, 145–158. [CrossRef]
18. Oldenborgh, G.J.V.; Krikken, F.; Lewis, S.; Leach, N.J.; Lehner, F.; Saunders, K.R.; Van Weele, M.; Haustein, K.; Li, S.; Wallom,
D.; et al. Attribution of the Australian Bushfire Risk to Anthropogenic Climate Change. Nat. Hazards Earth Syst. Sci. 2021, 21,
941–960. [CrossRef]
19. Sharples, J.J. An Overview of Mountain Meteorological Effects Relevant to Fire Behaviour and Bushfire Risk. Int. J. Wildl. Fire
2009, 18, 737–754. [CrossRef]
20. Sanderson, B.M.; Fisher, R.A. A Fiery Wake-up Call for Climate Science. Nat. Clim. Change 2020, 10, 175–177. [CrossRef]
21. Argañaraz, J.P.; Gavier Pizarro, G.; Zak, M.; Landi, M.A.; Bellis, L.M. Human and Biophysical Drivers of Fires in Semiarid Chaco
Mountains of Central Argentina. Sci. Total Environ. 2015, 520, 1–12. [CrossRef]
Remote Sens. 2022, 14, 1645 23 of 27

22. Rasilla, D.F.; García-Codron, J.C.; Carracedo, V.; Diego, C. Circulation Patterns, Wildfire Risk and Wildfire Occurrence at
Continental Spain. Phys. Chem. Earth 2010, 35, 553–560. [CrossRef]
23. Ma, W.; Feng, Z.; Cheng, Z.; Chen, S.; Wang, F. Identifying Forest Fire Driving Factors and Related Impacts in China Using
Random Forest Algorithm. Forests 2020, 11, 507. [CrossRef]
24. Liang, H.; Zhang, M.; Wang, H. A Neural Network Model for Wildfire Scale Prediction Using Meteorological Factors. IEEE Access
2019, 7, 176746–176755. [CrossRef]
25. Jenkins, M.E.; Bedward, M.; Price, O.; Bradstock, R.A. Modelling Bushfire Fuel Hazard Using Biophysical Parameters. Forests
2020, 11, 925. [CrossRef]
26. Guo, F.; Su, Z.; Wang, G.; Sun, L.; Tigabu, M.; Yang, X.; Hu, H. Understanding Fire Drivers and Relative Impacts in Different
Chinese Forest Ecosystems. Sci. Total Environ. 2017, 605–606, 411–425. [CrossRef]
27. Su, Z.; Hu, H.; Wang, G.; Ma, Y.; Yang, X.; Guo, F. Using GIS and Random Forests to Identify Fire Drivers in a Forest City, Yichun,
China. Geomat. Nat. Hazards Risk 2018, 9, 1207–1229. [CrossRef]
28. Blanchi, R.; Leonard, J.; Haynes, K.; Opie, K.; James, M.; Oliveira, F.D. de Environmental Circumstances Surrounding Bushfire
Fatalities in Australia 1901-2011. Environ. Sci. Policy 2014, 37, 192–203. [CrossRef]
29. Nolan, R.H.; Boer, M.M.; Collins, L.; Resco de Dios, V.; Clarke, H.; Jenkins, M.; Kenny, B.; Bradstock, R.A. Causes and
Consequences of Eastern Australia’s 2019–20 Season of Mega-Fires. Glob. Change Biol. 2020, 26, 1039–1041. [CrossRef]
30. Dowdy, A.J.; Mills, G.A.; Finkele, K.; De Groot, W. Australian Fire Weather as Represented by the McArthur Forest Fire Danger Index
and the Canadian Forest Fire Weather Index; Centre for Australian Weather and Climate Research: Melbourne, Austrialia, 2009.
31. Bradstock, R.A.; Auld, T.D. Soil Temperatures During Experimental Bushfires in Relation to Fire Intensity: Consequences for
Legume Germination and Fire Management in South-Eastern Australia. J. Appl. Ecol. 1995, 32, 76. [CrossRef]
32. Storey, M.A.; Bedward, M.; Price, O.F.; Bradstock, R.A.; Sharples, J.J. Derivation of a Bayesian Fire Spread Model Using Large-Scale
Wildfire Observations. Environ. Model. Softw. 2021, 144, 105127. [CrossRef]
33. Knight, I.K.; Sullivan, A.L. A Semi-Transparent Model of Bushfire Flames to Predict Radiant Heat Flux. Int. J. Wildl. Fire 2004, 13,
201–207. [CrossRef]
34. Potter, B.E. Atmospheric Interactions with Wildland Fire Behaviour I. Basic Surface Interactions, Vertical Profiles and Synoptic
Structures. Int. J. Wildl. Fire 2012, 21, 779–801. [CrossRef]
35. Taylor, J.; Webb, R. Meteorological Aspects of the January 2003 South-Eastern Australia Bushfire Outbreak. Aust. For. 2005, 68,
94–103. [CrossRef]
36. Guo, H.; Wang, L.; Liang, D. Big Earth Data from Space: A New Engine for Earth Science. Sci. Bull. 2016, 61, 505–513. [CrossRef]
37. Olden, J.D.; Lawler, J.J.; Poff, N.L. Machine Learning Methods Without Tears: A Primer for Ecologists. Q. Rev. Biol. 2008, 83,
171–193. [CrossRef]
38. Sharma, A.; Dey, S. A Comparative Sudy of Feature Selection and Machine Learning Techniques for Sentiment Analysis. In
Proceedings of the 2012 ACM Research in Applied Computation Symposium, San Antoino, TX, USA, 23–26 October 2012; pp.
1–7.
39. Hosseini, F.S.; Choubin, B.; Mosavi, A.; Nabipour, N.; Shamshirband, S.; Darabi, H.; Haghighi, A.T. Flash-Flood Hazard
Assessment Using Ensembles and Bayesian-Based Machine Learning Models: Application of the Simulated Annealing Feature
Selection Method. Sci. Total Environ. 2020, 711, 135161. [CrossRef]
40. Masmoudi, S.; Elghazel, H.; Taieb, D.; Yazar, O.; Kallel, A. A Machine-Learning Framework for Predicting Multiple Air Pollutants’
Concentrations via Multi-Target Regression and Feature Selection. Sci. Total Environ. 2020, 715, 136991. [CrossRef]
41. Guan, D.; Yuan, W.; Lee, Y.K.; Najeebullah, K.; Rasel, M.K. A Review of Ensemble Learning Based Feature Selection. IETE Tech.
Rev. 2014, 31, 190–198. [CrossRef]
42. He, Q.; Jiang, Z.; Wang, M.; Liu, K. Landslide and Wildfire Susceptibility Assessment in Southeast Asia Using Ensemble Machine
Learning Methods. Remote Sens. 2021, 13, 1572. [CrossRef]
43. Polikar, R. Ensemble Machine Learning; Springer: Boston, MA, USA, 2012; ISBN 9781441993250.
44. Fan, J.; Yue, W.; Wu, L.; Zhang, F.; Cai, H.; Wang, X.; Lu, X.; Xiang, Y. Evaluation of SVM, ELM and Four Tree-Based Ensemble
Models for Predicting Daily Reference Evapotranspiration Using Limited Meteorological Data in Different Climates of China.
Agric. For. Meteorol. 2018, 263, 225–241. [CrossRef]
45. Elith, J.; Leathwick, J.R.; Hastie, T. A Working Guide to Boosted Regression Trees. J. Anim. Ecol. 2008, 77, 802–813. [CrossRef]
46. Genuer, R.; Poggi, J.; Tuleau-malot, C.; Genuer, R.; Poggi, J.; Variable, C.T.; Forests, R.; Genuer, R.; Poggi, J.; Tuleau-malot, C.
Variable Selection Using Random Forests. Pattern Recognit. Lett. 2010, 31, 2225–2236. [CrossRef]
47. Cruz, M.G.; Sullivan, A.L.; Gould, J.S.; Sims, N.C.; Bannister, A.J.; Hollis, J.J.; Hurley, R.J. Anatomy of a Catastrophic Wildfire: The
Black Saturday Kilmore East Fire in Victoria, Australia. For. Ecol. Manag. 2012, 284, 269–285. [CrossRef]
48. Whittaker, J.; Haynes, K.; Handmer, J.; Mclennan, J. Community Safety during the 2009 Australian ‘Black Saturday’ Bushfires: An
Analysis of Household Preparedness and Response. Int. J. Wildl. Fire 2013, 22, 841–849. [CrossRef]
49. Cai, W.; Cowan, T.; Raupach, M. Positive Indian Ocean Dipole Events Precondition Southeast Australia Bushfires. Geophys. Res.
Lett. 2009, 36, 039902. [CrossRef]
50. Kala, J.; Evans, J.P.; Pitman, A.J. Influence of Antecedent Soil Moisture Conditions on the Synoptic Meteorology of the Black
Saturday Bushfire Event in Southeast Australia. Q. J. R. Meteorol. Soc. 2015, 141, 3118–3129. [CrossRef]
51. Ashton, D.H. The Environment and Plant Ecology of the Hume Range, Central Victoria. Proc. R. Soc. Vic. 2000, 112, 185–272.
Remote Sens. 2022, 14, 1645 24 of 27

52. Peterson, D.; Wang, J.; Ichoku, C.; Hyer, E.; Ambrosia, V. A Sub-Pixel-Based Calculation of Fire Radiative Power from MODIS
Observations: 1. Algorithm Development and Initial Assessment. Remote Sens. Environ. 2013, 129, 262–279. [CrossRef]
53. Justice, C.O.; Giglio, L.; Korontzi, S.; Owens, J.; Morisette, J.T.; Roy, D.; Descloitres, J.; Alleaume, S.; Petitcolin, F.; Kaufman, Y. The
MODIS Fire Products. Remote Sens. Environ. 2002, 83, 244–262. [CrossRef]
54. Zhang, Y.; Lim, S.; Sharples, J.J. Modelling Spatial Patterns of Wildfire Occurrence in South-Eastern Australia. Geomat. Nat.
Hazards Risk 2016, 7, 1800–1815. [CrossRef]
55. Laurent, P.; Mouillot, F.; Yue, C.; Ciais, P.; Moreno, M.V.; Nogueira, J.M.P. Data Descriptor: FRY, a Global Database of Fire Patch
Functional Traits Derived from Space-Borne Burned Area Products. Sci. Data 2018, 5, 132. [CrossRef]
56. Briones-Herrera, C.I.; Vega-Nieva, D.J.; Monjarás-Vega, N.A.; Briseño-Reyes, J.; López-Serrano, P.M.; Corral-Rivas, J.J.; Alvarado-
Celestino, E.; Arellano-Pérez, S.; álvarez-González, J.G.; Ruiz-González, A.D.; et al. Near Real-Time Automated Early Mapping of
the Perimeter of Large Forest Fires from the Aggregation of VIIRS and MODIS Active Fires in Mexico. Remote Sens. 2020, 12, 2061.
[CrossRef]
57. Bureau of Meteorology Atmospheric High-Resolution Regional Reanalysis for Australia. Available online: http://www.bom.gov.
au/research/projects/reanalysis/ (accessed on 6 January 2022).
58. Su, C.H.; Eizenberg, N.; Steinle, P.; Jakob, D.; Fox-Hughes, P.; White, C.J.; Rennie, S.; Franklin, C.; Dharssi, I.; Zhu, H. BARRA
v1.0: The Bureau of Meteorology Atmospheric High-Resolution Regional Reanalysis for Australia. Geosci. Model Dev. 2019, 12,
2049–2068. [CrossRef]
59. Rew, R.; Davis, G. NetCDF: An Interface for Scientific Data Access. IEEE Comput. Graph. Appl. 1990, 10, 76–82. [CrossRef]
60. Fernandez-Manso, A.; Quintano, C.; Roberts, D.A. Burn Severity Analysis in Mediterranean Forests Using Maximum Entropy
Model Trained with EO-1 Hyperion and LiDAR Data. ISPRS J. Photogramm. Remote Sens. 2019, 155, 102–118. [CrossRef]
61. Key, C.H.; Benson, N.C. Landscape Assessment (LA): Sampling and Analysis Methods. In FIREMON: Fire Effects Monitoring and
Inventory System; Lutes, D.C., Keane, R.E., Caratti, C.H., Key, N.C., Sutherland, S., Eds.; Rocky Mountain Research Station, USDA
Forest Service: Fort Collins, CO, USA, 2006; pp. LA-1–LA-55.
62. Pelletier, F.; Eskelson, B.N.I.; Monleon, V.J.; Tseng, Y.C. Using Landsat Imagery to Assess Burn Severity of National Forest
Inventory Plots. Remote Sens. 2021, 13, 1935. [CrossRef]
63. Keeley, J.E. Fire Intensity, Fire Severity and Burn Severity: A Brief Review and Suggested Usage. Int. J. Wildl. Fire 2009, 18,
116–126. [CrossRef]
64. Miller, J.D.; Knapp, E.E.; Key, C.H.; Skinner, C.N.; Isbell, C.J.; Creasy, R.M.; Sherlock, J.W. Calibration and Validation of the
Relative Differenced Normalized Burn Ratio (RdNBR) to Three Measures of Fire Severity in the Sierra Nevada and Klamath
Mountains, California, USA. Remote Sens. Environ. 2009, 113, 645–656. [CrossRef]
65. Parks, S.A.; Dillon, G.K.; Miller, C. A New Metric for Quantifying Burn Severity: The Relativized Burn Ratio. Remote Sens. 2014, 6,
1827–1844. [CrossRef]
66. Tran, B.N.; Tanase, M.A.; Bennett, L.T.; Aponte, C. Evaluation of Spectral Indices for Assessing Fire Severity in Australian
Temperate Forests. Remote Sens. 2018, 10, 1680. [CrossRef]
67. Cansler, C.A.; McKenzie, D. How Robust Are Burn Severity Indices When Applied in a New Region? Evaluation of Alternate
Field-Based and Remote-Sensing Methods. Remote Sens. 2012, 4, 456–483. [CrossRef]
68. Roy, D.P.; Boschetti, L.; Trigg, S.N. Remote Sensing of Fire Severity: Assessing the Performance of the Normalized Burn Ratio.
IEEE Geosci. Remote Sens. Lett. 2006, 3, 112–116. [CrossRef]
69. Martini, L.; Faes, L.; Picco, L.; Iroumé, A.; Lingua, E.; Garbarino, M.; Cavalli, M. Assessing the Effect of Fire Severity on Sediment
Connectivity in Central Chile. Sci. Total Environ. 2020, 728, 139006. [CrossRef]
70. Miller, J.D.; Thode, A.E. Quantifying Burn Severity in a Heterogeneous Landscape with a Relative Version of the Delta Normalized
Burn Ratio (DNBR). Remote Sens. Environ. 2007, 109, 66–80. [CrossRef]
71. Holsinger, L.M.; Parks, S.A.; Saperstein, L.B.; Loehman, R.A.; Whitman, E. Improved Fire Severity Mapping in the North
American Boreal Forest Using a Hybrid Composite Method. Remote Sens. Ecol. Conserv. 2021, 112, 1–14. [CrossRef]
72. Zhang, X.; Wang, J.; Gao, F.; Liu, Y.; Schaaf, C.; Friedl, M.; Yu, Y.; Jayavelu, S.; Gray, J.; Liu, L.; et al. Exploration of Scaling Effects
on Coarse Resolution Land Surface Phenology. Remote Sens. Environ. 2017, 190, 318–330. [CrossRef]
73. Peng, D.; Wu, C.; Zhang, X.; Yu, L.; Huete, A.R.; Wang, F.; Luo, S.; Liu, X.; Zhang, H. Scaling up Spring Phenology Derived from
Remote Sensing Images. Agric. For. Meteorol. 2018, 256–257, 207–219. [CrossRef]
74. Parks, S.A.; Holsinger, L.M.; Voss, M.A.; Loehman, R.A.; Robinson, N.P. Mean Composite Fire Severity Metrics Computed with
Google Earth Engine Offer Improved Accuracy and Expanded Mapping Potential. Remote Sens. 2018, 10, 879. [CrossRef]
75. Parks, S.A.; Dobrowski, S.Z.; Panunto, M.H. What Drives Low-Severity Fire in the Southwestern USA? Forests 2018, 9, 165.
[CrossRef]
76. Cai, L.; Wang, M. Is the RdNBR a Better Estimator of Wildfire Burn Severity than the DNBR? A Discussion and Case Study in
Southeast China. Geocarto Int. 2020, 1–15. [CrossRef]
77. DELWP Fire Severity 2009 Ground Control Plot Locations. Available online: https://discover.data.vic.gov.au/dataset/fire-
severity-2009-ground-control-plot-locations (accessed on 8 January 2022).
78. Gorelick, N.; Hancher, M.; Dixon, M.; Ilyushchenko, S.; Thau, D.; Moore, R. Google Earth Engine: Planetary-Scale Geospatial
Analysis for Everyone. Remote Sens. Environ. 2017, 202, 18–27. [CrossRef]
Remote Sens. 2022, 14, 1645 25 of 27

79. Kumar, L.; Mutanga, O. Google Earth Engine Applications since Inception: Usage, Trends, and Potential. Remote Sens. 2018, 10,
1509. [CrossRef]
80. Bui, D.T.; Tsangaratos, P.; Ngo, P.T.T.; Pham, T.D.; Pham, B.T. Flash Flood Susceptibility Modeling Using an Optimized Fuzzy
Rule Based Feature Selection Technique and Tree Based Ensemble Methods. Sci. Total Environ. 2019, 668, 1038–1054. [CrossRef]
81. Midi, H.; Sarkar, S.K.; Rana, S. Collinearity Diagnostics of Binary Logistic Regression Model. J. Interdiscip. Math. 2010, 13, 253–267.
[CrossRef]
82. Eskandari, S.; Amiri, M.; Sãdhasivam, N.; Pourghasemi, H.R. Comparison of New Individual and Hybrid Machine Learning
Algorithms for Modeling and Mapping Fire Hazard: A Supplementary Analysis of Fire Hazard in Different Counties of Golestan
Province in Iran. Nat. Hazards 2020, 104, 305–327. [CrossRef]
83. Michael, Y.; Helman, D.; Glickman, O.; Gabay, D.; Brenner, S.; Lensky, I.M. Forecasting Fire Risk with Machine Learning and
Dynamic Information Derived from Satellite Vegetation Index Time-Series. Sci. Total Environ. 2021, 764, 142844. [CrossRef]
[PubMed]
84. Heikkinen, R.K.; Luoto, M.; Araujo, M.B.; Virkkala, R.; Thuiller, W.; Sykes, M.T. Methods and Uncertainties in Bioclimatic
Envelope Modelling under Climate Change. Prog. Phys. Geogr. 2006, 30, 751–777. [CrossRef]
85. Oliveira, S.; Oehler, F.; San-Miguel-Ayanz, J.; Camia, A.; Pereira, J.M.C. Modeling Spatial Patterns of Fire Occurrence in
Mediterranean Europe Using Multiple Regression and Random Forest. For. Ecol. Manage. 2012, 275, 117–129. [CrossRef]
86. Ozigis, M.S.; Kaduk, J.D.; Jarvis, C.H.; da Conceição Bispo, P.; Balzter, H. Detection of Oil Pollution Impacts on Vegetation Using
Multifrequency SAR, Multispectral Images with Fuzzy Forest and Random Forest Methods. Environ. Pollut. 2020, 256, 113360.
[CrossRef]
87. Saha, S.; Arabameri, A.; Saha, A.; Blaschke, T.; Ngo, P.T.T.; Nhu, V.H.; Band, S.S. Prediction of Landslide Susceptibility in
Rudraprayag, India Using Novel Ensemble of Conditional Probability and Boosted Regression Tree-Based on Cross-Validation
Method. Sci. Total Environ. 2021, 764, 142928. [CrossRef]
88. Asadikia, A.; Rajabifard, A.; Kalantari, M. Systematic Prioritisation of SDGs: Machine Learning Approach. World Dev. 2021, 140,
105269. [CrossRef]
89. Jain, P.; Coogan, S.C.P.; Subramanian, S.G.; Crowley, M.; Taylor, S.; Flannigan, M.D. A Review of Machine Learning Applications
in Wildfire Science and Management. Environ. Rev. 2020, 28, 478–505. [CrossRef]
90. Chen, T.; Guestrin, C. XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22nd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, New York, NY, USA, 13–17 August 2016; pp. 785–794.
91. Conn, D.; Ngun, T.; Li, G.; Ramirez, C.M. Fuzzy Forests: Extending Random Forest Feature Selection for Correlated, High-
Dimensional Data. J. Stat. Softw. 2019, 91, 1–25. [CrossRef]
92. Mitchell, R.; Frank, E. Accelerating the XGBoost Algorithm Using GPU Computing. PeerJ Comput. Sci. 2017, 3, e127. [CrossRef]
93. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [CrossRef]
94. Degenhardt, F.; Seifert, S.; Szymczak, S. Evaluation of Variable Selection Methods for Random Forests and Omics Data Sets. Brief.
Bioinform. 2019, 20, 492–503. [CrossRef]
95. Friedman, J.H.; Meulman, J.J. Multiple Additive Regression Trees with Application in Epidemiology. Stat. Med. 2003, 22,
1365–1381. [CrossRef]
96. Bentéjac, C.; Csörgő, A.; Martínez-Muñoz, G. A Comparative Analysis of Gradient Boosting Algorithms; Springer: Amsterdam,
The Netherlands, 2021; Volume 54, ISBN 0123456789.
97. Ramirez, C.M.; Abrajano, M.A.; Alvarez, R.M. Using Machine Learning to Uncover Hidden Heterogeneities in Survey Data. Sci.
Rep. 2019, 9, 16061. [CrossRef]
98. Li, J.; Zhou, D.; Qiu, W.; Shi, Y.; Yang, J.J.; Chen, S.; Wang, Q.; Pan, H. Application of Weighted Gene Co-Expression Network
Analysis for Data from Paired Design. Sci. Rep. 2018, 8, 622. [CrossRef]
99. Jung, Y. Multiple Predicting K-Fold Cross-Validation for Model Selection. J. Nonparametr. Stat. 2018, 30, 197–215. [CrossRef]
100. Murtaugh, P.A. Performance of Several Variable-Selection Methods Applied to Real Ecological Data. Ecol. Lett. 2009, 12, 1061–1068.
[CrossRef]
101. Lafayette, L.; Sauter, G.; Vu, L.; Meade, B. Spartan Performance and Flexibility: An HPC-Cloud Chimera. In Proceedings of the
OpenStack Summit, Barcelona, Spain, 27 October 2016.
102. Zhang, T.; Han, L.; Chen, W.; Shahabi, H. Hybrid Integration Approach of Entropy with Logistic Regression and Support Vector
Machine for Landslide Susceptibility Modeling. Entropy 2018, 20, 884. [CrossRef]
103. Hong, H.; Liu, J.; Zhu, A.X. Modeling Landslide Susceptibility Using LogitBoost Alternating Decision Trees and Forest by
Penalizing Attributes with the Bagging Ensemble. Sci. Total Environ. 2020, 718, 137231. [CrossRef]
104. Guyon, I.; Elisseeff, A. An Introduction to Variable and Feature Selection. J. Mach. Learn. Res. 2003, 3, 1157–1182. [CrossRef]
105. Langfelder, P.; Horvath, S. WGCNA: An R Package for Weighted Correlation Network Analysis. BMC Bioinform. 2008, 9, 559.
[CrossRef]
106. Yang, L.; Wei, W.; Chen, L.; Chen, W.; Wang, J. Response of Temporal Variation of Soil Moisture to Vegetation Restoration in
Semi-Arid Loess Plateau, China. Catena 2014, 115, 123–133. [CrossRef]
107. Boisramé, G.; Thompson, S.; Stephens, S. Hydrologic Responses to Restored Wildfire Regimes Revealed by Soil Moisture-
Vegetation Relationships. Adv. Water Resour. 2018, 112, 124–146. [CrossRef]
Remote Sens. 2022, 14, 1645 26 of 27

108. Ambadan, J.T.; Oja, M.; Gedalof, Z.; Berg, A.A. Satellite-Observed Soil Moisture As an Indicator of Wildfire Risk. Remote Sens.
2020, 12, 1543. [CrossRef]
109. Vinodkumar; Dharssi, I. Evaluation and Calibration of a High-Resolution Soil Moisture Product for Wildfire Prediction and
Management. Agric. For. Meteorol. 2019, 264, 27–39. [CrossRef]
110. Jia, S.; Kim, S.H.; Nghiem, S.V.; Kafatos, M.C. Live Fuel Moisture Estimation Using SMAP Soil Moisture and MODIS Vegetation
Indices in Southern California, USA. In Proceedings of the 2019 IEEE International Geoscience and Remote Sensing Symposium,
Yokohama, Japan, 28 July–2 August 2019; pp. 6126–6129. [CrossRef]
111. Wright, B.R.; Clarke, P.J. Relationships between Soil Temperatures and Properties of Fire in Feathertop Spinifex (Triodia schinzii
(Henrard) Lazarides) Sandridge Desert in Central Australia. Rangel. J. 2008, 30, 317–325. [CrossRef]
112. Hirschi, M.; Seneviratne, S.I.; Alexandrov, V.; Boberg, F.; Boroneant, C.; Christensen, O.B.; Formayer, H.; Orlowsky, B.; Stepanek, P.
Observational Evidence for Soil-Moisture Impact on Hot Extremes in Southeastern Europe. Nat. Geosci. 2011, 4, 17–21. [CrossRef]
113. Alexander, L. Climate Science: Extreme Heat Rooted in Dry Soils. Nat. Geosci. 2011, 4, 12–13. [CrossRef]
114. Miralles, D.G.; Teuling, A.J.; Van Heerwaarden, C.C.; De Arellano, J.V.G. Mega-Heatwave Temperatures Due to Combined Soil
Desiccation and Atmospheric Heat Accumulation. Nat. Geosci. 2014, 7, 345–349. [CrossRef]
115. Sharples, J.J.; McRae, R.H.D.; Weber, R.O. Wind Characteristics over Complex Terrain with Implications for Bushfire Risk
Management. Environ. Model. Softw. 2010, 25, 1099–1120. [CrossRef]
116. Davies, L.; Reeder, M.J.; Lane, T.P. A Climatology of Atmospheric Pressure Jumps over Southeastern Australia. Q. J. R. Meteorol.
Soc. 2017, 143, 439–449. [CrossRef]
117. Beer, T. The Interaction of Wind and Fire. Bound. Layer Meteorol. 1991, 54, 287–308. [CrossRef]
118. Andrews, P.L.; Cruz, M.G.; Rothermel, R.C. Examination of the Wind Speed Limit Function in the Rothermel Surface Fire Spread
Model. Int. J. Wildl. Fire 2013, 22, 959–969. [CrossRef]
119. Cruz, M.G.; Alexander, M.E.; Wakimoto, R.H. Development and Testing of Models for Predicting Crown Fire Rate of Spread in
Conifer Forest Stands. Can. J. For. Res. 2005, 35, 1626–1639. [CrossRef]
120. Miller, C.; Hilton, J.; Sullivan, A.; Prakash, M. SPARK—A Bushfire Spread Prediction Tool. IFIP Adv. Inf. Commun. Technol. 2015,
448, 262–271. [CrossRef]
121. Tolhurst, K.; Shields, B.; Chong, D. Phoenix: Development and application of a bushfire risk management tool. Aust. J. Emerg.
Manag. 2008, 23, 47–54.
122. Kochanski, A.; Jenkins, M.A.; Sun, R.; Krueger, S.; Abedi, S.; Charney, J. The Importance of Low-Level Environmental Vertical
Wind Shear to Wildfire Propagation: Proof of Concept. J. Geophys. Res. Atmos. 2013, 118, 8238–8252. [CrossRef]
123. Bedia, J.; Herrera, S.; Gutiérrez, J.M.; Zavala, G.; Urbieta, I.R.; Moreno, J.M. Sensitivity of Fire Weather Index to Different
Reanalysis Products in the Iberian Peninsula. Nat. Hazards Earth Syst. Sci. 2012, 12, 699–708. [CrossRef]
124. Beer, T.; Gill, A.M.; Moore, P.H. Australian Bushfire Danger under Changing Climatic Regimes; CSIRO Publishing: Clayton, Australia,
1988.
125. Deb, P.; Moradkhani, H.; Abbaszadeh, P.; Kiem, A.S.; Engström, J.; Keellings, D.; Sharma, A. Causes of the Widespread 2019–2020
Australian Bushfire Season. Earth Future 2020, 8, 1671. [CrossRef]
126. Dowdy, A.J.; Mills, G.A.; Finkele, K.; de Groot, W. Index Sensitivity Analysis Applied to the Canadian Forest Fire Weather Index
and the McArthur Forest Fire Danger Index. Meteorol. Appl. 2010, 17, 298–312. [CrossRef]
127. Stephenson, A.G.; Shaby, B.A.; Reich, B.J.; Sullivan, A.L. Estimating Spatially Varying Severity Thresholds of a Forest Fire Danger
Rating System Using Max-Stable Extreme-Event Modeling. Am. Meteorol. Soc. 2015, 54, 395–407. [CrossRef]
128. Harris, S.; Anderson, W.; Kilinc, M.; Fogarty, L. The Relationship between Fire Behaviour Measures and Community Loss: An
Exploratory Analysis for Developing a Bushfire Severity Scale. Nat. Hazards 2012, 63, 391–415. [CrossRef]
129. Jefferson, U.; Carmenta, R.; Daeli, W.; Phelps, J. Characterising Policy Responses to Complex Socio-Ecological Problems: 60 Fire
Management Interventions in Indonesian Peatlands. Glob. Environ. Change 2020, 60, 102027. [CrossRef]
130. Schwartz, N.B.; Uriarte, M.; Gutiérrez-Vélez, V.H.; Baethgen, W.; DeFries, R.; Fernandes, K.; Pinedo-Vasquez, M.A. Climate,
Landowner Residency, and Land Cover Predict Local Scale Fire Activity in the Western Amazon. Glob. Environ. Change 2015, 31,
144–153. [CrossRef]
131. Simon, G.L.; Dooling, S. Flame and Fortune in California: The Material and Political Dimensions of Vulnerability. Glob. Environ.
Change 2013, 23, 1410–1423. [CrossRef]
132. Ghorbanzadeh, O.; Blaschke, T.; Gholamnia, K.; Aryal, J. Forest Fire Susceptibility and Risk Mapping Using Social/Infrastructural
Vulnerability and Environmental Variables. Fire 2019, 2, 50. [CrossRef]
133. Hamilton, M.; Fischer, A.P.; Ager, A. A Social-Ecological Network Approach for Understanding Wildfire Risk Governance. Glob.
Environ. Change 2019, 54, 113–123. [CrossRef]
134. Cattau, M.E.; Harrison, M.E.; Shinyo, I.; Tungau, S.; Uriarte, M.; DeFries, R. Sources of Anthropogenic Fire Ignitions on the
Peat-Swamp Landscape in Kalimantan, Indonesia. Glob. Environ. Change 2016, 39, 205–219. [CrossRef]
135. Tibbits, A.; Whittaker, J. Stay and Defend or Leave Early: Policy Problems and Experiences during the 2003 Victorian Bushfires
Stay and Defend or Leave Early: Policy Problems and Experiences during the 2003 Victorian Bushfires. Environ. Hazards 2011, 7,
7891. [CrossRef]
136. Blanchi, R.; Lucas, C.; Leonard, J.; Finkele, K. Meteorological Conditions and Wildfire-Related House Loss in Australia. Int. J.
Wildl. Fire 2010, 19, 914–926. [CrossRef]
Remote Sens. 2022, 14, 1645 27 of 27

137. Kc, U.; Aryal, J.; Hilton, J.; Garg, S. A Surrogate Model for Rapidly Assessing the Size of a Wildfire over Time. Fire 2021, 4, 20.
[CrossRef]
138. Yin, C.; Xing, M.; Yebra, M. Relationships between Burn Severity and Environmental Drivers in the Temperate Coniferous Forest
of Northern China. Remote Sens. 2021, 13, 5127. [CrossRef]