Despite store than that closer to the surface, thus

Despite covering only approximately 3% of the Earth’s terrestrial
surface, peatlands contain between 32 and 46% of the global
soil carbon (C) pool (Page and Hooijer, 2016) equating to 500-700 Gt of C
(Scharlemann et al., 2014) and have been widely
regarded as a globally important C pool for some
time (Rieley et al., 1993; Maltby, 1986). The importance of peatland stability
was noted due to climate change implications,
as the C pool is so large (Page et al., 2002), and persistent anthropogenic
influences in peatlands such as deforestation and drainage has led to soils which are particularly vulnerable to fires
(Watson et al., 2000). ‘Smouldering combustion’ peat
fires are ignited very easily, and can occur in wet conditions however it is the
lowering of the water table and subsequent
drying of soils which has increased the frequency and magnitude of burning (Turetsky et al., 2014).

 

There is an increasing concern about the impacts that wildfires
have upon carbon dioxide (CO2) emissions and other particulates into
the atmosphere (Tansey et al., 2008). In Indonesia, the peat carbon pool
accounts for nearly 75% of the country’s total forest carbon pool, but is
predicted to continue to decrease in size due to climatic warming, land-use
change and fire (Page and Hooijer, 2016). Page et al., (2002) estimated that in 1997 peat fires in Central Kalimantan
released between 0.81 and 2.57 Gt of carbon, equivalent
to 13-40% of mean annual global carbon emissions from fossil fuels, and an
approximate carbon loss of 29 kg per square metre. Fires which
reach greater depths into peat soils can reach soil carbon which was a much more permanent carbon store than that
closer to the surface, thus have a great impact on the carbon cycle and climate
feedbacks (Turetsky et al., 2014).

We Will Write a Custom Essay Specifically
For You For Only $13.90/page!


order now

 

Optical remote sensing
of burnt areas has been used since the early 1970’s (Wilson et al., 1971), with numerous iterations of satellites due
to their limited operating periods and advancements in satellite technology. Remote
sensing can be used in burnt area mapping due to the difference in
near-infrared (NIR) reflectance of vegetation when compared to other surfaces,
caused by the interaction of chlorophyll with the electromagnetic spectrum (a
and Klemas, 2011). In the NIR spectral range, healthy vegetation displays a
high reflectance compared to burned surfaces, which display lower reflectance
than any other land type except water (Kohl et al., 2006). This technique was
first implemented in 1970, where the NIR spectral range was used to detect
fires in Montana, USA (Wilson et al., 1971), which resulted in the delineation
of 200 wildfires, with numerous fires detected before they were from lookout towers
on the ground.

 

Remote sensing allows for studies over longer time series, and
over larger study areas. For example, Landsat records date back to 1984 with
16-day temporal resolution globally (Parks et al., 2014). Mapping of burn scars can
reveal interesting conclusions which may not have been visible otherwise.

Verhegghen et al (2016) mapped burn scars in the Congo Basin, and found that
the origin of burn scars was near to accessible points for humans, suggesting
anthropogenic causation for much of the 2016 fires. Using remote sensing, it is
possible to extrapolate ground measurements to predict carbon flux in Indonesia
(Page et al., 2002; Konecny et al., 2016). There is still a plethora of
unanswered questions regarding fire management practices, and remote sensing
has been used to support new techniques and methods to mitigate against
continued fires (Atwood et al., 2016).

 

Traditionally, remote sensing has been used to monitor and map fires at
relatively low resolution. Spatial resolutions ranged from 300 m to 1 km using
sensors such as Advanced Very High Resolution Radiometer (AVHRR) and Moderate
Resolution Imaging Spectroradiometer (MODIS) (Mouillot et al., 2014), which
often proved too coarse to accurately represent burnt area extent, with high
variations in results between each burnt area product (Verhegghen et al.,
2016). The introduction of the Sentinel-2 satellite allows rapid and accurate
assessment of fires at high temporal and spatial resolutions (ESA, 2010).

 

Sentinel-2 is tailored for natural hazard management such as fires
(ESA, 2010) and features both NIR (band 8) and SWIR (band 12) bands at high
spatial and temporal resolutions at band 8 and band 12 respectively. The
spatial resolution is much higher than previous sensors, with a number of bands
such as red, green, blue and NIR having a 10m resolution. However, NBR requires
SWIR-II, S2 band 12, which has a spatial resolution of 20 m (ESA, 2016).

Previous sensors do not provide enough knowledge about which vegetation formation
is burning due to a coarse spatial resolution, and in their Congo Basin study,
Verhegghen et al (2016, p.19) found S2 data to have ‘great potential for
tropical forest monitoring applications, and in Europe the GSE RISK-EOS service
network as part of the S2 mission, will provide 10-15m resolution burn scar
products (ESA, 2010)

 

One issue with optical remote sensing of burnt area in tropical
peatlands is the availability of cloud-free scenes due to the regional climate.

Many of the images contain high levels of cloud cover thus limited useable
imagery exists. For this reason, numerous studies have been conducted using
Synthetic Aperture Radar (SAR) (Kasischke et al., 1994; Dahdal, 2011) and LiDAR
(Englhart et al., 2013) in recent times. A further challenge arises regarding
the identification of burns and burn severities. Some land types can show
behaviour similar to burnt areas (Kohl et al., 2006) such as urban areas, thus
a multitemporal approach is suggested alongside the removal of urban areas and
any surface water.

 

There are numerous methods to detect burnt area using satellite
sensors. Optical remote sensing focuses on vegetation indicators such as the
Normalized Difference Vegetation Index (NDVI) described by Tucker (1979), which
is calculated using the near infra-red (NIR) and the red band as shown in
equation 1.

                                                         ( 1 )

Recently, numerous
studies have argued for an improved index. The Normalized Burn Ratio (NBR)
proposed by Key and Benson (2002) is very similar to NDVI, however the red band
is replaced with the short-wave infrared (SWIR) band. Key and Benson (2002)
found a direct correlation to field estimates of fire effects to the NBR
results in their study. Using the SWIR band instead of a red
band enhances the spectral response of vegetation affected by fire (Lopez
Garcia and Caselles, 1991; Eidenshink et al., 2007). As noted by Jones and
Vaughan (2010, p. 305), NBR “gives a good correlation with damage” in
comparison to NDVI, hence is preferable in burned area mapping. There is a
focus on change detection using NBR, which is often implemented to show burn
extent and severity. This differenced index, differenced NBR (dNBR), is a
transferable measure with potential to compare burns on a regional scale.

 

Differenced NBR
requires two scenes, which presents a few issues. Scene pairs should represent
similar phenology and should not show any land cover changes (Quintano et al.,
2018). In peatland regions, there are often areas of seasonal water (Rieley et
al., 1996; Borneo Nature Foundation, 2017) hence scene pairs should ideally be
chosen in the same season. This issue is two-fold, as the limited availability
of cloud-free imagery makes scene pairs of the same season even harder to
acquire.

 

Recently, there have
been a number of studies to suggest S2 and L8 can be combined to provide rapid
assessment of fires (Mandanici and Bitelli, 2016;
Korhonen et al., 2017; Van der Werff and Van der Meer, 2016; Novelli et al.,
2017). As concluded by Quintano et al., 2018, the combination of Landsat
and Sentinel-2 data may increase the speed at which an initial burnt extent and
severity map can be released after a wildfire. There is therefore a need to
test the response of Sentinel-2 and Landsat 8, as systematic differences in NBR
between each sensor will affect severity maps and emissions predictions.

 

To assess suitability, comparisons can be made between S2 and L8. Although
there has been very little study on the comparison of NBR between S2 and L8, a correlation
(Pearson’s coefficient) ranging between 0.93 and 0.99 for NDVI between the two
sensors has been reported, with the range in value depending on location
(Mandanici and Bitelli 2016). There are, however, some issues have been found
between NIR bands which need addressing (Mandanicini and Bitelli, 2016).

 

There is also some confusion over the choice of NIR band for
comparisons between S2 and LS, as the choice of band 8 or 8a can differ depending
on the Landsat satellite being used for comparison (Mandanici and Bitelli, 2016).

This issue has not received much attention, and has not been considered for NBR
comparison. However, there is found to be very little influence of the choice
of NIR band in NDVI calculations as shown in Figure 2.

Figure 2. Comparison of band 8 and 8a in the calculation of NDVI. The red
line signifies a linear regression. Source: Mandanici and Bitelli (2016).

Although the effects of NIR band choice are unlikely to affect the results
greatly (Mandanici and Bitelli, 2016) preliminary tests of band 8 and 8a in NBR
calculations in Central Kalimantan found that band 8 produced NBR results most
similar to Landsat 8. A linear regression of the relative response function
between S2 band 8 and the respective L8 band (5) reported results which vary
region to region, and it was concluded that the response of each NIR band
requires further study (Mandanici and Bitelli, 2016). However, SWIR is found to
match very strongly between sensors (Mandanici and Bitelli, 2016), and assuming
a consistent SWIR between the two sensors, a systematically lower NIR will
result in a lower NBR value, showing a more severe burn.

Due to the uncertainty of remotely sensed data an accuracy assessment
is usually required, preferably with comparisons to ground reference data. If
an accuracy assessment takes the form of classifications of land cover, in this
case burnt and unburnt area, error matrices are often used (Jones and Vaughan,
2010).