*3.4. Combination of Results from Different Levels*

Table 8 lists how much area of New South Wales and Victoria is covered by the relevant climate zones. Features are considered relevant if they show a trend matching or being close to statistical significance regarding burn extent or severity. Two of these zones—BSk and Cfa—cover significant portions of the two states.


**Table 8.** Percentage of area covered by relevant climate zones, regarding New South Wales and Victoria.

Table 9 lists, equivalent to Table 8, how much of the area of the relevant climate zones is covered by relevant ecological units. It can be seen that unit 1712 covers 36.3% of the area of the BSk climate zone in New South Wales, and even 52.3% of this zone in Victoria. However, this unit contains mostly cropland, and thus the fire activity has to be attributed in large part to agricultural burnings. While it is interesting to note that the burn severity rises on agricultural areas, this study targets the activity of potentially harmful wildfires, and is thus not concerned with controlled, anthropogenic fires. Ecological unit 1712, and with it the BSk climate zone, is therefore considered largely irrelevant for this study. The remaining zone is Cfa, which features a temperate climate with hot summers, and without a dry season. Unit 2268, "Hot Wet Mountains on Non-Carbonate Sedimentary Rock with Mostly Needleleaf/Evergreen Forest", shows the highest trends of all analyzed ecological units, both regarding fire extent and fire severity. The unit features an increasing trend of 0.26% on annual average regarding extent, and 0.39% regarding burn severity. Both trends are shown to be robust, indicated by their statistical significance. A similar trend can also be seen for unit 1652, "Warm Wet Mountains on Metamorphic Rock with Mostly Needleleaf/Evergreen Forest", which shows an increasing burn severity of 0.2% on annual average.

**Table 9.** Percentage of area covered by relevant ecological units, regarding New South Wales and Victoria.


Apart from the agricultural area, all units located within this climate zone featuring increasing severity trends contain needleleaf or evergreen forest.

This development is also discernible in Figure 11, which shows the affected vegetation types and the respective size in hectares for each of the four exposed states and territories, regarding the period of 2000 to 2020. Yearly land cover information of the ESA CCI-LC (Climate Change Initiative—Land Cover) dataset ([39]) has been used in order to derive the present vegetation types for each respective year. For New South Wales and Victoria, it can be seen that forests represent the predominantly affected vegetation type of the burnings in

2019/2020. The presented figures and statistics indicate that ecologically valuable, woody vegetation is increasingly affected in the study area.

**Figure 11.** Total yearly burnt area amount in million hectares for each state, subdivided by land cover type.

#### **4. Discussion**

While the inter-annual variability in fire activity complicates and in some cases prohibits the derivation of statistically significant trends, several expressive conclusions could be drawn for some of the investigated climate zones and ecological units.

Steady increases regarding burn severity could be found for the climate zones BSk and Cfa, which cover significant parts of New South Wales and Victoria. These development could be traced back to several ecological units, residing inside these climate zones. One of those, number 1712, is mostly characterized by agricultural activities and thus considered less relevant. The other ones, located within the Cfa climate zone, indicate pronounced increases in burn severity regarding needleleaf/evergreen forest. This is supported by a time series study of land use/land cover information. In general, the results show that woody vegetation is increasingly affected in New South Wales and Victoria.

Equivalent conclusions have been drawn by Tran et al., who analyzed fire severity for Victoria [15] regarding the period of 1987 to 2017. The authors furthermore stated that the consequences for ecosystem dynamics might be critical, as temperate forests usually adapted to fire could be damaged irreversibly through higher severity burnings.

Several points need to be taken into account regarding the methodology of this study:

First, note that the inter-comparison of the analyzed classes is only possible in a limited manner. The measure of fire severity has a very different expressiveness between arid, tropical and savannah land cover classes, for example. Hammill et al. also found that determining fire severity from satellite imagery for sedge-swamp or heath surface cover is only possible with lower accuracy compared to forests and woodlands [72]. Results are therefore distorted when study areas cover different ecosystem types, meaning that the robustness of the results increases with rising homogeneity of the study area. While the effect of mixed signals cannot be fully eliminated in a large-scale study, it can be mitigated

by analyzing regions of homogeneous climatic conditions or fine-scale ecological regions, as it is done here.

Second, the derivation of burnt area perimeters as well as the assessment of burn severity rely heavily upon the *NDV Idi f f* . This index has been utilized in numerous investigations, and was validated in a variety of studies, for example [73–75]. The index has been used for decades to assess fire severity, and is also actively used today. For instance, it was recently utilized by Mathews et al. as well as Storey et al. to analyze the burn severity of the wildfires in California in 2020 [76,77]. Tran et al., 2020 investigated indices which are commonly used to assess fire severity, regarding the study area of Victoria/Australia. They identified the *NDV I* as one of the optimal spectral indices for mapping fire severity, regarding the forest types of this study area [15,75].

Another index frequently utilized is the Normalized Burn Ratio (*NBR*), which is similar the *NDV I* but relies on the NIR and Short Wave Infrared (SWIR) band combination instead of red and NIR in case of the *NDV I*. This index could be shown to perform similar to the *NDV I* regarding high severity fires, but was superior regarding fire events featuring rather low severity [78]. The reason that the DLR burnt area dataset does not utilize the *NBR* is that this dataset is primarily based on the Sentinel-3 OLCI instrument, which does not feature a band in the SWIR domain. The MODIS instrument does have a SWIR band, however. This one is only available at at a reduced resolution of 500 m, though, opposed to 250 m regarding the red and NIR band. For the conduction of the study, it was decided to utilize the MODIS bands equivalent to the ones available in Sentinel-3 OLCI. This allows a homogeneous methodology at the best available spatial resolution.

Apart from rule-based approaches based on spectral indices, methodologies from the domain of Machine Learning are increasingly used in wildfire science. Collins et al. (2018) [79] used a Random Forest classifier for the determination of burn severity classes, and found a higher detection accuracy compared to index-based approaches. This proceeding, however, requires preceding steps of careful selection and preparation of training data, as well as the actual training of a Neural Network regarding the area of interest and input data to be used. A comprehensive overview of the requirements is given by Collins et al. (2020) [80]. The methodology invoked for the DLR dataset has been designed to be applicable with a variety of optical sensors, and to be operational globally without a preceding training step.

Third, the analyzed time period covers only the months from November to February for each analyzed year, which is the time span the majority of the disastrous burnings happened in the 2019/2020 fire season. The confinement to a subsection of the available input data became necessary because of the massiveness of the complete dataset, which could not have been processed within a reasonable time frame. However, this time range was found to be representative for the fire season regarding the state of Victoria by Tran et al. (2020) [15]. Still, this confinement represents a sub-optimal precondition, as important differences in the seasonality of fire across the study area might be ignored.

Fourth, developments regarding burn severity are dependent on different input factors, and can easily be misinterpreted. These developments can be caused by shifts in the affected vegetation coverage. Woody vegetation features a higher biomass amount compared to shrubland, which will result in a higher severity value when burnt. Furthermore, the spatial extent regarding affected land cover types plays a crucial role, since it proportionally influences the resulting average value. An increase in area of affected woody vegetation can be overcompensated by an even higher increase in area of affected shrubland vegetation.

Finally, it has to be stated that the analyzed time span of 20 years is rather short, with respect to gaining sufficient insight into climate related, long-term developments. This limitation is due to the availability of suitable satellite imagery of the MODIS and OLCI sensors. The available data time range does not allow conclusions regarding the question whether dramatic fire events occur more frequently than in earlier decades. For future studies, it is therefore planned to also incorporate data of the Advanced Very High

Resolution Radiometer (AVHRR) optical satellite sensor [81], which would allow to perform analysis based on a time series covering more than 40 years.

#### **5. Conclusions**

The pronounced, inter-annual variability in fire activity together with the spatial dynamics of wildfires often prohibits the derivation of statistically significant trends. The majority of the dramatic burnings occurring mostly in New South Wales and Victoria in the 2019/2020 fire season must be regarded as an exception. However, several meaningful, robust trends regarding fire severity and extent could be derived for some of the affected area, mostly located in the coastal area of northern New South Wales.

Two different climate zones have been found to be responsible for the rising burn severity trends in New South Wales and Victoria. The trends within the BSk zone, which is defined by cold, arid steppe conditions, is mostly due to fire activity in the ecological unit 1712, which contains mostly cropland. The fire activity in this zone is therefore attributed mainly to agricultural burnings, which are not examined in this study. The coastal Cfa climate zone, featuring temperate conditions with hot summers and without a dry season, however, was shown to be increasingly affected by potentially harmful wildfires. The rising trends of fire extent and severity could be traced back to several ecological units. All these units, except for one which is used agriculturally, share the characteristic of being covered by needleleaf/evergreen forest. While the extensive burnings of the 2019/2020 fire season clearly are exceptional, some of the fire activity took place in these forested areas, and is thus regarded to be in parts connected to a steady, long-term upward trend in fire extent and severity.

It is concluded that the forested regions of the Australian East coast residing within the Cfa climate zone (temperate, no dry season, hot summer) will most likely be increasingly affected by wildfire activity in the future. Specifically, this refers to the area covered by, first, ecological unit 2268 (Hot Wet Mountains on Non-Carbonate Sedimentary Rock with Mostly Needleleaf/Evergreen Forest), which features a mean annual increase of 0.26% in fire extent and 0.39% in fire severity. Secondly, this addresses the area covered by ecological unit 1652 (Warm Wet Mountains on Metamorphic Rock with Mostly Needleleaf/Evergreen Forest), which shows a mean annual increase of 0.25% in fire extent and 0.22% regarding fire severity.

The DLR-GZS burnt area dataset, on which this study is based, could be shown to be a valuable asset for wildfire related studies, such as burn severity time series analysis. To the knowledge of the authors, it is the only large-scale, decadal burnt area dataset including detailed burn severity information to this point.

**Author Contributions:** Conceptualization, M.N., G.S. and T.R.; methodology, M.N.; software, M.N.; validation, M.N. and N.M.; formal analysis, M.N.; investigation, M.N.; resources, M.N.; data curation, M.N.; writing—original draft preparation, M.N.; writing—review and editing, M.N., N.M., G.S. and T.R.; visualization, M.N.; supervision, G.S. and T.R.; project administration, M.N., G.S. and T.R.; funding acquisition, G.S. and T.R. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research received no external funding.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** The MODIS MOD09/MYD09 product is available from the NASA Land Processes Distributed Active Archive Center (LP DAAC), accessed on 8 April 2020 http: //e4ftl01.cr.usgs.gov. Sentinel 3A/B OLCI data can be obtained from the Copernicus Open Access Hub, accessed on 10 June 2020 https://scihub.copernicus.eu. Finally, Active Fire data used as auxiliary information in the described methodology can be acquired from the NASA Fire Information for Resource Management System (FIRMS), accessed on 17 May 2020 https://firms.modaps.eosdis. nasa.gov.

**Conflicts of Interest:** The authors declare no conflict of interest.
