*Article* **Using Long-Term SAR Backscatter Data to Monitor Post-Fire Vegetation Recovery in Tundra Environment**

#### **Zhiwei Zhou 1,2,\*, Lin Liu 2, Liming Jiang 1, Wanpeng Feng 3,4 and Sergey V. Samsonov <sup>5</sup>**


Received: 24 August 2019; Accepted: 21 September 2019; Published: 25 September 2019

**Abstract:** Wildfires could have a strong impact on tundra environment by combusting surface vegetation and soil organic matter. For surface vegetation, many years are required to recover to pre-fire level. In this paper, by using C-band (VV/HV polarization) and L-band (HH polarization) synthetic aperture radar (SAR) images acquired before and after fire from 2002 to 2016, we investigated vegetation change affected by the Anaktuvuk River Fire in Arctic tundra environment. Compared to the unburned areas, C- and L-band SAR backscatter coefficients increased by up to 5.5 and 4.4 dB in the severely burned areas after the fire. Then past 5 years following the fire, the C-band SAR backscatter differences decreased to pre-fire level between the burned and unburned areas, suggesting that vegetation coverage in burned sites had recovered to the unburned level. This duration is longer than the 3-year recovery suggested by optical-based Normalized Difference Vegetation Index (NDVI) observations. While for the L-band SAR backscatter after 10-year recovery, about 2 dB higher was still found in the severely burned area, compared to the unburned area. The increased roughness of the surface is probably the reason for such sustained differences. Our analysis implies that long records of space-borne SAR backscatter can monitor post-fire vegetation recovery in Arctic tundra environment and complement optical observations.

**Keywords:** arctic tundra fire; vegetation recovery; C- and L-band SAR; SAR backscatter

#### **1. Introduction**

Climate change is occurring rapidly at high latitudes, where surface air temperature has increased at twice the rate of the rest of the globe in the past several decades [1–3]. Such warming has resulted in pronounced environmental changes in the Arctic [4–6]. During that period, it has caused a dramatic increase in the frequency of large wildfires in boreal and tundra regions [7–9]. Wildfires strongly affect ecosystem in tundra environments, such as carbon balance of tundra [10,11], land surface albedo [12], active layer thickness change [13], vegetation shifts [14], permafrost thaw and thermokarst development [13,15], and even wildlife habits [16].

By burning surface vegetation and soil organic matter, fire can rapidly transfer large stocks of aboveground biomass and below-ground soil carbon to the atmosphere and change the ecosystem carbon balance in the short-term. Additionally, the surface albedo balance is also changed by burning some, or all, of the insulating layer of moss and soil organic matter, which shields the underlying permafrost from warm summer temperatures [12]. Then the active layer of burned zone is destabilized in summer, which may cause decomposition of deeper soil carbon stored below ground and results in additional carbon transfer to the atmosphere in the long-term after the initial combustion [17–19]. Thermokarst development induced by fire may accelerate the above process and results in carbon transfer to the atmosphere and to the tundra gully as dissolved organic carbon [20,21]. Since soil carbon underlain by permafrost in the Arctic is nearly twice the total amount of carbon as is in the atmosphere [22,23], increased fire frequency and severity in Arctic tundra could act as positive feedback to global climate warming [24].

Wildfire regimes could exert landscape-scale controls on vegetation structure and composition [25]. Wildfire also directly affects seedling recruitment success and migration [26]. Fire may alter vegetation successional trajectory and prepare seedbeds for shrub or tree invasion. Therefore, increased fire frequency and severity in Arctic tundra is expected to facilitate both shrub expansion and tree migration [25,27]. Wildfires also consume caribou forage lichens and exert strong effects on post-fire lichen recovery for decades [28]. Such a change in successional trajectory could alter the structure of the vegetation and ecosystem functions, the ecosystem may undergo a regime shift into an alternate state, and ecosystem services may be degraded [29].

Thus, information on post-fire vegetation regrowth in tundra is of great interest for scientists and land managers dealing with environmental issues (e.g., carbon budget, climate impact, ecosystem services). For monitoring post-fire vegetation recovery in expansive and remote tundra area, satellite remote sensors could be suitable tools to conduct the temporal analysis, since they are suitable to continuously collect data for long-term monitoring and temporal analysis in a large area. Tanase et al. [30] suggest that the suitability of remote sensing data is determined by two factors: (a) sensitivity to the changes in radiometric response due to the development of vegetation and (b) the spatial resolution of satellite imagery, which provides the ability to identify the features of the response patterns.

Normalized Difference Vegetation Index (NDVI) images derived from optical satellite observations have been most frequently used measurements for monitoring, analyzing, and mapping vegetation changes in time and space. However, NDVI is more sensitive to changes in leaf area than to changes in structure and overall biomass. The relationship between NDVI and leaf area index (LAI) varies both seasonally and inter-annually. Furthermore, NDVI saturates at high values of LAI [31]. This implies that NDVI is only sensitive to vegetation development in the first few years after fire disturbance. In a study on arctic tundra fires recovery by NDVI, Barrett et al., [14] showed that NDVI quickly recovered to the range of pre-fire level in 3 years.

Imaging radars may have the potential to complement traditional optical data in mapping vegetation recovery in the terms of vegetation structure [32–39], because the radar backscattered signal is highly sensitive to structural properties of the vegetation. In addition, radar sensors have all-weather, and day and night observing capability (independent of solar radiation). Furthermore, as long as more than 25 years' archives of spaceborne synthetic aperture radar (SAR) data and planned future SAR missions, as well as global coverage, provide potential to monitor post-fire vegetation recovery in Arctic tundra.

The radar backscatter from vegetation is the result of complex interactions between microwave electromagnetic energy and the ground and vegetation scattering components in a resolution cell. The intensity of backscatter from vegetation depends on several factors: (1) the amount and geometric properties of vegetation on site, (2) surface dielectric and roughness properties, and (3) radar frequency, polarization, and look direction [40–42]. Generally, due to the penetration capability, short wavelengths of 2–6 cm, such as X-band mostly interact with leaves/needles, twigs, and small branches, and C-band mostly interacts with leaves and small secondary branches [43]. In contrast, long wavelengths of approximately 24 cm, such as L-band, radar waves interact mostly with primary branches, tree trunks, and even underground surface [41]. There is an approximately linear response of backscatter to increasing forest biomass reaching saturation as a function of forest type and structure [42]. Therefore, radar backscatter is expected to change before and after a fire, as the fire removes most of the vegetation

components and may leave the ground or tree trunks exposed. Specifically, the backscatter increases with burn severity for L-band at HH polarized (SAR data acquired under wet condition) [44], and for X- and C-bands at HH and VV polarizations, while the backscatter decreases with burn severity for all frequency at cross-polarization [45]. The backscatter increases with incidence angle at co-polarization and decreases with incidence angle at cross-polarization in burned sites [45]. For example, using C-band SAR (HV polarization), previous studies show 3–6 dB increase in backscatter coefficients between fire scars and adjacent undisturbed forest and seasonal changes in backscatter coefficients in a burned forest [30,43].

Although SAR data usage for forest recovery after wildfires is well studied, it is limited used for tundra vegetation recovery. Using L-band HV data, it is estimated that approximately 40–50 years are required for the recovery of forest structure after disturbance [30]. The effects of wildfire on tundra vegetation are poorly studied by using SAR data. As post-fire vegetation regrowth in tundra environment was mostly investigated by fieldwork periodically [8,16,25,46], or projection based on post-fire soil biota [26], remotely sensed data are rarely used [14]. Using nearly two-decades ERS-1/2 C-band dataset (VV polarization), Jenkins et al. [47] demonstrated the capability of SAR backscatter in detection and monitoring of fire disturbance in the Alaska tundra. The study suggested that there was a 3-dB backscatter coefficient increase between burned and unburned area, and also four to five years of landscape recovery was observed by SAR backscatters.

The objective of this study is to monitor post-fire vegetation recovery in Arctic tundra environment by using long-term SAR backscatter data. In this study, we not only use C-band dataset but also L-band dataset. Compared to NDVI information, radar backscatter information provides new insights on vegetation recovery regarding the structure and density [30,48].

#### **2. Materials and Methods**

#### *2.1. Study Area*

The Anaktuvuk River Fire (ARF) burned 1039 km2 of Arctic tundra on the North Slope of the Brooks Range in the late summer and early fall of 2007. Figure 1a shows the fire perimeters obtained from the Alaska Interagency Coordination Center Large Fire Database [49]. This fire has an order of magnitude larger than the average fire size in the historic record for the North Slope [16]. The ARF resulted in the release of approximately 2.1 <sup>×</sup> <sup>10</sup><sup>3</sup> kg of carbon to the atmosphere, an amount similar in magnitude to the annual net carbon sinks for the entire Arctic tundra biome averaged over the last quarter of the twentieth century [10]. The fire occurred primarily in the Arctic Foothills physiographic province of the North Slope, where the ecosystem is described as upland, shrubby tussock tundra. The climate in the central North Slope is cold in winter (−25 ◦C mean high in January) and relatively cool in summer (20 ◦C mean high in July, vegetation growing season). The mean annual temperature and precipitation of this study area are −10 ◦C and 30 cm, respectively [10,16].

Figure 1b shows differenced Normalized Burn Ratio (dNBR) calculated from a pair of pre-fire Landsat Enhanced Thematic Mapper-plus (ETM+) image acquired on 30 June 1999, and post-fire Landsat Thematic Mapper (TM) image acquired on 14 June 2008 [50]. The dNBR values indicate burn severity within the fire boundary: 11% of the area did not burn (dNBR < 300), while within the burned area, approximately 47% of the ARF burned at high severity (dNBR > 600) and 35% and 18% at moderate to low severity (300 < dNBR < 600) [16]. Three sites located in severely and moderately burned, and unburned area were chosen to examine the SAR backscatter change before and after fire (Figure 1b). These three sites were placed less than 7 km away to each other and had similar pre-fire vegetation features [51].

**Figure 1.** The location and burn severity of the Anaktuvuk River Fire (ARF). (**a**) the False-color Landsat-5 image acquired on 17 July 2008, the pink polygon is the fire boundary and the red polygon of upper left map shows the location of the fire area in Alaska. (**b**) differenced Normalized Burn Ratio (dNBR) within the burn scar [50]. The red, black, and white triangles (labeled as 'a', 'b', and 'c') mark the locations of severely and moderately burned, and unburned sites to be examined in our analysis, respectively.

Prior to the fire, 54% of the burned area was classified as upland moist acidic tundra (soil pH < 5.5), 15% as moist nonacidic tundra (soil pH > 5.5), and 30% as shrubland [16]. The vegetation in the burned area is dominated by the tussock-forming sedges [50]. The dominant shrub types include willow, dwarf birch, Labrador tea, blueberry, and prostrate shrubs lowbush cranberry, bearberry, and crowberry [16].

#### *2.2. Remote Sensing Data*

We only used SAR amplitude images acquired by C-band from the European Remote Sensing-2 (ERS-2) and RADARSAT-2 (RS-2) satellites, L-band from Advanced Land Observing Satellite-1/2 (ALOS-1/2) Phased Array-type L-band SAR (PALSAR) sensors. To compare the difference of SAR amplitude before and after the fire, these images were acquired from 2002 to 2017, with different incidence angles and polarization modes (Table 1). The ERS-2 and ALOS-1 PALSAR datasets contain images acquired before and after the fire, and the RS-2 and ALOS-2 PALSAR datasets spanned nearly 10 years after the fire. ERS-2, RS-2, and ALOS-1 PALSAR Level-1.5 Geo-referenced amplitude SAR images used in this study were provided by the Alaska Satellite Facility (ASF) [52].

All SAR images were firstly radiometrically calibrated and then radiometric variability associated with topography corrected using a Digital Elevation Model (DEM). Next, speckle noise was reduced by using a 3 × 3 window Lee filter. Finally, all SAR images were geocoded to WGS-84 projection. The ERS-2 and ALOS-1 images were processed by the MapReady package (v3.1.24) [53]. The RS-2 and ALOS-2 were processed by the Sentinel Application Platform (SNAP, v5.0). The resulting images were calibrated to sigma nought (σ ◦ , backscatter coefficient) format. A 30 m ASTER DEM was used for radiometric terrain correction caused by local terrain variability in both MapReady and SNAP. Figure 2 shows a series of SAR backscatter coefficient images in both C- and L-bands that spanned the period of before and after the fire. Before the fire, there was no difference in backscatter between the burned and unburned areas (visually represented by brightness in Figure 2a,b). An obviously brighter (increased backscatter) fire scars can be observed in the burned area in 2008 in both the C- and L-band SAR images (Figure 2c,d). Moreover, fire scars become lighter five years after the area burned in C-band (Figure 2e) and after 10 years in L-band SAR (Figure 2f).

**Table 1.** List of synthetic aperture radar (SAR) images used in this study. Note: the resolution is equal in range and azimuth direction.


**Figure 2.** Images of backscatter coefficient that span before and after the fire. The white polygon marks the fire boundary. (**a**,**b**) present the images acquired from ERS2 and ALOS-1 before the fire. (**c**–**f**) present the post-fire images acquired from ERS2, ALOS-1, RS-2, and ALOS-2, respectively.

Due to the inherent speckle noise caused by the nature of the SAR systems, a single pixel of SAR data should not be used to directly relate to field variable. Instead, a group of pixels in azimuth and range direction (called multi-looking process) are usually averaged to mitigate speckle noise. In this study, for all C- and L-band data, a mean of approximately 150 m × 150 m (e.g., 3 × 3 and 12 × 12 looks in both azimuth and range direction for RS-2 and ALOS-1 based on the spatial resolution) area centered on the unburned, moderately burned, and severely burned sites are chosen as the control, moderate, and severe burn sites, respectively (see their locations in Figure 1b). Five-year (2002–2007) pre-fire data were used to compare with 6-year (2008–2010, 2014–2017) post-fire data.

To aid the interpretation of changes of SAR backscatters, NDVI of the whole fire zone and soil moisture collected from the severely burned, moderately burned, and unburned zones are used as ancillary information. We used the NDVI products generated from the Moderate Resolution Imaging Spectroradiometer (MODIS) and distributed by the Land Processes Distributed Active Archive Center (LP DACC) of NASA/USGS [54]. More specifically, the MOD13 Q1 products obtained during growing seasons (June, July, and August) from 2002 to 2016 were used. The accuracy of this product is now within ±0.025, and the derived NDVI accuracy is within ±0.025 [55]. We chose these days because mid-July is close to the end of the growing season. The MOD13 Q1 product is a global product with a spatial resolution of 250 m and a temporal interval of 16 days.

#### *2.3. SAR Backscattering Mechanisms before and after a Fire*

In order to evaluate the sensitivity of radar backscatter to the post-fire vegetation regrowth, it is important to identify the dominant scattering mechanisms before and after a fire. Figure 3 illustrates the dominant backscatter scattering mechanisms in tundra environment, including pre-fire volume scattering (Figure 3a) and post-fire surface scattering (Figure 3b) [43]. Before a fire, the C-/L-band radar waves interact with the canopy of the tussock-forming sedges, in this case, most of the C-/L-band radar backscatter originates from the volume scattering with the canopy layer (Figure 3a). Thus, the backscatter is more sensitive to canopy changes than the changes of ground moisture and roughness. After a fire, the vegetation and/or part of organic soil are burned, leaving the ground exposed. In this case, radar waves primarily interact with the ground surface (Figure 3b). In such case of surface scattering, the backscatter strongly depends on the roughness and soil moisture content of the ground. In a few years after a fire, the radar waves interact more and more with newly regenerated vegetation. Then the effects of soil moisture and roughness on backscatter gradually diminish because of the attenuation due to the growing vegetation. Once the vegetation density and structure reach the pre-fire level, the volume scattering becomes the dominant mechanism again (Figure 3a).

**Figure 3.** Dominant radar backscattering mechanisms in tundra environment. (**a**) volume scattering within tussock-forming sedges before a fire, (**b**) surface scattering from rougher ground after a fire.

#### **3. Results**

#### *3.1. NDVI Changes*

Figure 4 shows the NDVI on 12 or 13 July between 2002 and 2016. Generally, sparse vegetation may result in moderate NDVI values (approximately 0.2–0.5). High NDVI values (approximately 0.6–0.9) correspond to dense vegetation. A clear decrease of surface greenness can be observed following the ARF in 2007 (Figure 4g), and the greenness quickly recovered by 3 years after the fire (Figure 4i).

**Figure 4.** Maps of Normalized Difference Vegetation Index (NDVI) from July 2002 to July 2016. The black polygon in each figure marks the fire boundary. Note that the NDVI image on 13 July 2014 is contaminated by cloud cover.

Figure 5 presents a quantitative analysis of NDVI changes at the severely burned and unburned sites. Each site presents an area of approximately 1.25 km × 1.25 km. At both sites, NDVI showed a clear growing pattern (between 0.5 and 0.7) and reached the highest value in mid-July or early August (Figure 5a). The pre-fire NDVI values were similar at both sites. After the fire, an obvious reduction of NDVI can be observed at the burned site in 2008 (Figure 5a), and the difference between the severely burned and unburned sites was up to 0.2 (Figure 5b). Then a quick recovery occurred in the subsequent 3 years. The post-fire NDVI difference reached zero again from 2010 onward.

**Figure 5.** Time series of NDVI at the severely burned and unburned sites (**a**) and their differences (**b**) severely burned minus unburned. The vertical red line marks the end of the ARF.

#### *3.2. SAR Backscatter Changes*

In Figure 6, we present plots of mean SAR backscatter pre- and post-fire trend. Figure 7 shows Cand L-band backscatter difference (only in growing season) between burned and unburned, and its corresponding statistics analysis is plotted in Figure 8. Similar annual and inter-annual temporal trend can be observed for all sites and bands. The backscatter coefficient changes seasonally from −18 dB to −9 dB for L-band (Figure 7a) and from −25 dB to −12 dB for C-band (Figure 7b).

**Figure 6.** Time series of backscatter coefficients at the severely and moderately burned and unburned sites. (**a**) for ALOS-1 and ALOS-2, (**b**) for ERS-2 and RS-2. The red vertical line marks the end of the fire. The gray bars denote the growing seasons (June, July, August). The gray dashed lines separate ALOS-1 and ALOS-2 in (**a**), ERS-2 and RS-2 in (**b**).

**Figure 7.** Time series of backscatter difference between burned and unburned sites. (**a**) for ALOS-1 and ALOS-2, (**b**) for ERS-2 and RS-2.

**Figure 8.** Statistical summary of the backscatter differences. For each box, the central line is the median, the upper and lower edges of the box are the 25th and 75th percentiles, the whiskers represent the minimum and maximum values, and the plus symbols mark the outliers. (**a**,**b**) present ALOS-1/2 and ERS-2/RS-2 backscatter difference between severe and unburned area, (**c**,**d**) present the ALOS-1/2 and ERS-2/RS-2 backscatter difference between moderate and unburned area.

We observe the following temporal characteristics of backscatter changes before the fire. Firstly, the temporal trend characteristics of severely and moderately burned sites are very similar to those of the unburned site for both C- and L-bands (Figure 6). Secondly, in the growing seasons before the fire, most of the C- and L-band backscatter coefficients at the severely and moderately burned sites showed

little dynamics: close to 0 dB variation for L-band and <1 dB for C-band (Figures 7 and 8). Thirdly, the backscatter coefficients from both C- and L-band showed strong seasonal changes, increased in summer and decreased in winter. This is mainly caused by seasonal changes in vegetation, soil moisture, and soil freeze/thaw status. In the fall, the soil and vegetation freeze and the leaves fall from deciduous shrubs, reducing the soil moisture and overall backscattering, which causes backscatter to decrease. During the summer, the trend of soil and vegetation changes reverses, resulting in increasing the soil moisture and over backscattering, causing backscatter increase in summer. Combining the above three points, these suggest that the backscatter is mainly from the similar vegetation canopy property in the summer.

We observe the following temporal characteristics of backscatter coefficient changes after fire. Firstly, comparing to backscatter coefficient value in growing season before fire, a clear increase in backscatter coefficient value occurred in growing season after fire for both C- and L-band between 2008 and 2010 (Figures 7 and 8). During the growing season, the mean backscatter coefficient differences between severely burned and unburned site is nearly 3 dB (with maximum of 5.5 dB) and 1 dB (with maximum of 2 dB) for ERS-2 and RS-2 C-band, respectively; nearly 3 dB (with maximum of 4.4 dB) and 2 dB (with maximum of 2.5 dB) for ALOS-1 and ALOS-2 L band, respectively. The differences between moderately burned and unburned site were relatively lower for both C- and L-bands. Secondly, different change patterns for C- and L-bands can be observed between 2008 and 2016. C-band backscatter coefficient in growing season decreased from 2008 to 2010 (Figure 7b). Furthermore, a clear drop of backscatter values can be found from ERS-2 records to RS-2 records. This decline is contributed by polarization mode (HH to HV) and incidence angle (23 degree to 27 degree) change after the fire. However, there was no decreasing trend after 2012 (Figure 7b), and three sites present almost similar temporal change pattern after 2012 (Figure 6b). This suggests that after 5-year regrowth, radar waves from the severely and moderately burned sites interacted with similar vegetation canopy level as unburned site again.

For L-band, there was no decreasing trend of backscatter coefficient in severely and moderately burned sites between 2008 and 2010 during the growing season (Figures 6a and 7a). This suggests the vegetation in the severely and moderately burned sites did not recover to the similar state as in unburned site in 2010. The radar signal is primarily contributed by the roughness and soil moisture content of the ground in this stage. However, a slow decreasing trend can be observed between 2008 and 2016, and about 2 dB difference between unburned and moderately/severely sites still can be found in 2016 (Figures 7a and 8a,c). This is mainly because the radar signal interacts with the regenerated vegetation canopy, and the effects of roughness and soil moisture on backscatter gradually diminished. We should note that there is 1.2 degrees (at mid swath) difference in the incidence angle between ALOS-1 and ALOS-2 SAR acquisitions. Since they are in the same HH polarization modes, the different observation angle could have limited influence on SAR backscatter values after the fire.

#### **4. Discussion**

#### *4.1. Inconsistency between NDVI Trend and Field Observations*

In the ARF burned area, the NDVI changes suggest approximately 3 years of vegetation recovery. Barrett et al. [14] reported a similar 3-year NDVI recovery after another tundra fire on the western North Slope of Alaska. However, field investigations over three years after the ARF found the vegetation cover in the burned areas was still sparser than in the unburned areas (35–50% cover for graminoid and 33–42% cover for shrubs). Moreover, field investigations found that the most significant change in vegetation species in the burned areas was graminoid, whose cover increased from 11% in 2008 to 45% in 2010 [16]. Shrub cover also increased in the burned areas, but slowly, from 4% to 11% for deciduous shrubs, and from 4% to 13% for ericaceous shrubs [16]. These field data suggest that the vegetation did not return to pre-fire level after three years of recovery, which contradicts the NDVI observations.

#### *4.2. SAR Backscatter before and after the Fire*

The change of surface components due to fire causes SAR backscatter coefficient change. Our results show that the backscatter value of severe burn area increased up to 5.5 dB and 4.4 dB after the fire for C- and L-band, respectively. Similar backscatter coefficient differences are also found in different Alaska tundra fire sites using ERS-1 C-band dataset [47]. Over the three years of recovery, the L-band backscatter coefficient showed less change than the C-band (Figures 7 and 8). One possible reason is that regenerated vegetation is not dense as pre-fire. L-band radar waves can penetrate these newly regrown land cover, backscatter mainly still depends on roughness and soil moisture of the ground. While C-band backscatter from severely and moderately burned sites indicate a slow decrease trend with the regrowth of vegetation, but it does not return to the same level as the pre-fire level. C-band radar waves interact with new regenerated vegetation, resulting in reduction of the effect of roughness and soil moisture on the backscatter. Combining C- and L-band backscatter different changes over the three years of recovery, we could infer that: (1) the vegetation recovery did not return to the pre-fire status in vegetation structure and density; (2) C-band is more sensitive than L-band to the vegetation recovery in tundra environment.

After nearly 10 years of vegetation regrowth, the backscatter from RS-2 C-band and ALOS-2 L-band showed different change pattern. The C-band backscatter coefficient of all the sites showed almost the same changing pattern after 2011. This implies that the C-band radar waves returned from three sites interacts with a similar surface property after 2011. In other words, the vegetation property in severe burn and moderate burn had recovered to a similar level as in unburned site in 2012. The L-band backscatter from the severe burn and moderate burn sites was still about 2 dB higher than that from the unburned site. Using multitemporal airborne Light Detection and Ranging (LiDAR) dataset acquired in 2009 and 2014, Jones et al. [15] found surface roughness increased in burned area caused by thermokarst development. Therefore, the soil moisture and increased roughness probably contribute to the higher backscatter value in severely and moderately burned sites for ALOS-2 L-band. As L-band has stronger penetration capability, ground roughness and moisture have stronger effects on the returned radar signal than C-band. Combining these observations, we are able to infer that it takes 5 years for vegetation in the burned area to recover to a similar level as unburned site. We note that Jenkins et al. (2014) also suggested that 4–5 years are required for landscape recovery in tundra environment.

#### *4.3. Limitations of SAR Data*

However, compositions or new species vegetation recovered after the fire are unable to differentiate by SAR data. Fieldwork is necessary to investigate the vegetation compositions or shifts. A severe burn is able to facilitate colonization by new species when there is strong combustion of the pre-fire seedbed and high plant propagates. A 17-year study of post-fire vegetation recovery in Arctic tundra shows that there is abundance of graminoid species and an absence of Betulanana [14]. With an increase in fire activity and the conditions that contribute to increased fire severity, fire-related shifts in vegetation type will become more common [14]. Therefore, combining SAR imagery and ground observations could be an effective way to investigate the post-fire vegetation recovery in Arctic tundra environment.

#### **5. Conclusions**

Post-fire vegetation recovery in Arctic tundra environment is important to the balance of tundra ecosystem, such as carbon storage, active layer thickness, and vegetation composition. Effective measurements are required to monitoring the post-fire vegetation regrowth. In this paper, we investigate the potential of C- and L-band SAR backscatter for monitoring post-fire vegetation recovery after the 2007 Anaktuvuk River fire. We found that vegetation in burned area has probably returned to the pre-fire level after 5-year recovery, and C-band is more suitable to monitor such recovery process. The ground soil moisture is a major component of the total backscatter power for both C- and L-bands

after fire. Yet, SAR data alone cannot quantify the vegetation composition and shifts, and additional field observations are required toward a thorough understanding of the composition.

**Author Contributions:** Conceptualization, Z.Z. and L.L.; data curation, Z.Z.; formal analysis, Z.Z. and L.L.; funding acquisition, L.L. and L.J.; investigation, Z.Z. and L.L.; methodology, Z.Z.; project administration, L.L. and L.J.; resources, L.L., W.F. and S.V.S.; supervision, L.L.; validation, Z.Z.; visualization, Z.Z.; writing—original draft, Z.Z.; writing—review and editing, L.L., L.J., and W.F.

**Funding:** This work is supported by the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDA19070104), the National Natural Science Foundation of China (Grant No. 41804017), the Key Research Program of Frontier Sciences, CAS (QYZDB-SSW-DQC027 and QYZDJ-SSW-DQC042), the National Key R & D Program of China (2017YFA0603103), the Hong Kong Research Grants Council Grants CUHK14300815, and CUHK Direct Grant for Research 4053206.

**Acknowledgments:** We thank Benjamin Jones for providing the dNBR data and the Canadian Space Agency for providing the RS-2 ScanSAR Wide images.

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

#### **References**


© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

### *Technical Note* **Quantitative Analysis of Forest Fires in Southeastern Australia Using SAR Data**

**Aqil Tariq 1, Hong Shu 1, Qingting Li 2, Orhan Altan 3, Mobushir Riaz Khan 4, Muhammad Fahad Baqa <sup>5</sup> and Linlin Lu 5,\***


**Abstract:** Prescribed burning is a common strategy for minimizing forest fire risk. Fire is introduced under specific environmental conditions, with explicit duration, intensity, and rate of spread. Such conditions deviate from those encountered during the fire season. Prescribed burns mostly affect surface fuels and understory vegetation, an outcome markedly different when compared to wildfires. Data on prescribed burning are crucial for evaluating whether land management targets have been reached. This research developed a methodology to quantify the effects of prescribed burns using multi-temporal Sentinel-1 Synthetic Aperture Radar (SAR) imagery in the forests of southeastern Australia. C-band SAR datasets were specifically used to statistically explore changes in radar backscatter coefficients with the intensity of prescribed burns. Two modeling approaches based on pre- and post-fire ratios were applied for evaluating prescribed burn impacts. The effects of prescribed burns were documented with an overall accuracy of 82.3% using cross-polarized backscatter (VH) SAR data under dry conditions. The VV polarization indicated some potential to detect burned areas under wet conditions. The findings in this study indicate that the C-band SAR backscatter coefficient has the potential to evaluate the effectiveness of prescribed burns due to its sensitivity to changes in vegetation structure.

**Keywords:** prescribed burns; SAR; fire impact; radar burn ratio; post-fire restoration; change detection

#### **1. Introduction**

Wildfires are a global agent for environmental change [1,2], and prescribed burning is an important strategy to minimize the harmful impact of wildfires globally [3]. Prescribed burning has proven effective in various regions, including Southern Europe [4], North America [5,6], and Australia [7–9]. In order to meet particular environmental protection targets, controlled fires are required for prescribed burning. A fire is set under specific environmental conditions, over an explicit region, and with a specific length, intensity, and spread rate [3]. Prescribed burning was first implemented in forests with the aim of minimizing wildfire risks in Australia in the late 1950s [9]. The prescribed fire is used for: (i) fuel reduction to moderate the adverse effects of wildfires on biodiversity, water, and soil [10–12], and (ii) land management [13], including logging restoration, forest maintenance, and conservation of biodiversity [14].

The intensity of prescribed burns varies between low to moderate in southeastern Australia [15]. The rate of fire spread depends on the seasons. For example, there are

**Citation:** Tariq, A.; Shu, H.; Li, Q.; Altan, O.; Khan, M.R.; Baqa, M.F.; Lu, L. Quantitative Analysis of Forest Fires in Southeastern Australia Using SAR Data. *Remote Sens.* **2021**, *13*, 2386. https://doi.org/10.3390/rs13122386

Academic Editors: Alfonso Fernández-Manso and Carmen Quintano

Received: 10 May 2021 Accepted: 17 June 2021 Published: 18 June 2021

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

fewer fire ignitions in spring, which is from September to December, whereas there are more ignitions in autumn, from March to May, when weather conditions are mild. Broadscale adoption of prescribed burning as a conservation measure has altered the structure and pattern of the fuel period, substantially decreasing the frequency and magnitude of unplanned fires, along with the severity of impacts on land, environmental factors, and vegetation in this area [9,11]. However, the effects of prescribed burning should be monitored to confirm whether defined consumption, area coverage, and fire intensity levels were achieved. In various habitats, coarse woody debris is a crucial structural feature for forest biodiversity and functioning [16], and can be impacted by the burning season, severity, and frequency of prescribed fire treatments [17,18]. Historical data on prescribed burning effects expressed as intensity or severity of fire are important for determining the degree to which the goals of land/wildfire management have been accomplished and also for informing the reporting process aimed at quantifying atmospheric emissions from fire.

The degree or magnitude of environmental changes caused by fire can be defined as the severity of fire and can be calculated by bio-loss [19]. These environmental trends are generally measured by severity indices, including the Composite Burn Index (CBI), to assess the rate of burning for vegetation areas in various sections [20]. Nevertheless, field investigations are expensive and their application over large areas is impractical. Remote sensing methods have been widely used to collect information on post-fire landscape conditions over broad spatial scales [21–23], and spectral indices from multi-spectral remote sensing data are commonly used for assessing burn severity at the spatial grain of the sensor [24,25]. However, reflectance-based indices have difficulty in yielding accurate results due to the influence of multiple factors, such as forest sensitivity, plant phenology, or solar elevation, rather than structure [26,27].

Considering their sensitivity to scattering elements and capacity to provide direct measurements of vegetation structure, active sensors such as radar and LiDAR may overcome the above limitations [28,29]. Various SAR-based techniques are useful to examine the effects of fire on the three-dimensional structures of canopy and the strata of lower plants, as radar waves have the ability to penetrate through canopy. The possibility to estimate the effects of forest fires via radar-based measurements, such as polarimetric decomposition coefficients, interferometric coherence, and backscatter intensities, has been demonstrated in previous studies [30,31]. These studies obtained satisfactory estimation accuracies in wildfire-affected forests and emphasized the need for model calibration with reference datasets. Burned area algorithms based on temporal differences between preand post-fire radar backscatter values may also suffer from the temporal decorrelation effect that must also be accounted for. A recent study also revealed that undesirable effects caused by topography and varying environmental circumstances could be minimized with bi-temporal approaches [30].

Since prescribed burns primarily affect surface fuels and understory vegetation, their effect on radar backscattering is largely unknown and may vary from that of wildfires. The feasibility of analysis methods developed for wildfires has barely been evaluated when applied to controlled burns. Due to the introduction of thresholds for distinguishing different groups of fires categorized by their severity, their applications are further complicated. In addition, it is difficult to distinguish between high-intensity wildfire events and controlled burns, where the impact of fire ranges from low to moderate. Since SAR data provide a wealth of information on fire effects even for fires at low severity levels, C-band SAR data hold great potential for assessing the impacts of prescribed burns [22]. The objectives of this study include developing and applying a framework for forest fire analysis in areas impacted by controlled burns, and evaluating the effect of SAR sensor polarization and environmental conditions on analysis in comparison with wildfires.

#### **2. Study Area**

This study was conducted in Croajingolong National Park in southeastern Australia (36◦13'30" S to 39◦24'42" S, 149◦47'53" E to 150◦18'12") (Figure 1) [32]. One of only twelve world biosphere reserves in Australia, Croajingolong National Park in the state of Victoria is adjacent to Nadgee Nature Reserve in New South Wales. The study area has an oceanic Mediterranean climate with mean annual precipitation of about 1200 mm, which can be divided into a warm and dry season (October to April) and a cool and wet season (May to September) with >70% of precipitation [33]. The monthly average temperature can be as high as 27.2 ◦C, usually in January [34]. The above-ground biomass (AGB) ranges between 250 and 350 t/ha for mature Croajingolong National Park, with a further tall woody understory of 10–20 t/ha and woody debris of 50–70 t/ha. Southeastern Australia's forests in Croajingolong are distinguished by breeding plants (70% of the species, specifically the main species of tree) from post-fire stems [35]. With high-intensity fires, the trees may withstand defoliation, destruction of stem, and total top scorch, and some particular trees regenerate apically dominant crowns within five years. Within a month of burning, epicormics shoots emerge when fire-destroyed branches and upper stem parts weakened by fire eventually split and fall within the first two years. Fire-defoliated crowns of trees are typically destroyed, with shoots of epicormics restricted to the lower stem. With 35% to 45% of the total trees impacted by complete crown scorch, crown mortality also occurs. Depending on the species, crown replacement is between 30% and 95% within five years, and it is impossible to differentiate plants with substituted tops from those that have not experienced top death. In the lesser plant layers, up to 85% of classes can achieve flowering in the first two years after prescribed burns. Therefore, in the first year after the fire, certain improvements in forest composition can be predicted, as leaves emerge from the stem and main branches of both eucalyptus and shrub plants. Between 11 November and 30 December 2019, a scheduled burn was carried out, and approximately 88,500 hectares of tall open karri forest were burned in Croajingolong (Figure 1).

**Figure 1.** Geographic location of the study area: (**a**) the study area in Australia, (**b**) main forest distribution in Victoria, Australia, and (**c**) geographical location and main forest distribution in the study area.

The goal of the planned burn was to moderate the danger of wildfires by reducing fuel such as dead wood, leaf litter, bark, and shrubs, preventing extreme scorching or defoliation of tree crowns in the forest with young regrowth. The flow diagram in Figure 2 summarizes the approaches developed in this study.

**Figure 2.** General framework of this study.

#### **3. Materials and Methods**

#### *3.1. Radar Data*

The Sentinel-1 Single Look Complex (SLC) datasets were obtained from the Copernicus Open Access Hub and used in this study (Table 1). The dual polarization mode of the fine-beam (polarization of VV and VH) datasets (orbit: 29794, frame: 716, path: 147) were collected with a 36.4◦ incident angle at the middle of the swath. For burn severity estimation, two SAR images (6 October 2019 and 22 February 2020) obtained under different environmental conditions (dry and wet) were used. The post-fire tree growth was examined using a third SAR imagery dataset (18 October 2020), which was collected ten months after the fire (Table 1). Sentinel-1 datasets were obtained over the same season (australautumn) to reduce the possible variations caused by plant phenology and to ensure that environmental settings are as comparable as possible.

**Table 1.** The obtained SAR data and cumulative precipitation (mm) three days prior to image acquisition. Historical median values for temperature and precipitation over the study period are also provided.


a: Precipitation registered partially during the acquisition date. b: Precipitation reported the following day after image acquisition.

The SLC data were co-registered to the first image of the series (6 October 2019) using a lookup table generated from the images' orbital state vectors. An image-to-image cross-correlation algorithm was used to estimate the residual offsets present in the lookup table [36]. Least square and third-degree polynomial function regression methods were used to solve the errors in the lookup table, and offsets in the table were modeled. Multilook (eight looks in azimuth and four in range) techniques were applied to the co-registered images to achieve a spatial resolution of 30 m. The SAR intensity data was normalized using an incident field of a derived DEM to achieve the gamma naught backscatter coefficient [3]. Using the Range-Doppler approach, the SAR images were orthorectified to UTM 50S/WGS 84 projection with topography information from DEM and SAR data orbital details as input [37]. The preprocessing steps were mainly performed using ESA'S SNAP8.0.3 software.

After preprocessing, the SAR images were grouped into three stacks, one for each polarization (*VH*/*VV*), and one for normalized difference backscatter intensity (NDBI). The NDBI is a normalized ratio between *VV* and *VH* backscatter coefficients (Equation (1)). For each image stack, three images collected on different dates were processed.

$$NDBI = \frac{VV - VH}{VV + VH} \tag{1}$$

The radar burn ratio (RBR) basically works as the post- to pre-fire ratio of the backscattering coefficients (Equation (2)):

$$\text{RBR}\_{xy} = \frac{\gamma^0 Postfire\_{xy}}{\gamma^0 Prefire\_{xy}} \tag{2}$$

where *xy* describes a particular polarization based on *VV* or *VH* and *NDBI*, and the normalized backscatter coefficient based on a linear scale is γ<sup>0</sup> (gamma naught). In both dry and wet conditions, the indices of RBR were computed based on the two polarization (*VV*, *VH*) images and *NDBI*. The RBR was also computed using the mean of all pre- and post-fire images. Image histograms were analyzed, and the first (1) and last (99) percentiles were masked to eliminate outliers. In total, nine RBR images were created.

#### *3.2. Reference Data*

A total of 139 geo-located field photographs and an aerial orthophoto with 17 cm spatial resolution were used to create reference data (i.e., 730 points of different severity levels). The orthophoto was taken on 22 February 2020 (two months after the burn) with a Hasselblad H4D-60 camera and georeferenced and orthorectified by Parks and Wildlife Service, WA. Field images were taken between September and October 2020 (ten months after the fire) and corresponded to areas with varying levels of effects. Parks and Wildlife Service from the Department of Agriculture provided the aerial image and field photos. We further explain three burn severity levels and regrowth of vegetation in Figure 3.


**Figure 3.** Details of severity levels and regrowth of vegetation in the study area: (**a**) burned area with low to moderate severity, (**b**) burned area with high severity, (**c**) unburned area, and (**d**) regrowth of vegetation.

Within the prescribed burn boundary, a standard point grid with 300 m spacing was used to select the sampling points. While the unburned region within the burn area was small, to represent unburned trees, the points found outside the fire border were also sampled. There were 730 points sampled proportional to the area in each single class. The distribution of sample points was stratified by severity classes. The dataset was divided into two sets after each plot was allocated to a severity class. The first set was used to calibrate the model, while the second was used to validate it. The points for validation were selected randomly, leaving all the unselected points for calibration. The results were not influenced by the various calibration sample sizes for each class, since median value was used for model calibration for every single class.

#### *3.3. Ancillary Data*

The Meteorology Bureau, Government of Australia, provided regular precipitation weather data in the nearest four meteorological stations, i.e., Point Hicks (Lighthouse), Genoa (Fools Haven), Mallacoota, and Gabo Island lighthouse (http:/www.bom.gov.au/ climate/data/, accessed on 10 May 2021) (Table 1). In order to explain the environmental conditions (dry or wet) related to radar backscatter variability due to changes in soil moisture, the data for cumulative precipitation for the three days prior to the SAR acquisition date were obtained. Low spatial variability of soil moisture in the top few centimeters is related to high temperatures (high evapotranspiration potential) and low precipitation. The dates 6 October 2019 and 22 February 2020 were labeled as dry and wet, respectively. Since the cumulative precipitation was low and the precipitation was registered at least one day before the SAR acquisition date, the image used to assess post-fire tree survival (18 October 2020) was considered dry. The Shuttle Radar Topography Mission Digital Elevation Model (DEM) data with 30 m spatial resolution was used for the orthorectification of SAR data.

#### *3.4. Analysis and Modeling of Prescribed Burns*

RBR values were derived for all sample points (dry, wet, mean), taking into consideration each type of polarization/index (VV, VH, and NDBI) with different environmental conditions. To describe the relationship between prescribed burn effects and RBR under different environmental conditions and polarizations, the mean values with standard deviation values were analyzed. Two modeling approaches were combined for the extraction of prescribed burn effects using the RBR images. The first modeling approach was suggested in [29] (the STAND model below). An initial strategy was developed and tested to ease organizational execution (the NORM model). Note that linear scale operations of a mathematical approach were conducted on the SAR images, while a logarithmic scale (base 10) was used for data analysis (decibels, dB), which is a standard procedure in radar remote sensing.

#### 3.4.1. Standard Model (STAND)

There are three continents with seven different forest types included in the original RBR system, translating into the large variations of RBR ranges. By standardization of mathematical data (having z-scores), which rescales the values of RBR to a particular range, compensation for different RBR ranges was achieved. It is important to note that RBR standardization is not exclusively appropriate for the small area examined in this study. The original RBR system, however, was retained for reference purposes. Furthermore, using standard values (z-scores, distance among the significance of the sample values and the mean of the population) is helpful to interpret the RBR values over various environmental conditions. Standardization was carried out using Equation (3):

$$\mathbf{s}^{RRR} = \frac{(\mathbf{x} - \boldsymbol{\mu})}{\sigma} \tag{3}$$

where the standardized RBR value is s*RBR*, *x* represents the pixel value, and *σ* and *μ* are the mean and standard deviation values of RBR applied to adjacent non-fire-affected areas of forest.

For standardization, the unburned pixels were used to show the difference compared to the original process, as the burned region typically displayed gradient ranges from moderate to low, which then leads to the high variability of the backscatter. Every pixel value shows the difference in backscatter present in the pixel and backscatter mean of unburned areas with the use of unburned area for standardization, allowing for better data analysis (the difference with unburned conditions). To measure the μ and σ values, a total of 3524 pixels in 5 polygons across the prescribed burn were used to account for spatial variability in the forest. Once the images with uniform characteristics (s*RBR*) were collected, for each single image, the mean value of each class related to severity was determined using the backscatter values derived from the calibration dataset coordinates. Thresholds dividing various severity classes were calculated as the median value of the mean (s*RBR*) of neighboring severity classes [30]. The median was perceived to be a stable central tendency test (where outliers are less sensitive) in the case of SAR image analyses marked with speckle noise.

#### 3.4.2. Normalized Model (NORM)

This model uses the normalized values (*nRBR*, Equation (4)) of backscattering coefficients. The benefit of using *nRBR* lies in its simplified calculation and interpretation, as the index does not require standardization:

$$m^{RBR} = \frac{\chi^0 Postfire\_{xy} - \chi^0 Prefire\_{xy}}{\chi^0 Postfire\_{xy} + \chi^0 Prefire\_{xy}} \tag{4}$$

where *xy* is a definite polarization (i.e., VV or VH) or NDBI, and γ<sup>0</sup> (gamma naught) is the linear-scale normalized backscattering coefficient. As in the previous model, the median values were used for computing thresholds between groups.

#### 3.4.3. Model Validation

Using the thresholds obtained for each model, images were classified using s*RBR* or *nRBR*, including unburned, low to medium, and high severity. A total of 18 maps were produced (Appendix A, Figures A1–A3). In order to minimize the influence of isolated pixels, a major filter (3 × 3) was employed on all maps.

Confusion matrices were constructed using the validity points by removing the modeled fire grades. In order to evaluate the agreement between estimated fire impact and reference, overall accuracy (*OA*), Cohen's kappa (kc), omission error (*OE*), and commission error (*CE*) were calculated using the confusion matrices (Equations (5)–(8)) [38].

$$OA = \frac{TP + TN}{TP + FP + FN + TN} \tag{5}$$

$$kappa = \frac{OA - p\_\varepsilon}{1 - p\_\varepsilon} \tag{6}$$

$$CE = \frac{FP}{TP + FP} \tag{7}$$

$$OE = \frac{FN}{TP + FN} \tag{8}$$

where *TP* = true positive, *TN* = true negative, *FP* = false positive, *FN* = false negative, and *pe*<sup>=</sup> (*TP*+*FN*)·(*TP*+*FP*)+(*FP*+*TN*)·(*FN*+*TN*) *TP*+*FP*+*FN*+*TN* . *OA* is the total number of accurately predicted values divided by the total number of predictions. The omission error of one class (OE) is the proportion of real samples of that class that were misclassified (they were class A in the ground truth, but we classified them as B, C, or D). The commission error of one class (CE) is the proportion of all classified points in that class that were misclassified (they were classified as class A, but they belong to class B, C, or D in ground truth). The average values of CE and OE in each class were calculated by adding up all false positives, false negatives, true positives, and true negatives. The perception of kappa values relies on various considerations, with no agreement (kc < 0), mild agreement (0 to 0.20), good agreement (0.21 to 0.40), modest agreement (0.41 to 0.60), significant agreement (0.61 to 0.80), and nearly perfect agreement (0.81 to 1) [39].

#### **4. Results**

In Figure 4a, *RBRVH* decreased as the intensity of the fire increased under dry conditions. For the cross-polarized channel between the intensity classes of low to moderate and unburned, an increasing factor in soil moisture due to wet conditions shows a fairly complex relationship with the detection of minor differences. In dry conditions, *RBRVV* was less effective in distinguishing between fire severity classes but showed the potential under wet conditions to distinguish unburned from burned areas, as explained in Figure 4b. Wet conditions produced the highest dynamic range between unburned and high severity (2 dB) for *RBRVV*, which may be due to increased backscatter from the field. Under both environmental conditions, the *RBRNDBI* showed an increase with fire severity, as explained in Figure 4c.

In the process of retrieving the effect based on prescribed burns, two modeling methods produced comparable results. The most reliable predictions were obtained in dry conditions, *RBRVH*, using the STAND model (OA = 82.3%; kc = 0.78). The results are illustrated in Figure 5.

Figure 6 and Table 2 illustrate validation results for both approaches (STAND and NORM). The gap between the two models was relatively small. Omission and commission errors for the severity class of low to moderate were the largest (30% and 35%, respectively). Errors did not exceed 20% for unburned and high-severity classes. For the cross-polarized pathway (*RBRVH*), elevated humidity contributed to a significant loss of sensitivity to intensity levels (Table 2). In contrast to the analysis in dry conditions for the cross-polarization channel, *RBRVV* and *RBRNDBI* displayed lower OA estimation accuracies (70–80%).

**Figure 4.** The relationship between RBR and intensity levels under various environmental conditions for VH polarization (**a**), VV polarization (**b**), and NDBI (**c**). Median values (horizontal line), percentiles (25% and 75%, box edges), non-outlier values range (whiskers), and outlier values (circles) are seen in box plots.

**Figure 5.** Estimated severity map using VH polarization data in dry conditions and the STAND model.

**Figure 6.** Omission and commission errors (average through classes): (**a**) dry conditions, (**b**) wet conditions, and (**c**) mean conditions.


**Table 2.** Validation results for severity estimates under different environmental conditions obtained for the VH polarization.

The NORM model outperformed the STAND model for wet images and the mean between dry and wet images (Table 2). Although VH polarization provided the most reliable fire severity forecasts (OA = 82.3%; kc = 0.78) in dry conditions, VV was of particular importance in wet conditions. *RBRVV* obtained more accurate results than VH and NDBI under wet conditions, with the NORM model showing the least errors (OA = 76.6%; kc =0.69). The NORM model and *RBRVH* generated the most reliable results (OA = 74.4%; kc = 0.62) for the mean between dry and wet images. It was also possible to differentiate between unburned and burned areas in those circumstances (Appendix A, Figures A1–A3).

#### **5. Discussion**

Previous experiments have demonstrated that the impacts of fire can be measured by integrating pre- and post-fire datasets. This study focused on the areas impacted by the incidents of wildfires that are normally marked as high. Our study improves previous knowledge on the accuracy of RBR over prescribed burns to minimize ground fuels with limited effects on the upper canopy. Since C-band SAR data was highly sensitive to scatter from the top part of the canopy [40], the response of RBR to the lower effect levels in areas controlled by prescribed burns should be assessed.

This work proved, even in low to moderate severity, that the cross-polarized channel and dry conditions are the most suitable combination for a fire effect evaluation. The fire sensitivity was found to be poor in the co-polarized channel, specifically in dry conditions, and increased under wet conditions, which is consistent with previous studies using X-, C-, and L-band data over different types of woods [26,28,41]. Since the ground surface shows more susceptibility with increased backscattering, the VV polarization method and NDBI provide the ability to detect burned areas under wet conditions. A simplified modeling technique was applied and yielded comparable accuracies.

The RBR structure and the thresholds suggested in [30] were used to extract severity maps for the study area. The benefit of using a standard calibration for prescribed burns was demonstrated by maps validated using reference data in this study. The calibrated models greatly improved average efficiency from 60% to 82% and the Kappa coefficient ranged from 0.21 to 0.78. Similarly, the overall accuracy exceeded the minimal value required to map the severity of prescribed burns (82.3%) [42]. However, further studies are required to assess the accuracy attained in other forest types and ecosystems. The results obtained in WA are marginally higher compared to approaches based on a high-resolution optical dataset (OA value range of 78–83%) [15]. The application of SAR polarimetric and phase information and deep learning methods achieved high accuracies for burnt area mapping from Sentinel-1 SAR data [29,31]. Although the accuracies of the proposed method were relatively lower, it might be more suitable for operational forest fire monitoring with the low demand for computational resources. Moreover, the proposed modeling strategy extracted different severity levels of prescribed burns.

Among the limitations of this study, it is important to point out that five levels of fire severity were considered in previous studies [15], although only three classes were included here. For prescribed burns, fewer groups were considered to distinguish between the unburned (failed burns), burning areas at low to moderate severity levels, and burned areas at a high severity level, where crown tree damage or decreasing trends in stems may decrease productivity of the forest. Using fewer groups may affect the outcomes, as simulation techniques could be more stable, and hence the accuracy will increase. As compared to field investigations, another weakness was due to the use of reference data obtained from satellite image interpretation. Such a strategy can produce noise and lead to the reduction in potential for discrimination. In addition, variations in environmental conditions before and after the burn cause uncertainties in remote sensing detection results. The variations in soil moisture in multi-temporal images may inhibit the evaluation of the impact of fire using SAR datasets [43]. By combining multiple SAR images, the impact of environmental variations can be minimized each time before and after a fire event [9,44].

#### **6. Conclusions**

This study expanded our knowledge of SAR data for impact assessment of wildfires through an innovative example, prescribed burns. The outcome of prescribed burning was analyzed by estimating fire effects (severity) on forest flora in Victoria, Australia. The freely available Sentinel-1 C-band SAR data, with a 12-day temporal coverage anda5m spatial resolution, was applied because of its sensitivity to structural changes in vegetation after forest fires. A fire intensity assessment was carried out for various polarization and environmental conditions using multi-temporary SAR indices. A modeling approach for prescribed burns that can be implemented operationally was developed. The validation results showed that the VH polarization data under dry conditions provided reliable fire severity forecasts with an overall accuracy of 82.3% (kc = 0.78). Although VH polarization provided the most reliable estimation results in dry and mean between dry and wet images, VV polarization data obtained optimal results in wet conditions. Based on our analysis, the NORM models outperformed the STAND models in wet and mean between dry and wet images. The modeling strategy mostly retained the accuracy of the initial RBR method while reducing computational complexity. This study provided a novel technique to assess the effectiveness of controlled burns for forest management and decision-makers in fireprone areas. The analysis framework could be useful for forest conservation and fire control decision-makers, not only in southeastern Australia, but also in other areas where burning is used as a land management method. Future research will involve the deployment of multiple SAR wavelength channels and the integration of Sentinel-1 data with optical sensor data to further improve fire severity estimates.

**Author Contributions:** A.T.: Conceptualization, Writing—review and editing, Methodology, Software, Formal analysis, Visualization, Data curation, Writing—original draft, Investigation; H.S.: Supervision, Conceptualization, Writing—review and editing, Methodology, Software, Formal analysis, Visualization, Data curation, Writing—original draft, Investigation; Q.L.: Supervision, Conceptualization, Writing—review and editing, Methodology, Software, Formal analysis, Visualization, Data curation, Writing—original draft, Investigation, Funding; O.A.: Visualization, Validation, Investigation, Writing—review and editing; M.R.K.: Writing—review and editing, Methodology, Formal analysis, Visualization, Investigation; M.F.B.: Visualization, Validation, Investigation, Writing review and editing; L.L.: Supervision, Writing—review and editing, Methodology, Software, Formal analysis, Visualization, Data curation, Writing—original draft, Investigation, Funding. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work is supported by the Strategic Priority Research Program of the Chinese Academy of Sciences (Grant No. XDA23100304), the National Natural Science Foundation of China (Grant No. 41871345), and the National Natural Science Foundation of China (Grant No. 42071321).

**Data Availability Statement:** The data presented in this study are available upon request from the corresponding author. The data are not publicly available due to the thesis that is being prepared from these data.

**Acknowledgments:** The authors wish to thank the editors and anonymous reviewers for their valuable comments and helpful suggestions. The authors are grateful to GIS Department (Corporate Services | Department of Environment, Land, Water and Planning) for providing data and metadata related to the prescribed burn, from which the reference data used in this work were derived.

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

**Appendix A**

**Figure A1.** Dry condition.

**Figure A2.** Mean of dry and wet condition.

**Figure A3.** Wet condition.

#### **References**


MDPI St. Alban-Anlage 66 4052 Basel Switzerland Tel. +41 61 683 77 34 Fax +41 61 302 89 18 www.mdpi.com

*Remote Sensing* Editorial Office E-mail: remotesensing@mdpi.com www.mdpi.com/journal/remotesensing

MDPI St. Alban-Anlage 66 4052 Basel Switzerland Tel: +41 61 683 77 34

www.mdpi.com

ISBN 978-3-0365-5668-0