Wildfire can become a catastrophic natural hazard, especially during dry
summer seasons in Australia. Severity is influenced by various
meteorological, geographical, and fuel characteristics. Modified Mark 4
McArthur's Grassland Fire Danger Index (GFDI) is a commonly used approach to
determine the fire danger level in grassland ecosystems. The degree of curing
(DOC, i.e. proportion of dead material) of the grass is one key ingredient in
determining the fire danger. It is difficult to collect accurate DOC
information in the field, and therefore ground-observed measurements are rather
limited. In this study, we explore the possibility of whether adding
satellite-observed data responding to vegetation water content (vegetation
optical depth, VOD) will improve DOC prediction when compared with the
existing satellite-observed data responding to DOC prediction models based on vegetation greenness
(normalised difference vegetation index, NDVI).
First, statistically significant relationships are established between
selected ground-observed DOC and satellite-observed vegetation datasets (NDVI
and VOD) with an
Wildfire can be responsible for major environmental damage or changes to ecosystems (Cobb et al., 2016; Gazzard et al., 2016; Mistry et al., 2016). One of the important components in determining the severity of wildfire is fuel availability. Wildland fuels can vary considerably, both spatially and temporally (Stambaugh et al., 2011). Various interpretations and characterisations of fuel have been made in past studies as a key contribution to assessing wildfire potential (Hudec and Peterson, 2012; Jurdao et al., 2012; Sharples et al., 2009b; Stambaugh et al., 2011; Yebra et al., 2013). Fuel can also be quantified by its age or the time since the last fire (Bradstock et al., 2010). Since this paper uses numbers of abbreviations that may not be familiar to the reader, please refer to the Appendices for the list of acronyms.
In this study, we focus on the availability of combustible fuel in the aboveground biomass in grassland ecosystems; this fuel availability metric is referred to as the degree of curing (DOC). The DOC is the percentage of dead material in a grassland fuel bed; 100 % indicates a fully cured (dead) grassland fuel complex. The DOC has a direct influence on wildfire development in grasslands and thus it is an important input for fire danger indices and fire spread models, such as the McArthur Grassland Fire Danger Index (GFDI) (Gill et al., 2010) and the CSIRO grassland fire spread model (Cruz et al., 2015; Kidnie et al., 2015). Generally, fires are unable to spread across grasslands that are less than 50 % cured (Anderson et al., 2011). However, this lower limit has been revised, since a more recent study demonstrated that fire can spread in grassland with DOC as low as 20 % (Cruz et al., 2015). In climatological studies, DOC is often assumed to be 100 % (Pitman et al., 2007). This leads to an overestimation of areas experiencing high levels of fire danger and hence provides only a weak indication of where to focus resources of fire agencies. An accurate, spatially and temporally explicit estimate of DOC would provide more useful guidance to these agencies.
Measuring DOC in the field is a tedious and expensive task, especially when an accurate assessment of curing is required. Anderson et al. (2011) suggested that current methods for measuring DOC still present problems. The visual assessment method, which relies on field observers to estimate the general curing value based on their expertise and a visual guide, is subjective and can often be unrepresentative of DOC of the entire area. Destructive sampling approaches can provide accurate field-based observation, but is a labour intensive task. Thus, Anderson et al. (2011) offered a simple field-based method utilising a levy rod, based on the modified point quadrant method of pasture assessment; the approach involves counting the number of live and dead touches on a thin steel rod that was driven into the ground. It was suggested that this approach can be applied across Australia with higher accuracy than current visual assessment methods (Anderson et al., 2011).
Apart from ground measurement, DOC can be estimated using satellite remotely sensed data, but it is limited by the satellite sensors' capability, e.g. spatial resolution and various atmospheric interferences. Dilley et al. (2004) established a relationship between curing and normalised difference vegetation index (NDVI, a proxy of vegetation canopy greenness) by estimating live fuel moisture content from NDVI and relating it to curing via an exponential function using a finite difference Levenberg–Marquardt method (Dilley et al., 2004; Rouse et al., 1973). Newnham et al. (2011) showed that estimation of curing using a relative greenness (RG) approach that was based on NDVI distribution provided more accurate estimation of curing than a direct linear regression between curing and NDVI. Chladil and Nunez (1995) used curing derived from a soil dryness index model and NDVI to predict soil and fuel moisture content. Various optical-based vegetation indices computed from remote sensing reflectance products can also be developed into a satellite-based model integrated with ground observations to predict curing (Martin et al., 2015; Turner et al., 2011). These methods, though vastly different in their approaches, achieved good results for their set objectives, but they tend to focus on particular applications. It should also be noted that optical-based remote sensing products, including NDVI, are affected by cloud cover and aerosols. Some studies explicitly acknowledge challenges presented by cloud effects and when there are both forest and water bodies in the same NDVI pixel, which results in an erroneous grassland interpretation (Allan et al., 2003; Chladil and Nunez, 1995). However, if appropriately detailed aerosol data are available, atmospheric correction can mitigate the aerosol effect on NDVI.
Currently, there are satellite-based DOC products over Australia provided by the Bureau of Meteorology. The products have 500 m spatial resolution and 8-day temporal resolution and are based on two past studies (Martin et al., 2015; Newnham et al., 2010). There are five separate satellite-based DOC models; four are from Newnham et al. (2010) and one is from Martin et al. (2015). All satellite-based DOC models here are based on optical and near-infrared wavelength bands. We would like to investigate whether including a recent passive microwave-based satellite product can improve the DOC estimation over Australia.
A passive microwave-based remote sensing vegetation product, referred to as
vegetation optical depth (VOD), has been developed recently (Meesters et
al., 2005). VOD is primarily sensitive to vegetation water content,
including both leafy and woody components (Guglielmetti et al., 2007;
Jackson and Schmugge, 1991; Kerr and Njoku, 1990). Unlike the traditional
optical-based vegetation indices, such as NDVI, VOD is minimally influenced
by the atmospheric conditions due to its longer wavelength and stronger
penetration capacity (Jones et al., 2009). However, it has a coarser spatial
resolution (0.1
There are two objectives of this study. The first is to explore the possibility of whether adding the VOD (responding to vegetation water content) will improve DOC prediction when compared with existing NDVI-based (responding to vegetation greenness) DOC prediction model. The second is to implement the satellite-based DOC estimation (both our and existing models) into GFDI to investigate whether better fire severity predictions can be achieved in grassland environments.
The NDVI dataset used here is derived from the Moderate Resolution Imaging
Spectroradiometer (MODIS) 8-day surface reflectance product (MOD09A1)
on board the Terra satellite (Vermote and Vermeulen, 1999). The MOD09A1
product used here for computing NDVI is an 8-day product, which has less
noise than the daily product, and its spatial resolution is 0.005
The 8-day NDVI data are derived from the MODIS reflectance dataset using the
following:
The VOD dataset used here is retrieved from the Advanced Microwave Scanning Radiometer – Earth Observing System (AMSR-E) and derived using the Land Parameter Retrieval Model (LPRM) approach from which soil moisture and VOD are retrieved simultaneously (Meesters et al., 2005; Owe et al., 2001). Several assumptions are made in the LPRM approach, including canopy surface temperature equal to soil surface temperature, a constant single scattering albedo, same vegetation parameters for both horizontal and vertical polarisations, and minimal effect of surface roughness (Meesters et al., 2005; Owe et al., 2001). Uncertainties in soil moisture and VOD retrievals are expected with these assumptions. The evaluation of LPRM soil moisture over Australia showed that the temporal patterns of satellite-based and in situ soil moisture agree very well (Draper et al., 2009; Gevaert et al., 2016). This agreement suggests a reasonable separation of temporal patterns of soil moisture and VOD, while uncertainties may exist in the absolute magnitudes of these two variables.
Example vegetation optical depth (VOD) and normalised difference
vegetation index (NDVI) time series
VOD has a spatial resolution of 0.1
VOD data, which have a near daily temporal resolution, are converted to an
8-day average product to reduce noise and ensure complete coverage over
Australia (10 to 45
An example comparison time series of VOD and NDVI from July 2002 to June 2011
at one of the observed DOC sites, Silent Grove, WA (17.131
The global 0.05
A burned area product from MODIS is acquired for further evaluation of the recalculated GFDI from satellite-based DOC results. The monthly archived MODIS burned area map re-projected for Australia is obtained from Remote Sensing at the NCI site (Paget and King, 2008). There are two separate MODIS burned area products: the MCD45A1 and the MCD64A1. The MCD64A1 burned area product is preferred over MCD45A1, since it was proven to be more accurate (Andela and van der Werf, 2014; Padilla et al., 2015; Ruiz et al., 2014). Its spatial specification is exactly the same as the MODIS reflectance dataset, with temporal availability from August 2000 onwards. To ensure high quality of the burned pixels, only pixels with the valid data flag from the provided quality control file are included in the analysis. Over 99 % of pixels from mid-2002 to mid-2011 are classified as unburned. To reduce the number of prescribed burns and other low power anomalies detected by the burned area product, a fire radiative power (FRP, unit: MW) from MODIS active fire product (MCD14ML) is used to mask out low severity fires.
MCD12C1 land cover type map for Australia (Hansen et al., 2000). The locations of 23 valid observed degree of curing (DOC) sites are marked with crosses.
The observed grassland DOC data were provided by Bushfire and Natural Hazards
Cooperative Research Centre and its partner agencies (Project reference:
To identify robust relationships between the site-observed and remotely
sensed DOC, a number of site criteria must be met. Sites meeting these
criteria were used for calibration of the VOD and NDVI satellite-based DOC
models, while all of the valid records were used for evaluation. One major
factor in deciding the site selection is the land cover properties of the
observed DOC site. The 0.05
According to the land cover information for each 0.1
All sites were also examined to ensure a negative correlation between VOD
and the in situ DOC data. That is, since VOD is a proxy for water content in
aboveground biomass, an overall negative correlation between VOD and curing
is expected. If this is not the case, then there is likely some other
activity within the 0.1
In addition, there are eight sites (Umbigong, ACT – native grass; Kilcunda, VIC – improved pasture; Tooradin, VIC – improved pasture; Tooradin North, VIC – improved pasture; Caldermeade Park, VIC – improved pasture; Kaduna Park, VIC – improved pasture; Hobart Airport, TAS – native grass; and Jerona, QLD – native grass) in which VOD data are not available. Most of these are due to sites being located near the coast or a large body of water, where the VOD signal is strongly influenced by the water itself. With the remaining 23 out of 37 sites, several site selection criteria were applied for the calibration phase. The criterion used here to maintain consistency in observation time series requires sites to have at least eight consecutive records – records are considered consecutive when they are separated by no more than 15 days. Only the consecutive series of records within the selected sites are included in the analysis for the calibration phase. This ensures that the derived model contains the temporal evolution of DOC within years. Only 5 out of 23 sites are retained for this group, containing a total of 122 (out of 238 total) observations. The selected sites are Majura, ACT (improved pasture); Tidbinbilla, ACT (mixed grass); Ballan, VIC (improved pasture); Murrayville 1, VIC (native grass); and Murrayville 2, VIC (improved pasture). Multiple linear regression models of VOD and NDVI were then calibrated with the observed curing from the final selected sites.
To further assess the usability of the satellite-based curing acquired from
the VOD and NDVI model, the GFDI is computed. Additional meteorological data
needed for GFDI computation are dry bulb or maximum temperature, 15:00 relative
humidity, maximum wind speed, and fuel load (Purton, 1982). Since
fuel load is often set as a constant value of 0.45 kg m
As indicated by past studies (Dilley et al., 2004; Peterson et al., 2008),
NDVI has a significant relationship with live fuel moisture content and DOC.
In addition, the NDVI dataset has a spatial resolution of 0.005
Utilising both VOD and NDVI datasets, the following multiple linear
regression equation for estimating DOC can be expanded from Eq. (3) as
We acquire existing satellite-based DOC products available from the Australian Bureau of Meteorology and compare their performance with our model. There are five
models available, four based on Newnham et al. (2010) and one based
on Martin et al. (2015). We decided to test only one of Newnham's
models – the one with the best overall RMSE (Method B) –and Martin's model
(MapVic). Both Method B and MapVic DOC models are described in Eqs. (5)
and (6), as shown below:
To compare both Method B and MapVic model performance with our model, we evaluate them using the same observed DOC sites and evaluation methods. We also computed recalculated GFDI with both Method B and MapVic DOC and assessed their burned area prediction capability.
Scatter plot of observed degree of curing (DOC) from five calibration sites against VOD, NDVI, and combined VOD and NDVI terms.
Several revisions of GFDI were made by past studies (Noble et al., 1980;
Purton, 1982). The GFDI revision used in this paper is modified Mark 4 GFDI,
since it is the grassland fire danger assessing system that is generally
being used by the Bureau of Meteorology (Sharples et al., 2009b). Originally,
the fire danger rating system was presented in a circular slide rule. A
mathematical representation of modified Mark 4 GFDI was derived
from the circular meter and can be expressed as follows (Purton, 1982):
The GFDI is computed on the basis of meteorological input data and either a constant DOC at 100 % or satellite-based dynamic curing values. These different GFDI datasets along with the burned area data (MCD64A1) can be used to examine the changes due to variable DOC spatially and temporally. By pairing up burned and unburned pixels with their associated GFDI pixel, we can assess the number of burned and unburned pixels for each GFDI severity level. Using histogram and receiver operating characteristic (ROC) analysis, the difference between original GFDI with constant DOC at 100 % and recalculated GFDI with satellite-based dynamic DOC can be assessed (DeLong et al., 1988; Zweig and Campbell, 1993).
Across all selected observed DOC sites (excluding the forest areas) from
July 2002 to June 2011, the
Scatter plot of residual observed degree of curing (DOC) from five calibration sites that are unexplained by NDVI against VOD and combined VOD and NDVI terms.
Calibration and evaluation of satellite-based degree of curing (DOC) models derived from vegetation optical depth (VOD) and normalised difference vegetation index (NDVI). Evaluating existing estimated DOC models, Method B (Newnham et al., 2010) and MapVic (Martin et al., 2015), are also listed below.
However, at this level of
These results can be compared to those obtained using existing remotely
sensed DOC estimates which are also shown in Table 1. The MapVic DOC has a
lower
Example satellite-based and site-observed degree of curing (DOC) time
series comparison at Silent Grove, WA (17.131
Using the relationship between VOD, NDVI, and observed DOC from the first model, as stated in Eq. (10), we calculated satellite-based DOC for Australia. Figure 5 presents maps of satellite-based DOC data averaged over the summer periods (December, January, February, DJF) for the years 2002–2003 and 2010–2011. From mid-2002 to mid-2011, the overall average curing for the Australian summer period is the highest during 2003 and the lowest during 2011. Note that the pixels that are classified as any forest types are masked out in white. Comparison time series between satellite-based and site-observed DOC at Silent Grove, WA (same location as shown in VOD and NDVI example comparison in Fig. 1), is also shown at the top of Fig. 4 as an example.
To determine the amount of spatial variation in DOC across Australia, we
computed the standard deviation of all valid DOC estimates across the
continent within a single time step. All areas that are indicated as forests
by the land cover type map are excluded from the analysis. The spatial
variation time series can then be plotted for the available time period of
mid-2002 to mid-2011, as shown in Fig. 6. Note that the continental mean
spatial DOC standard deviation is 20.39 %. This indicates that there is
significant spatial variability in DOC that persists across all years, and
contains a small seasonal component. For a normally distributed variable,
95 % of values would lie within 2 standard deviations, which is
In addition, based on time series of satellite-based curing data, Fig. 7 reveals the spatial distribution of standard deviations calculated for each pixel. It shows that most of the strong temporal variation occurs in the south, especially in the southeast and southwest of Australia. Several areas in the middle of the continent that have unexpectedly high variation are likely due to rare inundation events. The continental mean temporal standard variation is at 11.88 %. Together, Figs. 6 and 7 show the variability in DOC that will impact calculations of fire danger indices.
Spatial standard deviation of estimated degree of curing (DOC) time series from 4 July 2002 to 26 June 2011.
The spatial plot for maximum summer recalculated GFDI from the DOC multiple linear regression model is shown in Fig. 8, where panels a and b are the maps for summer 2003 and summer 2010, respectively. The magnified regions – for example fire events in Weston Creek, ACT, in 2003 and Toodyay, WA, in 2010 – can be seen in panels c and d. The fire locations are marked with a red plus sign. White pixels are forest areas that were masked using the land cover map. Overall, summer 2003 has 4.51 % more areas indicated as severe or higher GFDI than summer 2010. MCD64A1 burned area map (Fig. 9) also suggested that summer 2003 had 91.45 % more severe wildfire counts than summer 2010. It should be noted that high GFDI values do not guarantee a fire as there is no accounting for ignition sources; rather, a higher GFDI value indicates higher probability of fire ignition. If a grassland fire were to start it would spread faster compared to low GFDI values, given no fire suppression activity. Further complicating comparison of Figs. 8 and 9 is the presence of prescribed burns that are deliberately done during low to moderate GFDI conditions; in addition, some fires shown in Fig. 9 occur in forested areas where GFDI is not applicable. Nevertheless, they provide a picture of the interannual spatial variability in both GFDI and burned area.
Spatial standard deviation of estimated degree of curing (DOC) by season, month, and land cover type from 4 July 2002 to 26 June 2011.
Temporal standard deviation of estimated degree of curing (DOC) map from 4 July 2002 to 26 June 2011.
The time series plots of recalculated GFDI at Weston Creek, ACT, and
Toodyay, WA, for the 2003 and 2010 fire events are
shown in Fig. 10. The black line represents the recalculated GFDI from
variable DOC, while the dashed, light green line is for original GFDI with
constant DOC at 100 %. These locations are marked with red plus signs on the spatial maps (Fig. 8). Note that the original GFDI time
series peaks every year, whereas the recalculated GFDI with variable curing
time series shows sudden peaks in the days near major fires. The Weston
Creek fire was part of the 2003 Canberra bushfire complex, where multiple
fires merged and rapidly propagated from 18 to 22 January 2003, burning 1600 km
Using a burned area observation dataset from MODIS (MCD64A1), we test the
effectiveness of GFDI with variable curing in increasing the probability
that fires will occur in high GFDI severity levels compared to the
probability that fires will occur in low–moderate GFDI severity levels. Low-intensity fires, such as prescribed burns, are removed from the burned area
observation by using the FRP provided in MODIS active fire product (MCD14ML)
to mask out burned area that have low FRP. We assume any burned area with
FRP lower than 100 MW to be unburned (associated with low–moderate GFDI
risk). At each burned and unburned daily data point, the corresponding daily
GFDI was calculated. The GFDI histogram in Fig. 11 shows the frequency of
satellite-based recalculated GFDIs and constant-based (DOC
Maximum estimated Grassland Fire Danger Index (GFDI) for the summers
(December, January, February) of 2002–2003
We can evaluate the performance in correctly assigning burned and unburned area for both recalculated and reference GFDI by using the concept of ROC, as described earlier. Assume that the MCD64A1 burned area map represents the true condition and that the GFDI severity level represents the predicted condition, where the prediction is positive when GFDI level is classified as high or above for a burned area and low–moderate for an unburned area. Table 3 shows the contingency table, including both type I (unburned area with high or above GFDI level; false positive) and type II (burned area with low–moderate GFDI level) errors. Though recalculated GFDI has a lower true positive rate of correctly assigning burned area than reference GFDI (0.86 vs. 0.95), it is much better at assigning unburned area correctly, i.e. lower false positive rate (0.38 vs. 0.53). Overall accuracy for recalculated GFDI is higher than the reference GFDI (0.62 vs. 0.47).
MCD64A1 burned area map (Ruiz et al., 2014) during summer (December,
January, February) 2002–2003
Grassland Fire Danger Index (GFDI) time series plot at Weston Creek,
ACT, from July 2002 to June 2011
Referenced and recalculated Grassland Fire Danger Index (GFDI) severity and burned–unburned area contingency table for satellite-based degree of curing (DOC) derived from vegetation optical depth (VOD) and normalised difference vegetation index (NDVI). Reference GFDI is computed from constant DOC at 100 %, while recalculated GFDI is computed from satellite-based DOC.
Both Method B and MapVic DOC are then used to compute recalculated GFDI and
compare with burned area observation dataset in the same manner as our DOC
model. From the ROC analysis in Table 4 for Method B and MapVic recalculated
GFDI, we found that even though Method B has the best DOC evaluation results
(highest
Grassland Fire Danger Index (GFDI) severity level histograms at burned and unburned areas over Australia during 4 July 2002 to 26 June 2011 where the dark and light blue shaded bars are recalculated GFDI with satellite estimated variable degree of curing (DOC), while the green and yellow shaded with diagonal hatch bars are reference GFDI with constant DOC at 100 %.
Previous studies that derived satellite-based DOC have mostly relied solely
on NDVI as a predictor. In the study by Newnham et al. (2011) various forms
of NDVI were used, including a straight NDVI linear regression and relative
NDVI, as shown in Eq. (2). Their results suggested that a linear regression
model based on NDVI alone reproduced site observations with an
An earlier study for estimating DOC directly with NDVI yielded even smaller
RMSE of up to 6.3 %, but that particular study is focused on data from
only three different sites, within a limited study area of 1 km
Recalculated Grassland Fire Danger Index (GFDI) severity and burned–unburned area contingency table for degree of curing (DOC) computed with Method B (Newnham et al., 2010) and MapVic (Martin et al., 2015) model. Reference GFDI is computed from constant DOC at 100 %, while recalculated GFDI is computed from satellite-based DOC.
Though the overall recalculated GFDI from Method B DOC is the best (overall accuracy from best to worst is 0.85 for Method B, 0.67 for our DOC model, 0.61 for MapVic, and 0.48 for GFDI with 100 % constant DOC), we found that it is the worst at detecting burned area correctly (true positive rate from best to worst is 0.88 for GFDI with 100 % constant DOC, 0.74 for MapVic, 0.69 for our DOC model, and 0.09 for Method B). Our VOD- and NDVI-based DOC model (first model) has a good balance in having the second best evaluation result and overall recalculated GFDI accuracy with a decent correct burned area detection rate. We note that GFDI is only an indicator for the level of fire risk and does not guarantee that a fire will occur, even at an extreme danger level. However, the improvement in accuracy indicates that inclusion of time- and space-varying DOC estimates makes it much more likely that areas identified at a GFDI severity level of high or above will burn than a low–moderate severity level area.
It is worth noting here that in an operational setting atmospheric
interference by clouds or smoke will cause gaps in the optical (NDVI) data,
though the VOD data remain unaffected. We also note that while the VOD data
used here were derived from the AMSR-E sensor, which is no longer
operational, VOD data derived from currently operating passive microwave
sensors, such as the Advance Microwave Scanning Radiometer 2 (AMSR2), could
be used in an operational setting. It should also be noted that VOD's
moderately coarse resolution of 0.1
Reducing the chance of incorrectly assigning unburned and burned areas correctly from the ROC analysis made here is purely based on using the burned area map as a true baseline. However, the burned area map may include fires that are deliberately lit in low–moderate conditions, such as prescribed burns and fires that the GFDI is not designed for, such as a fire that burns in forested regions. Prescribed burns and low-intensity fires are, however, minimised by applying low FRP threshold, using information from the MODIS active fire product. The ROC analysis result here is only used to reinforce the idea that using the reference GFDI with constant curing (100 %) leads to overestimating GFDI in some situations and might result in misleading fire danger warnings.
The satellite-based DOC produced here is also at a moderate spatial resolution, which is a limitation of many satellite products. However, DOC in reality can vary over spatial scales much finer than the satellite footprint (less than 500 m). As such, our model should only be used as a guide for dynamic, near daily assessment of grassland curing at coarse to moderate spatial scales. This is also true for other satellite-based DOC models, including Method B and MapVic models.
This study developed an alternative approach for estimating the grassland DOC using a relationship between the observed DOC and satellite-based VOD and NDVI. The satellite-based dataset was evaluated against the observed DOC data, which resulted in a comparable performance with the currently existing optical-based DOC estimation models. Despite the relatively coarse spatial resolution and temporal coverage of VOD and NDVI datasets used in this study, the satellite-based DOC dataset produced from our model has the potential to contribute to the preparedness of fire management agencies and improve fire spread modelling. With a comparable, and arguably more balanced, performance in correctly predicting burned and unburned area through GFDI than currently available satellite-based DOC models (i.e. Method B and MapVic), our model could provide an appealing alternative estimated DOC data for GFDI computations and fire risk modelling.
The following datasets and their associated sources or contact points are as
listed below:
MODIS MOD09A1, MCD16C1, MCD14ML, and MCD64A1 for NDVI computation, land
cover type map, active fire product, and burned area map for Australia are
freely available from NASA via Remote Sensing at NCI (mosaicing and
regridding by CSIRO):
AMSR-E VOD dataset for Australia is available upon request by contacting Yi
Liu: Method B and MapVic DOC products information can be found at
Observed DOC dataset via visual assessment, levy rod, and destructive
methods is available upon request from the following Bushfire and Natural
Hazards Cooperative Research Centre legacy project: Maximum daily observed gridded temperature and vapour pressure dataset for
Australia are freely available from AWAP: ERA-Interim maximum daily reanalysis gridded wind components for Australia
are freely available from ECMWF:
WC, JE, and YL were involved in initial research planning and experiments design. For the rest of the study, including the analysis of the results, paper writing, and manuscript proofreading, all authors (WC, JE, YL, and JS) were involved.
The authors declare that they have no conflict of interest.
We would like to acknowledge the following parties:
NASA for making MODIS data freely available; NCI and CSIRO for making a freely available preprocessed MODIS data for Australia; Bureau of Meteorology for making freely available satellite-based DOC
dataset and continue to generate and distribute the data products set up by
the AWAP project; Bushfire and Natural Hazards Cooperative Research Centre and its partner
agencies for allowing us to use their observed DOC dataset from their legacy
project; AWAP for making daily Australia climate gridded dataset freely available; ECMWF for making ERA-Interim reanalysis gridded dataset freely available.
Edited by: J. G. Pinto
Reviewed by: Célia Gouveia and one anonymous referee