Spectral analysis of fire severity in north Australian tropical savannas

Remote Sensing of Environment 136 (2013) 56–65

Contents lists available at SciVerse ScienceDirect

Remote Sensing of Environment journal homepage: www.elsevier.com/locate/rse

Spectral analysis of fire severity in north Australian tropical savannas Andrew C. Edwards a, b, c,⁎, Stefan W. Maier b, Lindsay B. Hutley b, Richard J. Williams d, Jeremy Russell-Smith a, b a

Bushfires NT, PO Box 37346, Winnellie NT 0821, Australia Research Institute for the Environment and Livelihoods, Charles Darwin University, Darwin NT 0909, Australia Bushfires Cooperative Research Centre, 340 Albert St, East Melbourne Vic. 3002, Australia d CSIRO Sustainable Ecosystems, PMB 44, Winnellie NT 0822, Australia b c

a r t i c l e

i n f o

Article history: Received 25 October 2012 Received in revised form 19 April 2013 Accepted 20 April 2013 Available online 22 May 2013 Keywords: MODIS NBR/CBI

a b s t r a c t This paper reports on the application of fire severity studies describing the immediate post-fire spectral responses of fire affected vegetation and substrates, to remotely sensed mapping of fire affected tropical savanna vegetation in northern Australia. Hyperspectral data were collected from a helicopter coincident with accurately located sites where detailed ground sampling was undertaken based on adaptation of standard methods such as the GeoCBI. Ground sampling revealed the importance of models that characterise both photosynthetic and non-photosynthetic vegetation including scorched foliage. The proportion of charred material was not significantly correlated with fire severity categories. Models were assessed particularly in relation to spectral bands of the MODIS sensor given its high observation frequency and global application in fire detection and mapping studies. Significant and inverse fire severity relationships were observed with the near infrared and two short wave infrared bands, demonstrating support for a model like the widely used differenced normalised burn ratio (ΔNBR). However, model assessment using Akaike's Information Criteria suggests the most parsimonious model is the pre- and post-fire difference in MODIS channel 6 (1628–1652 nm). The resultant models have direct application in fire severity mapping products for fire-prone tropical savanna vegetation in northern Australia. © 2013 Elsevier Inc. All rights reserved.

1. Introduction In Australia, between 300,000 km 2 and 700,000 km 2 is affected by fire annually, with the great majority occurring in the northern savannas (Maier & Russell-Smith, 2012, chap. 4; Russell-Smith & Yates, 2007). Savanna fires are recognised as a significant global source of greenhouse gas emissions, with the contribution of biomass burning from Australian savannas ranked either second or third after Africa (Cook et al., 2010; van der Werf et al., 2010). In Australia, nearly all savanna fires are attributable to anthropogenic sources (Russell-Smith & Yates, 2007). In recent decades there has been a shift across substantial parts of northern Australia towards a late dry season-dominated fire regime characterised by high severity events, low levels of landscape patchiness, high total fuel consumption and consequently increased greenhouse gas (GHG) emissions (Russell-Smith et al., 2009; Yates et al., 2008). Given the extent of Australia's flammable tropical savannas (~ 2 million km 2), low population density (b0.01 person.km −2) and associated lack of infrastructure, remote sensing of fire extent and impact has become an essential land management tool (Dyer et al., ⁎ Corresponding author at: Bushfires NT, PO Box 37346, Winnellie NT 0821, Australia. Tel.: +61 427270835. E-mail address: [email protected] (A.C. Edwards). 0034-4257/$ – see front matter © 2013 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.rse.2013.04.013

2001). Sophisticated automated methods for mapping burnt (i.e. fire affected) areas globally have been developed accounting for the anisotropy of land surfaces, angular sensing and illumination variations through bi-directional reflectance model-based detection (Maier, 2005, 2010; Roy et al., 2002, 2005). Models applicable to MODIS sensor data are currently used for mapping fire affected areas in the north Australian region; these provide an appropriate landscape mask within which to classify fire severity. The high frequency of MODIS acquisitions (approximately two observations per day) is adequate for difference imaging, especially in the context of accounting for marked and rapid phenological changes in tropical savanna habitats post-fire, such as re-flushing and leaf fall (Pereira, 2003). In addition, the spatial and spectral resolution of MODIS imagery is currently superior to other sensor data that meet these first two conditions. While burnt area mapping products have been widely used in a variety of savanna conservation and management applications (Dubinin et al., 2010; French et al., 2008; Fule et al., 2003; Hudak & Brockett, 2004; Karau & Keane, 2010; Keane et al., 2001; Mitri & Gitas, 2002; Röder et al., 2008; Rogan & Franklin, 2001; RoldanZamarron et al., 2006; Sa et al., 2003; Stroppiana et al., 2002; Ustin et al., 2004), to date tools for remotely sensed mapping of fire severity have not been available. Rather, fire severity has been inferred from coarse assessments of fire seasonality where, as demonstrated by Russell-Smith and Edwards (2006) in Australian mesic savannas,

A.C. Edwards et al. / Remote Sensing of Environment 136 (2013) 56–65

fires early in the annual ~ 7-month burning season (pre-August) are typically of low intensity/severity, and fires later in the year (August–November) are generally of higher intensity/severity. Fire severity mapping is an essential requirement for enhanced biodiversity and greenhouse gas emissions monitoring and assessment, both in fire-prone savanna landscapes across northern Australia (Meyer et al., 2008; Russell-Smith et al., 2006, 2009, 2012; Strand et al., 2007; Williams et al., 2002; Woinarski et al., 2010) and internationally (Keeley, 2009; Strand et al., 2007). If acquired in a timely fashion, it also provides land managers with intra-seasonal information regarding the effects of their imposed fire management. 1.1. Describing fire severity Whereas fire intensity is a physical measure of the energy released from biomass burning, fire severity describes the direct impact of the fire, and its post-fire effects on combustible material (Keeley, 2009; Lentile et al., 2006). The categorical classification of fire severity will be different in different contexts, a continuous index provides objectivity. In this paper, we will initially apply a categorical classification to the 9 fire events using a severity index as defined by RussellSmith & Edwards (2006). We will then use ground measurements to create a more powerful continuous index that can be applied to remote sensing analyses and categorised for various applications. Fire severity can be quantified within the habitat structure both horizontally and vertically (Russell-Smith & Edwards, 2006). The horizontal component reflects fire patchiness, the internal heterogeneity of burnt areas, while the standing component reflects the degree of charring and scorching of photosynthetic and non-photosynthetic plant material in different strata. Rapid assessments describing these components can provide large and useful datasets for the calibration and validation of remotely sensed data. Studies attempting to rapidly characterise fire severity are generally habitat specific and rely on a single measure for assessment including: minimum branch diameter (Pérez & Moreno, 1998; Whight & Bradstock, 1999; Williams et al., 2006); tree mortality (Chappell & Agee, 1996); crown damage (Russell-Smith & Edwards, 2006); scorch and char height (Williams et al., 1998); soil burn depth (Chafer, 2008; Schimmel & Granstrom, 1996); fuel moisture conditions (Ferguson et al., 2002); and white ash production (Smith & Hudak, 2005; Smith et al., 2005). For north Australian fire management applications fire severity can be defined with respect to a small number of readily observable classes using ground patchiness and leaf scorch proportional height (refer Figures in Edwards, 2009; Russell-Smith & Edwards, 2006). The Composite Burn Index (CBI) of Key and Benson (2006), re-interpreted by De Santis and Chuvieco (2009) as the GeoCBI, are indices used to calibrate and validate the results of a remote sensing index, the change in the Normalised Burn Ratio (ΔNBR), which is applied specifically to multi-temporal Landsat imagery. The United States Forest Service (USFS) has developed this system into a national fire severity mapping program. However, each Landsat image is calibrated individually, limiting its application for real-time fire management. A number of other studies have attempted to classify fire severity from a single image or through change detection techniques, often using ΔNBR, with paired satellite images or aerial photography of a single fire event (Chafer, 2008; Chafer et al., 2004; Chuvieco and Congalton, 1988; Hammill and Bradstock, 2006; Noonan et al., 2002; Smith et al., 2005; Roldan-Zamarron et al., 2006; White et al., 1996). As in the CBI/ΔNBR system, site data are applied to calibrate the fire severity classifications. The result is evaluated for accuracy with either a new independent set of data or a random subset. The result is a one-off categorisation and the whole process must be repeated for each image. Full automation is not possible. The above coupled approaches indicate clear differentiation between fires affecting ground layer vegetation versus the canopy, but they often report inadequacies with respect to describing categories,

57

or the continuum, of effects in the understorey. The simpler classification into severe (upper canopy affected) and non-severe fires has greatest application in an iterative, intra-seasonal assessment of prescribed burning activities, and the greatest potential accuracy (Holden et al., 2009). However, for the purposes of conservation management planning and the calculation of greenhouse gas emissions, greater discrimination of fire effects on all strata is required. Fire scars persist in tropical savannas for weeks or months at most (Bowman et al., 2003; Pereira, 2003). However, the persistence of fire severity indicators is unknown. Accordingly, acquisition frequency and immediacy become critical features of any automated burnt area or fire severity mapping system. The spectral and spatial resolution of MODIS data are appropriate to mapping fire and its effects in north Australian tropical savannas (Maier, 2010; Maier & RussellSmith, 2012). MODIS imagery acquisition characteristics resolve persistence issues in tropical savanna habitats given near-twice daily image availability. The highly accurate geo-location of pixels and wide swath make processing more efficient, particularly in comparison to other moderate resolution sensors (Justice et al., 2002). This study reports on the spectral characterisation of fire severity in mesic (> 1000 mm rainfall p.a.) north Australian savanna systems. In particular we, (1) describe methods for detailed field sampling and the rapid assessment of fire severity, (2) define categorical and continuous indices of fire severity, (3) characterise the ground variables, (4) quantify their change as a function of fire severity, (5) test the most useful spectral variables describing fire severity, and (6) statistically test a suite of reflectance algorithms to map fire severity from satellite based remotely sensed data. 2. Methods 2.1. Study area Australia's tropical savanna region extends from Broome in Western Australia, across the “Top End” of the Northern Territory through the areas surrounding the Gulf of Carpentaria to Cape York in Queensland to the east (Fig. 1). Over that broad region, rainfall is highly seasonal with most rain falling from October to April. During the fire season (May to September), the understorey progressively cures with increasing biomass flammability. Mid and upper storeys may provide between 50 and 80% of the available fine fuel through leaf fall (Cook, 2003), with the remainder comprising grass and other herbaceous fuels. Planned burning takes place in the early months (April to June/July), referred to here as the Early Dry Season (EDS), where fuel and soil moisture are still sufficiently high to restrict fire spread, and fire intensity and severity is generally low. Mid Dry Season (MDS) fires, July/August, are generally of higher intensity but may still be controlled (Russell-Smith et al., 2007). By the Late Dry Season (LDS), September–November, typically there has been many months without rain, relative humidity is low, and fuels are completely cured. At this time, wildfires are prevalent, mostly from anthropogenic but also lightning sources, and fire suppression is difficult. 2.2. Study locations Site data were collected in three areas (Table 1, Fig. 1). The CSIRO site in Darwin contained two sampling areas unburnt for 8 and 25 years, respectively. The sampling area at the Howard Springs eddy covariance study site (Fig. 1; Beringer et al., 2007) consisted of two blocks, one with regular, but low severity fires, and a neighbouring plot that was also burnt regularly, but with a regime dominated by moderate to high fire severities (Beringer et al., 2003). At both Howard Springs sites the understorey is dominated by tall annual and perennial grasses (O'Grady et al., 2000), with fire occurring every second year. The third sampling area was near the community of Kabulwarnamyo, on the Arnhem Plateau (Fig. 1). The area is intensively managed as part of a

58

A.C. Edwards et al. / Remote Sensing of Environment 136 (2013) 56–65

Fig. 1. Location maps: (a) the location of the study area in northern Australia; (b) the experimental area at the CSIRO complex is in Darwin, the Howard Springs experimental sites ~30 km east of Darwin, the community of Kabulwarnamyo in the West Arnhem Land region ~300 km east of Darwin.

traditional indigenous estate, with a fire regime of mixed severity (Yibarbuk et al., 2001), generally occurring two out of three years (Edwards & Russell-Smith, 2009). The understorey is dominated by perennial and woody species, with little annual grass (Lynch & Wilson, 1998). 2.3. Data collection Initial post-fire canopy photographs (from October 2006) illustrated significant re-flushing and extensive leaf fall (Fig. 2b). These rapid and marked phenological changes, also present in other savanna studies (Trigg & Flasse, 2000), indicated that consistent and timely post-fire data collection was required. Reflectance data and coincident ground data were collected subsequently. For the sake of consistency, collection of spectra was restricted to the day after the fire to coincide with the temporal difference in MODIS image acquisition. Ground site evaluation was restricted to the second day post-fire, so that sampling was prior to any influential phenological change,

but after fire effects such as leaf scorch were obvious and clearly measureable. 2.4. Reflectance data Site-specific hyperspectral reflectance data in the optical range of the electromagnetic spectrum were collected post fire using a hand-held spectrometer from a helicopter, reducing the problem of accounting for atmospheric effects and co-location with ground sampling. A hand-held ASD Field Spec Pro was used to record spectra in the range of 350–2500 nm with a 25° field of view at approximately 91 m above ground, resulting in a ground sampling diameter of approximately 40 m (Taylor, 2004). Incident radiation was measured by pointing the spectrometer at a white reference (WR) panel fabricated from spectralon (Jackson et al., 1992). Site selection was randomly stratified, reflectance spectra were averaged 25 times for a total of 10 samples per site, tagged with GPS location information, and saved as a unique file.

Table 1 Location and structural descriptions of sampled sites. Basal area was calculated from diameter of stems >5 cm measured at 1.3 m above stem base. Foliage projective cover (FPC) is a measure of the proportion of orthogonal foliage shadow on the ground. No.

Site-fire history location

Characteristics (area)

Basal area (m2 ha−1)

No. stems (ha−1)

FPC (%)

1 2 3 4

CSIRO site — 8 years unburnt 12°24′35″S 130°55′12″E CSIRO site — 25 years unburnt 12°24′35″S 130°55′12″E Howard Springs Burn Plot — moderate/high fires 2006,07,08 12°29′26″S 131° 8′26″E Howard Springs Control Plot – Low/moderate fire 2006; – Mostly excluded 2007; – Low severity 2008; – Moderate severity 2009. 12°29′25″S 131°9′18″E Kabulwarnamyo, west Arnhem Land — mixed low/moderate severity 12°45′53″S 133°50′41″E

Open Open Open Open

10.7 8.9 11.8 10.0

1336 960 500 320

53 64 24 19

12.8

296

47

5

forest savanna (8.7 ha) forest savanna (2 ha) forest/woodland savanna (300 ha) forest/woodland savanna (270 ha)

Open forest savanna (~1000 ha)

A.C. Edwards et al. / Remote Sensing of Environment 136 (2013) 56–65

59

Fig. 2. (a) Photograph of a typical ground transect, where scorch height can be compared against the technician (approximately 1.6 m high), (b) photograph of a ground site sampled one week after a fire event, in November 2006. The fire was of high severity and illustrates strong evidence of re-flushing by cycads (Cycas sp.), and abundant scorched leaf fall from canopy species.

Spectra were collected from a helicopter at the three sites, associated with nine separate fire events that spanned a range of fire severities from low and patchy to high, determined by scorch height as for Russell-Smith and Edwards (2006). No extreme high fire severity events occurred during this sampling period. During October to November 2006, the efficacy of the ground-based sampling methodology was assessed, providing spectral characteristics of the savanna landscape and general changes in end-member components for a range of severities. In this method development phase, physical properties of the savanna tree–grass–soil matrix were examined and data collection methods refined. Prior to ground data collection, GPS waypoints of sites geographically coincident with spectra were downloaded from the ASD software. The coordinates collected during the 10 measurements were averaged to find a central point. 2.5. Ground data A 50 m north–south transect, centred on the GPS location, was laid out for ground sampling. Care was taken not to disturb the western side of the transect, as ground cover measurements, including char and ash, were made on this side of the transect. Ground data comprised descriptions of the site, its location, habitat type, vegetation structure, soil colour and type, recent observable fire history, and fire severity (following Russell-Smith and Edwards, 2006) (Table 1). Vegetation was described by three-dimensional vegetation structure and woody species floristic attributes. Assembled data were used to calculate stem basal area, canopy area and volume, Foliage Projective Cover (FPC) and Leaf Area Index (LAI). The method also incorporated measurements of scorch and char height to estimate fire line intensity (W m −1), using previous empirically derived correlations for regional savannas (Williams et al., 1998). In conjunction with ground patchiness measurements, these metrics were used to develop the composite fire severity index, described below. After the fire, quadrat variables (Table 2) were used to quantify the contribution of four strata to the reflectance signal, namely: upper canopy (>5 m); mid storey (2 to 5 m); lower storey (0.5 to 2 m); ground cover (b 0.5 m). Cover types in the upper and ground

Table 2 Ground variables used to assess fire severity. Variable name

Variable description

%green %litter %scorch %char %dry grass %bare soil %ash

The proportion The proportion The proportion The proportion The proportion The proportion The proportion

of photosynthetic vegetation of non-photosynthetic vegetation of scorched photosynthetic vegetation of charred vegetation of senescent grass of exposed soil of mineralised ash, i.e. carbon removed

strata were estimated using a sighting tube (Hill, 1993) via a line intercept method sampled at 1 m intervals along the 50 m transect. The sighting tube was not used to measure the cover of the mid and lower strata since vegetation in the vicinity of the user was too readily disturbed and the fire affected trees and shrubs in these strata were sparse. Instead, variables in the mid and lower strata were estimated as a percentage of total cover in ten 5 × 5 m quadrats along the 50 m transect. Error in these measurements was greater (±20%) than the sighting tube measurements (±2%), however the proportion of measureable material in the lower and mid storeys combined was only 11% on average, compared to 27% in the upper storey and 62% in the ground storey excluding bare soil. Three patches of ~ 200 ha were selected at the Howard Springs site representing (1) unburnt woodland (10–30% canopy cover) and seasonal swamp/grassland (b10% canopy cover), (2) an adjacent savanna open forest (30–50% canopy cover) burnt in July and of low to moderate severity, where ground data were collected on the two days following fire, and (3) woodland savanna (20–30% canopy cover) burnt in October at high fire severity. At the high severity woodland savanna site, ground data were collected a week after the fire. In subsequent sampling, ground data were always collected within 48 h to reduce the impact of re-flushing, and leaf fall from fire affected foliage. 3. Data analysis 3.1. Outline Steps taken in data analyses are illustrated in Fig. 3. In summary, components of the reflectance and ground data were assessed and correlated with continuous fire severity indices, explained below. Firstly, an analysis of variance (ANOVA) was conducted to assess if vegetation structure varied between sites grouped on fire severity category. Secondly, a dependent continuous variable of fire severity was derived from field measurements for correlative analysis, which was then statistically categorised using cluster analysis for a simplified error assessment of the classification. Once the spectra were transformed, fire severity was correlated with ground and reflectance variables, and used to develop a list of a priori models. Akaike's Information Criterion (AICc) (Burnham & Anderson, 2002) was then used to determine the best and most parsimonious model/s. Finally, identified best models were regressed with MODIS sensor data to determine appropriate band combinations for categorising fire severity. 3.2. Process To determine the representativeness of the site data, sites were grouped in field-based fire severity classes, and the total basal area (m 2 ha −1) for each group of sites was compared. The continuous

60

A.C. Edwards et al. / Remote Sensing of Environment 136 (2013) 56–65

ANOVA of BA in each FS class

compared to the initial field based categorical fire severity values based on the field classification scheme of Edwards (2009) to assist in assigning the category labels. The un-transformed spectra were tabulated and scatterplots derived. Logarithmic transformations of each variable, other than the 2 indices, NDVI and NBR, were created based on a visual assessment of their frequency distributions, and skewness and kurtosis.

Test: groups of sites within fire severity categories are equally representative of northern Australian tropical savanna habitats H0 rejected

H0 accepted

re-assess data

Continuous fire severity variable derived from calculation of scorch height and patchiness of unburnt ground vegetation

Calculate Fire Severity Variables

Categorical fire severity variable derived from cluster analysis of continuous fire severity variable

Ranges of spectra averaged to represent sensor band ranges

Reflectance Spectra

Reflectance Variables

Assessment of Distribution

Transformation

Correlation

Model Development AICc

Regression

Ground Data

3.3. A priori wavelength selection

Proportions of main ground variables summed for 4 strata and rescaled

Ground Variables

Assess variables for non-linear relationships using scatterplots and for asymmetrical distributions of the variables with boxplots Select and apply transformation from histogram skewness and kurtosis Derive correlation matrix and assess for significant variables Develop candidate model set from a priori knowledge and results of correlation matrix Calculate AICc, model weights and R

2

Generalised linear modelling of best model/s

Fire Severity Model for Remote Sensing

Fire Severity Variables for Ground Sampling

Fig. 3. Flow diagram describing the analytical process and algorithm development based on ground variables and associated spectra. H0 = that sites are not equally representative of savanna vegetation.

dependent fire severity variable, FSI, was calculated for all sites. Scorch height was found to be the most simple and reliable metric of fire effect as it could affect all strata and affected a high proportion of material (unlike char height), therefore it also had the least percentage error in measurement. It is most readily discernible in site photographs (Russell-Smith & Edwards, 2006) and from a distance, so could therefore be useful for rapid assessment from a moving helicopter or other vehicle. To derive the FSI for respective study sites, scorch height was weighted by the horizontal proportion of unburnt organic ground layer material as follows: FSI ¼ SH  ð100−ð%PVG þ %NPVGÞÞ=100

We also modelled the application of a series of spectral ranges and derived indices from the literature with FSI, specifically with respect to wavelengths or ranges of wavelengths partly, or wholly, captured within the channels of the MODIS sensor (Table 3). The practical application of the NBR (Allen & Sorbel, 2008; Chafer, 2008; Cocke et al., 2005; Epting et al., 2005) with ground based calibration (De Santis & Chuvieco, 2009; Key & Benson, 2006) suggests the incorporation of MODIS channels 2 (NIR1) and 7 (SWIR2) into the candidate set of models. Neither NPV (which is mostly ground litter) nor soil had any unique spectral feature in the visible-near infrared (400 to 1100 nm) wavelength region (Nagler et al., 2003). Nagler et al. (2003) states that the average depth of ligno-cellulose absorption is featured at 2.1 μm and can be used for discriminating plant litter from soil. The Cellulose Absorption Index (CAI) is represented by more discrete ranges of wavelengths, CAI = 0.5(R2.0 + R2.2) − R2.1, than those available from MODIS channels. However 2.1 μm is encapsulated within MODIS channel 7 and differences in R2.0 or R2.2 and R2.1 potentially has an effect on the average MODIS channel 7 value. Smoke aerosol particles range in size from 100 to 1000 nm, making them efficient scatterers of solar radiation. Smoke from biomass burning contains high concentrations of black carbon making it an absorbing aerosol. Data assessed by Pereira (2003) from tropical savannas in Brazil and Zambia illustrate that aerosol transmittance does not exceed 0.1 and 0.5, in the visible to near infrared domain, respectively. However, transmittance increases exponentially into and through the short wave infrared domain to 0.75 and 0.9 at 2200 nm, suggesting the incorporation of MODIS channels 6 (SWIR1) and 7 (SWIR2) into a candidate set of models. MODIS bands centred at 858 and 1240 nm are applied in the normalized difference water index, NDWI [(R860 − R1240)/(R860 + R1240)] for vegetation water content estimation (Zarco-Tejada et al., 2003). The MODIS band R1240 is on the edge of the liquid water absorption, and R858 is used for normalisation as it is insensitive to water content changes. These features make this index potentially suitable for global monitoring of vegetation water content from MODIS (Zarco-Tejada et al., 2003). The effect of atmospheric water vapour and aerosol scattering on R858 and R1240 (MODIS bands 2 (NIR1) & 5 (NIR2)) causes only small perturbations on these reflectance bands for remote estimation of water content (Gao, 1996). In summary, these previous studies suggested the assessment of MODIS equivalent channels 2, 5, 6 and 7.

ð1Þ

where %PVG = the ground stratum % of photosynthetic vegetation (PV); %NPVG = the ground stratum % of non-photosynthetic vegetation (NPV); and SH = the mean maximum height of scorched leaves, i.e. non-pyrolised foliage affected by the heat of the flame still suspended in the canopy of a shrub or tree. The ground patchiness weighting factor proportionately reduces the volume of biomass consumed. A cluster analysis of the frequency distribution of FSI provided a re-interpreted categorical fire severity variable, FSIR. This categorical variable provides utility for error assessment and map production. Cluster analysis was undertaken using STATISTICA (Stat. Soft. Inc. 2004), and provides certainty that the analyses reflected the detailed measurements and not an arbitrary classification. The result was

Table 3 Known indices for discriminating physical phenomena, to develop an a priori candidate set of models for wavelength ranges from the MODIS sensor. Index/element

Components

References

NBR

(R2.1 − R0.86)/(R2.1 + R0.86) = (MODIS2 − MODIS7)/ (MODIS2 + MODIS7) (0.5 (R2.0 + R2.2) − R2.1) MODIS7 RSWIR = MODIS6 & MODIS7 (R0.86 − R1.24)/(R0.86 + R1.24) = (MODIS2 − MODIS5)/ (MODIS2 + MODIS5)

Key and Benson (2006)

CAI Smoke penetration NDWI

Nagler et al. (2003) Pereira (2003) Gao (1996) Zarco-Tejada et al. (2003)

A.C. Edwards et al. / Remote Sensing of Environment 136 (2013) 56–65

than Unburnt Woodland in the NIR, probably due to a lack of PV, but the relationship is inverted in the SWIR due to cellulose absorption, demonstrating the higher proportion of NPV. These inverse relationships potentially afford applying a ratio or normalised difference to two indicative ranges of the electromagnetic spectrum.

4. Results 4.1. Pre-assessment of spectra Analysis of spectra collected with a hand held spectrometer in the early experimental stage of the study (Fig. 4 (a)) represented openforest savanna with and without various fire severity effects. These data suggest that, comparatively, individual components such as scorched and unscorched trees, burnt and unburnt woodland, and burnt (low severity) and burnt (high severity) woodlands, produce markedly different spectra across the near and short wave infrared, but relatively small differences in the visible. An inverse relationship was observed between NIR and SWIR regions for unburnt woodland (July) and burnt woodland (July). This inverse relationship was not evident between scorched and unscorched trees, although there is marked difference in SWIR1 reflectance, that is negligible in SWIR2. The cured grassland (July) had less reflectivity

a 1.0

bluegreenred NIR1 1 2 3 4 3 4 1 2

NIR2

4.2. Correlation between fire severity and ground variables No site effect was observed between sites grouped by fire severity category (P > 0.8). The correlation matrix of ground and FSI (Table 4) shows that significant variables (p b 0.05) comprised %Green; %Scorch; %Litter; %Dry Grass and %Bare Soil, explaining 72%, 34%, 30%, 23% and 21% of the variability, respectively. The proportion of charred material, predominantly on the ground, was uncorrelated with FSI. Whereas the proportion of mineral ash was significantly correlated with FSI, the overall proportion of ash at sites was generally

SWIR1 5 6

5

61

SWIR2 7 7

Description ETM+ band MODIS channel MODIS bands

0.9

ETM+ bands 0.8

atmosphere transmittance

% reflectance

0.7

Unburnt Woodland (July) An Unscorched Tree (July) Cured Grassland (July) A Scorched Tree (July) Burnt Woodland (October-high severity) Burnt Woodland (July-low severity)

0.6 0.5 0.4 0.3 0.2 0.1 0.0 400

600

800

1000 1200 1400 1600 1800 2000 2200 2400

wavelength [nm]

% Reflectane

b

1 0.9

unburnt

0.8

low

0.7

moderate

0.6

high

0.5 0.4 0.3 0.2 0.1 0 3

4

1

2

5

6

7

NDVI

NBR

MODIS channel/index Fig. 4. (a) Spectroradiometer reflectance spectra during the initial field campaign in 2006. 30 measurements are averaged 10 times, average error ~10%. (b) The average value (n = 41) of reflectance representing MODIS channels 1 to 7, MODIS derived NDVI and NBR for unburnt low, moderate and high fire severity classes. Error bars represent the standard error around the mean. Lines are plotted for visual guidance only.

62

A.C. Edwards et al. / Remote Sensing of Environment 136 (2013) 56–65

Table 4 Correlation coefficients (R) of fire severity indices and transformed ground variables. Significant correlations in bold italics at p b 0.01; n = 36. Transformed ground variables

FSI SQRT (%Green) ASINH (%Scorch) SQRT (%Litter) %Char LN (%DryGrass) ASINH (%BareSoil)

FSI

SQRT (%Green)

ASINH (%Scorch)

SQRT (%Litter)

%Char

LN (%Dry Grass)

ASINH (%Bare Soil)

1.00

−0.87 1.00

0.62 −0.64 1.00

−0.63 0.66 −0.44 1.00

−0.02 −0.21 0.44 −0.34 1.00

−0.62 0.72 −0.41 0.45 −0.32 1.00

0.59 −0.64 0.11 −0.59 −0.17 −0.50 1.00

small — suggesting this may not be a useful variable given that no ash was discernible three days post-fire. 4.3. Model selection Akaike's Information Criterion for model selection with small values of n (AICc) (Burnham & Anderson, 2002) was used to determine the best models describing relationships between independent ground variables with FSI (Table 4). The percentage of photosynthetic vegetation (%green) explained the most variability with FSI (72%) and was used in combination with each of the other significant ground variables. ΔAICc values b 2 were observed for two models incorporating %green and %scorch, and a third including %litter, implying they were equally supported. Parsimony suggested the %green + %scorch model was superior, although %green × %scorch explained more of the variance with a similar probability of being the best model (Table 5). However, the most strongly supported model, incorporating the weighted probability of %green + %scorch + %litter, combines change in both photosynthetic and non-photosynthetic variables. 4.4. Correlation between fire severity and reflectance variables An assessment of the proportion of variance of the average reflectance value of all wavelengths within each MODIS equivalent band, after removing outliers, was b5%. Therefore, the application of a band equivalent reflectance (BER) function (Smith et al., 2010; Trigg & Flasse, 2000) was not considered necessary to create the MODIS equivalent bands. Relationships between field-based fire severity classes with reflectance values representing MODIS channels 1–7, and representative indices, are given in Fig. 4(b). Unburnt spectra were most separated from the burnt classes in the MODIS equivalent spectral ranges of channels 2 and 5. Ignoring the unburnt reflectance, in a two-class classification (severe v not-severe), moderate separation was observed in MODIS channels 5 and 6, and clear separation with MODIS NDVI and NBR indices. However, in a three-class system

(low, moderate, high), MODIS channels 2, 5, and 6 in combination were observed to separate the low from the moderate classes (which combine to create the not-severe class), providing an opportunity to derive a two stepped model for calibrating a classification as a useful operational tool. The correlation coefficient of the transformed spectral variables with FSI is given in Table 6. The near infrared (Ln-M2) is very poorly correlated, whilst both the red (Ln-m1) and SWIR2 (Ln-M7) are moderately correlated, which is reflected in the moderate correlation in NBR and NDVI. The highest correlation is with SWIR1 (Ln-M6). 4.5. Correlation between fire severity and a priori model selection AICc model results for MODIS band combinations selected a priori demonstrate that ΔAIC values for the top eight models were all b 2 (Table 7), implying equal support for described relationships. A strong interaction between MODIS channels 6 and 7 is evident in most models. There is also a strong influence from MODIS channel 2. However, it is evident that the most important variable is MODIS 6 (LnM6) as it occurs in all models where ΔAICc is b2. Adding other variables (i.e. as in the first five models in Table 7) does not enhance the model significantly. As a rule-of-thumb, if a variable is important, it should decrease AICc by > 2 (Richards, 2005) and while more complex models explain slightly increased variability, using extra variables in this instance is effectively redundant (Richards, 2005). Table 6 Correlation coefficient (R) of the log-transformed spectral variables (MODIS channels 1 to 7 and derived indices NDVI and NBR) versus FSI. Bold italicised correlations are significant at p b 0.05; n = 36.

Table 5 AICc model selection describing relationships between significant ground variables and the FSI, where ΔAIC is the difference between the model AICc value and the minimum AICc value of all models; wi is the probability of the model being the best of the set of models; and R2 describes the proportion of the variability explained by the model. Models where ΔAIC values are b2 can be considered equally best models (Burnham & Anderson, 2002). Model %Green %Green %Green %Green %Green %Green %Green %Green %Green %Green %Green

+ %Scorch + %Litter + %Scorch * %Scorch + %Litter + %Drygrass + %Baresoil * %Litter * %Drygrass * %Baresoil * %Scorch * %Litter

AICc

ΔAIC

wi

R2

156.50 157.33 157.75 158.76 158.95 160.79 160.97 161.17 162.64 163.42 164.66

0.00 0.83 1.24 2.26 2.45 4.29 4.47 4.67 6.13 6.91 8.16

0.31 0.20 0.16 0.10 0.10 0.04 0.03 0.03 0.01 0.01 0.01

0.77 0.74 0.76 0.73 0.71 0.72 0.72 0.74 0.72 0.72 0.79

Variable

FSI

Ln-M1 Ln-M2 Ln-M3 Ln-M4 Ln-M5 Ln-M6 Ln-M7 MODIS-NDVI MODIS-NBR

0.49 0.03 0.29 0.30 0.38 0.67 0.52 −0.39 −0.41

Table 7 AICc model selection results of the significant spectral variables (an explanation of terms is given in Table 5). Model LnM2 LnM6 LnM2 LnM6 LnM5 LnM6 LnM2 LnM5

+ LnM6 + LnM7 * LnM7 + LnM6 * LnM7 + LnM7 + LnM6 + LnM7 * LnM6 + LnM7 + LnM6 * LnM7

AICc

ΔAICc

wi

R2

181.20 181.22 182.15 182.37 182.83 182.94 183.08 183.17

0.00 0.02 0.95 1.18 1.64 1.74 1.88 1.97

0.20 0.19 0.12 0.11 0.09 0.08 0.08 0.07

0.54 0.54 0.56 0.49 0.52 0.44 0.55 0.55

A.C. Edwards et al. / Remote Sensing of Environment 136 (2013) 56–65

63

5. Discussion

5.2. Fire severity models

The indices, summarised in Table 4, suggested a candidate set of models incorporating MODIS channels 2, 5, 6 and 7 can be useful for fire severity classification. The global algorithm for systematic fire affected area mapping (Roy et al., 2002) also uses these same channels. The proportion of photosynthetic vegetation was found to be the main driver in all models describing fire severity effects from the ground data. However, NIR reflectance alone is very poorly correlated with FSI. Although, well known for discriminating burnt areas (Lopez-Garcia & Caselles, 1991; Veraverbeke et al., 2011), the NIR has demonstrated poor differentiation of fire effects in the savannas (Pereira, 2003) and other landscape types (Libonati et al., 2010; Mohler & Goodin, 2010) as it shows little internal variance in reflectance values. This narrow reflectance range renders the NIR inadequate for discriminating fire severity levels within burned areas. Effective discrimination of a spectral class and detailed quantification of its internal variability are conflicting goals. The fire severity models will not be required to detect burnt areas, they will be nested under the burnt area algorithm, and map the severity within previously mapped burnt areas. The model including the proportions of photosynthetic and non-photosynthetic vegetation (including scorched leaves) had the highest AICc score. The a priori model selection process suggested that in the optical spectrum short wave infrared reflectance was most able to penetrate smoke and to detect the proportion of non-photosynthetic fuel and moisture (Table 3). In these analyses we included the unburnt data as a control class, therefore the modelling included the NIR (Ln-M2), yet if excluded there is no significant decrease in AICc (0.02 for LnM6 * LnM7) and the correlation coefficient remains the same.

The proportions of PV and NPV, with the inclusion of the proportion of foliage scorched by fire, effectively NPV, produced the best model from a candidate set of models attempting to incorporate all significant end members. The proportion of charred material was not significantly correlated with fire severity and displayed only a weak correlation with the proportion of scorched leaves. These analyses clearly demonstrate the insignificance of the proportion of charred material as a driver in any modelling of fire effects in north Australian tropical savanna habitats. Smith et al. (2007) derived ΔNBR and %char from Landsat imagery using linear spectral un-mixing analysis and predicted the one-year post-fire canopy condition metric “percentage live trees” (Smith et al., 2007). Smith et al. (2007) hypothesise that the char fraction value may be considered a potential surrogate measure of the fire intensity. The analysis in this study found no such indication of its utility, except to suggest that in tropical savanna habitats it may indicate the difference between an extreme and not-extreme fire severity event, as completely charred leaves in the upper canopy might greatly increase the proportion of charred material if a significant proportion of PV became and remained charred but was not pyrolised. More data are required for extreme fire severity events to support this. In this study, charred material was not evident in the upper canopy at any site, and perhaps indicates its relative occurrence in these habitats. Initially, a visual assessment of the average value of each spectral variable within each fire severity category was undertaken. This assisted in developing a priori information regarding the ability of each variable to differentiate the burnt categories of the fire severity classes, i.e. excluding data from unburnt sites. The result of the assessment suggested the development of two algorithms to differentiate low/moderate (not-severe) from high severity (severe) events using ΔNBR. This index can provide an accurate and potentially useful operational output. Secondly, a simple algorithm using MODIS channel 6 was able to differentiate low from moderate fire severity events. Further a priori models were derived from the literature for physical phenomena that, individually or in combination, could be relevant to fire severity class differentiation. Eight models of the resultant candidate set highlighted the importance of the near infrared if incorporated into models with the two short wave infrared channels (channels 6 and 7). However, through the interpretation of the AICc analysis, there is no justification in employing such a complex model when a model exclusively using MODIS channel 6 appears to be equally effective, and best meets the requirement of parsimony.

5.1. Sources of error The increased reliance on remotely sensed information to provide end-users with accurate mapping requires high quality sampling for calibration and validation data. Typical associated errors include the adverse effects of the atmosphere, such as smoke from fire, geolocation error, between ground sites and image pixels, and ground sample size. The helicopter based sensor in this study markedly reduced all of these errors. Reflectance data collected by satellitebased sensors suffer from atmospheric effects, more so in the tropics (Kanniah et al., 2010), especially smoke aerosol effects occurring throughout the fire season. There are often issues with geo-location, resulting in data aggregation and image filtering. The accuracy of the co-location of the coupled sampling methods in this study removed a large component of geo-rectification error that would normally be associated with the application of field data to moderate (e.g. MODIS) and more so to high (e.g. Landsat) resolution satellite based remotely sensed data. A sampling methodology must also be scaled appropriately (Lentile et al., 2006) and in this study the ground sample size was ~ 10% of the spectral field-of-view and was always within the exact area of reflectance data acquisition. Additionally, the helicopter based instrument allowed timely collection of reflectance spectra on the day after a fire. Visual assessment of spectra and site photographs suggested that reflectance spectra are heavily influenced by phenological changes that continue to occur post-fire via leaf drop and accumulation of this leaf litter on the soil surface. Therefore, the sampling period was set at 24 and 48 h post-fire for spectra and ground data respectively. The collection methods for this study included all possible ground variables and optical spectra as suggested by the CBI, GeoCBI and other remotely sensed ground data collection methods. The sampling methodology was thorough and repeatable. Near-ground data acquisition of spectra overcame the short-comings of the data preprocessing that generally results in uncertainty (Roy et al., 2006).

6. Conclusion The detailed ground assessment underpinning these analyses provided us with a minimum set of variables to characterise fire effects and to determine the suitability and habitat placement of sites. Ground data collection for fire severity class calibration or validation should include measures of ground patchiness, scorch height, and tree height at a minimum, and should include basal area and projected foliage cover where possible. To further develop an operational and automated fire severity map product, the suitable post-fire period for application of the models needs to be determined. Advances in automated burnt area mapping in the last decade provide researchers with the opportunity to develop other automated fire effects algorithms solely within burnt areas. Differentiation of fire severity classes enables a more robust assessment of fire affected landscapes. In this study, the ground layer, sampled at unburnt sites (n = 5) across the 3 regions, consisted of 24% PV, 67% NPV (including senescent grasses) and 9% bare soil, whilst of the total PV, 44% occurs in the upper canopy, 11% in the mid storey (2–5 m), 8% in the lower storey (0.5–2 m) and 37% in the ground layer. Therefore 10% of all measureable area is upper canopy PV, and 15% is woody canopy

64

A.C. Edwards et al. / Remote Sensing of Environment 136 (2013) 56–65

(stems, branches, etc.). The inference here is that although %green (PV) is the variable most obviously affected by fire severity, it is not the dominant cover type in either the pre- or post-fire landscape. As a result, the change in NPV, including PV that becomes NPV, i.e. scorched leaves, strongly influences the reflectance signal. This study has determined that the short wave infrared most clearly detects the changes in the effect fire has on tropical savanna vegetation. The parsimonious model uses the band-slice simulating MODIS channel 6, however the model including the simulation of channels 6 and 7 explained the most variation. Although this research was specifically undertaken for potential application in the Eucalypt dominated tropical savannas of northern Australia, the results highlight the utility of the short wave infrared to discriminate fire effects on vegetation generally. Empirical models for any fire affected landscape could be readily developed with sufficient ground data for calibration. The first two important steps are to identify the requirements of the mapping, and then to work with this end-user group to characterise appropriate fire severity classes. Acknowledgements This project was funded by the Bushfire Cooperative Research Centre, Bushfires NT and Charles Darwin University. Special thanks also go to Cameron Yates, Michael Carter, Kevin Natt, Lee Humpris, Jess Jennings and Dean Yibarbuk for field support. References Allen, J. L., & Sorbel, B. (2008). Assessing the differenced Normalized Burn Ratio's ability to map burn severity in the boreal forest and tundra ecosystems of Alaska's national parks. International Journal of Wildland Fire, 17, 463–475. Beringer, J., Hutley, L. B., Tapper, N. J., & Cernusak, L. A. (2007). Savanna fires and their impact on net ecosystem productivity in north Australia. Global Change Biology, 13, 990–1004. Beringer, J., Hutley, L. B., Tapper, N. J., Coutts, A., Kerley, A., & O'Grady, A. P. (2003). Fire impacts on surface heat, moisture and carbon fluxes from a tropical savanna in northern Australia. International Journal of Wildland Fire, 12, 1–8. Bowman, D. M. J. S., Zhang, Y., Walsh, A., & Williams, R. J. (2003). Experimental comparison of four remote sensing techniques to map tropical savanna fire-scars using Landsat-TM imagery. International Journal of Wildland Fire, 12, 341–348. Burnham, K. P., & Anderson, D. R. (2002). Model selection and multi-model inference: A practical information theoretic approach. New York, USA: Springer-Verlag. Chafer, C. J. (2008). A comparison of fire severity measures: An Australian example and implications for predicting major areas of soil erosion. Catena, 74, 235–245. Chafer, C. J., Noonan, M., & MacNaught, E. (2004). The post-fire measurement of fire severity and intensity in the Christmas 2001 Sydney wildfires. International Journal of Wildland Fire, 13, 227–240. Chappell, C. B., & Agee, J. K. (1996). Fire severity and tree seedling establishment in abies magnifica forests, southern cascades, Oregon. Ecological Applications, 6, 628–640. Chuvieco, E., & Congalton, R. G. (1988). Mapping and inventory of forest fires from digital processing of TM data. Geocarto International, 3(4), 41–53. Cocke, A. E., Fule, P. Z., & Crouse, J. E. (2005). Comparison of burn severity assessments using Differenced Normalized Burn Ratio and ground data. International Journal of Wildland Fire, 14, 189–198. Cook, G. D. (2003). Fuel dynamics, nutrients and atmospheric chemistry. In A. Andersen, G. Cook, & R. Williams (Eds.), Fire in tropical savannas: the Kapalga experiment (pp. 47–58). New York: Springer. Cook, G. D., Williams, R. J., Stokes, C. J., Hutley, L. B., Ash, A. J., & Richards, A. E. (2010). Managing sources and sinks of greenhouse gases in Australia's rangelands and tropical savannas. Rangeland Ecology & Management, 63, 137–146. De Santis, A., & Chuvieco, E. (2009). GeoCBI: A modified version of the Composite Burn Index for the initial assessment of the short-term burn severity from remotely sensed data. Remote Sensing of Environment, 113, 554–562. Dubinin, M., Potapov, P., Lushchekina, A., & Radeloff, V. C. (2010). Reconstructing long time series of burned areas in arid grasslands of southern Russia by satellite remote sensing. Remote Sensing of Environment, 114, 1638–1648. Savanna burning: Understanding and using fire in northern Australia. Dyer, R., Jacklyn, P., Partridge, I., Russell-Smith, J., & Williams, R. J. (Eds.). (2001). Darwin, Northern Territory, Australia: Tropical Savannas Management Cooperative Research Centre Darwin. Edwards, A. C. (2009). Fire Severity Categories for the Tropical Savanna Woodlands of northern Australia. Melbourne, Victoria, Australia: Bushfire Cooperative Research Centre. Edwards, A., & Russell-Smith, J. (2009). Ecological thresholds and the status of fire-sensitive vegetation in western Arnhem Land, northern Australia: Implications for management. International Journal of Wildland Fire, 18, 127–146.

Epting, J., Verbyla, D., & Sorbel, B. (2005). Evaluation of remotely sensed indices for assessing burn severity in interior Alaska using Landsat TM and ETM+. Remote Sensing of Environment, 96, 328–339. Ferguson, S. A., Ruthford, J. E., McKay, S. J., Wright, D., Wright, C., & Ottmar, R. (2002). Measuring moisture dynamics to predict fire severity in longleaf pine forests. International Journal of Wildland Fire, 11, 267–279. French, N. H. F., Kasischke, E. S., Hall, R. J., Murphy, K. A., Verbyla, D. L., Hoy, E. E., et al. (2008). Using Landsat data to assess fire and burn severity in the North American boreal forest region: An overview and summary of results. International Journal of Wildland Fire, 17, 443–462. Fule, P. Z., Crouse, J. E., Heinlein, T. A., Moore, M. M., Covington, W. W., & Verkamp, G. (2003). Mixed-severity fire regime in a high-elevation forest of Grand Canyon, Arizona, USA. Landscape Ecology, 18, 465–486. Gao, B. C. (1996). NDWI—A normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sensing of Environment, 58, 257–266. Hammill, K. A., & Bradstock, R. A. (2006). Remote sensing of fire severity in the Blue Mountains: Influence of vegetation type and inferring fire intensity. International Journal of Wildland Fire, 15, 213–226. Hill, T. (1993). Taking the " " out of "Ground Truth": objective accuracy assessment. In, 12th Pecora Conference, South Dakota, USA: Sioux Falls. Holden, Z. A., Morgan, P., & Evans, J. S. (2009). A predictive model of burn severity based on 20-year satellite-inferred burn severity data in a large southwestern US wilderness area. Forest Ecology and Management, 258, 2399–2406. Hudak, A., & Brockett, B. (2004). Mapping fire scars in a southern African savannah using Landsat imagery. International Journal of Remote Sensing, 25, 3231–3243. Jackson, R. D., Clarke, T. R., & Moran, M. S. (1992). Bidirectional calibration results for 11 Spectralon and 16 BaSO4 reference reflectance panels. Remote Sensing of Environment, 40, 231–239. Justice, C. O., Giglio, L., Korontzi, S., Owens, J., Morisette, J. T., Roy, D., et al. (2002). The MODIS fire products. Remote Sensing of Environment, 83, 244–262. Kanniah, K., Beringer, J., Tapper, N., & Long, C. (2010). Aerosols and their influence on radiation partitioning and savanna productivity in northern Australia. Theoretical and Applied Climatology, 100, 423–438. Karau, E. C., & Keane, R. E. (2010). Burn severity mapping using simulation modelling and satellite imagery. International Journal of Wildland Fire, 19, 710–724. Keane, R. E., Burgan, R., & van Wagtendonk, J. (2001). Mapping wildland fuels for fire management across multiple scales: Integrating remote sensing, GIS, and biophysical modeling. International Journal of Wildland Fire, 10, 301–319. Keeley, J. E. (2009). Fire intensity, fire severity and burn severity: A brief review and suggested usage. International Journal of Wildland Fire, 18, 116–126. Key, C. H., & Benson, N. C. (2006). Landscape Assessment: Ground measure of severity, the Composite Burn Index; and Remote sensing of severity, the Normalized Burn Ratio. In D. C. Lutes, R. E. Keane, J. F. Caratti, C. H. Key, N. C. Benson, S. Sutherland, & L. J. Gangi (Eds.), FIREMON: Fire Effects Monitoring and Inventory System. Rocky Mountain Research Station, Ogden, UT: USDA Forest Service. Lentile, L. B., Holden, Z. A., Smith, A. M. S., Falkowski, M. J., Hudak, A. T., Morgan, P., et al. (2006). Remote sensing techniques to assess active fire characteristics and post-fire effects. International Journal of Wildland Fire, 15, 319–345. Libonati, R., DaCamara, C. C., Pereira, J. M. C., & Peres, L. F. (2010). Retrieving middle-infrared reflectance for burned area mapping in tropical environments using MODIS. Remote Sensing of Environment, 114, 831–843. Lopez-Garcia, M. J., & Caselles, V. (1991). Mapping burns and natural reforestation using thermatic mapper data. Geocarto International, 31–37. Lynch, B. T., & Wilson, P. L. (1998). Land Systems of Arnhem Land. Technical Report, no. 97/1, Darwin: Northern Territory: Department of Lands, Planning & Environment. Maier, S. W. (2005). Automatic burnt area mapping with MODIS. WASTAC Annual Report 2005, Western Australian Satellite Technology and Applications Consortium (WASTAC). Perth, Australia: WASTAC. Maier, S. W. (2010). Changes in surface reflectance from wildfires on the Australian continent measured by MODIS. International Journal of Remote Sensing, 31, 3161–3176. Maier, S. W., & Russell-Smith, J. (2012). Measuring and monitoring of contemporary fire regimes in Australia using satellite remote sensing. In R. A. Bradstock, A. M. Gill, & R. J. Williams (Eds.), Flammable Australia. Fire regimes, biodiversity and ecosystems in a changing world (pp. 79–95). Collingwood, Australia: CSIRO Publishing. Meyer, M. C., Russell-Smith, J., Murphy, B., Cook, G. D., Edwards, A., Yates, C., et al. (2008). Accounting and verification of greenhouse gas emissions from fire management programs in northern Australia. International Bushfire Research Conference 2008 — incorporating The 15th annual AFAC Conference. Adelaide, South Australia, Australia. Mitri, G. H., & Gitas, I. Z. (2002). The development of an object oriented classification model for operational burned area mapping on the Mediterranean island of Thasos using LANDSAT TM images. In Viegas (Ed.), Forest Fire Research & Wildland Fire Safety90-77017-72-0. Mohler, R. L., & Goodin, D. G. (2010). A comparison of red, NIR, and NDVI for monitoring temporal burn signature change in tallgrass prairie. Remote Sensing Letters, 1, 3–9. Nagler, P. L., Inoue, Y., Glenn, E. P., Russ, A. L., & Daughtry, C. S. T. (2003). Cellulose absorption index (CAI) to quantify mixed soil-plant litter scenes. Remote Sensing of Environment, 87, 310–325. Noonan, M., Chafer, C., & MacNaught, E. (2002). Mapping fire severity and extent using multi-temporal RADARSAT-1 and ERS-2 SAR data. In, 11th Australasian Remote Sensing and Photogrammetry Conference. O'Grady, A. P., Chen, X., Eamus, D., & Hutley, L. B. (2000). Composition, leaf area index and standing biomass of Eucalypt open forests near Darwin in the Northern Territory, Australia. Australian Journal of Botany, 48, 629–638.

A.C. Edwards et al. / Remote Sensing of Environment 136 (2013) 56–65 Pereira, J. M. C. (2003). Remote sensing of burned areas in tropical savannas. International Journal of Wildland Fire, 12, 259–270. Pérez, B., & Moreno, J. M. (1998). Methods for quantifying fire severity in shrubland-fires. Plant Ecology, 139, 91–101. Richards, S. A. (2005). Testing ecological theory using the information-theoretic approach: Examples and cautionary results. Ecology, 86, 2805–2814. Röder, A., Hill, J., Duguy, B., Alloza, J. A., & Vallejo, R. (2008). Using long time series of Landsat data to monitor fire events and post-fire dynamics and identify driving factors. A case study in the Ayora region (eastern Spain). Remote Sensing of Environment, 112, 259–273. Rogan, J., & Franklin, J. (2001). Mapping wildfire burn severity in Southern California forests and shrublands using enhanced Thematic Mapper Imagery. Geocarto International, 16, 91–106. Roldan-Zamarron, A., Merino-de-Miguel, S., Gonzalez-Alonso, F., Garcia-Gigorro, S., & Cuevas, J. M. (2006). Minas de Riotinto (south Spain) forest fire: Burned area assessment and fire severity mapping using Landsat 5-TM, Envisat-MERIS, and Terra-MODIS postfire images. Journal of Geophysical Research, 111. Roy, D. P., Boschetti, L., & Trigg, S. N. (2006). Remote sensing of fire severity: Assessing the performance of the normalized burn ratio. IEEE Geoscience and Remote Sensing Letters, 3, 112–116. Roy, D. P., Jin, Y., Lewis, P. E., & Justice, C. O. (2005). Prototyping a global algorithm for systematic fire-affected area mapping using MODIS time series data. Remote Sensing of Environment, 97, 137–162. Roy, D. P., Lewis, P. E., & Justice, C. O. (2002). Burned area mapping using multi-temporal moderate spatial resolution data — A bi-directional reflectance model-based expectation approach. Remote Sensing of Environment, 83, 263–286. Russell-Smith, J., & Edwards, A. C. (2006). Seasonality and fire severity in savanna landscapes of monsoonal northern Australia. International Journal of Wildland Fire, 15, 541–550. Russell-Smith, J., Edwards, A. C., Cook, G. D., Schatz, J., & Brocklehurst, P. (2006). Estimating greenhouse gas emissions from savanna burning in northern Australia. (Figuiera da Foz, Portugal). Fifth International Conference on Forest Fire Research. Forest Ecology and Management, Vol. 234 (Supplement 1). (pp. S152). Elsevier Press. Russell-Smith, J., Edwards, A. C., & Price, O. F. (2012). Simplifying the savanna: The trajectory of fire-sensitive vegetation mosaics in northern Australia. Journal of Biogeography. Russell-Smith, J., Murphy, B. P., Meyer, C. P., Cook, G. D., Maier, S., Edwards, A. C., et al. (2009). Improving estimates of savanna burning emissions for greenhouse accounting in northern Australia: Limitations, challenges, applications. International Journal of Wildland Fire, 18, 1–18. Russell-Smith, J., & Yates, C. P. (2007). Australian savanna fire regimes: Context, scales, patchiness. Fire Ecology, 3, 48–63 (Special Issue). Russell-Smith, J., Yates, C. P., Whitehead, P. J., Smith, R., Craig, R., Allan, G. E., et al. (2007). Bushfires ‘down under’: Patterns and implications of contemporary Australian landscape burning. International Journal of Wildland Fire, 16, 361–377. Sa, A. C. L., Pereira, J. M. C., Vasconcelos, M. J. P., Silva, J. M. N., Ribeiro, N., & Awasse, A. (2003). Assessing the feasibility of sub-pixel burned area mapping in miombo woodlands of northern Mozambique using MODIS imagery. International Journal of Remote Sensing, 24, 1783–1796. Schimmel, J., & Granstrom, A. (1996). Fire severity and vegetation response in the boreal Swedish forest. Ecology, 77, 1436–1450. Smith, A. M. S., Eitel, J. U. H., & Hudak, A. T. (2010). Spectral analysis of charcoal on soils: Implications for wildland fire severity mapping methods. International Journal of Wildland Fire, 19, 976–983. Smith, A. M. S., & Hudak, A. T. (2005). Estimating combustion of large downed woody debris from residual white ash. International Journal of Wildland Fire, 14, 245–248.

65

Smith, A. M. S., Lentile, L. B., Hudak, A. T., & Morgan, P. (2007). Evaluation of linear spectral unmixing and ΔNBR for predicting post-fire recovery in a North American ponderosa pine forest. International Journal of Remote Sensing, 28, 5159–5166. Smith, A. M. S., Wooster, M. J., Drake, N. A., Dipotso, F. M., Falkowski, M. J., & Hudak, A. T. (2005). Testing the potential of multi-spectral remote sensing for retrospectively estimating fire severity in African Savannahs. Remote Sensing of Environment, 97, 92–115. Strand, H., Höft, R., Strittholt, J., Miles, L., Horning, N., Fosnight, E., et al. (2007). Sourcebook on Remote Sensing and Biodiversity Indicators. Secretariat of the Convention on Biological Diversity, Montreal, Technical Series, no. 32. (pp. 203). Stroppiana, D., Pinnock, S., Pereira, J. M. C., & Gregoire, J. -M. (2002). Radiometric analysis of SPOT-VEGETATION images for burnt area detection in northern Australia. Remote Sensing of Environment, 82, 21–37. Taylor, F. (2004). Field guide for the ASD FieldSpec Pro — White Reference Mode. Southampton UK: Field Spectroscopy Facility — Natural Environment Research Council. Trigg, S., & Flasse, S. (2000). Characterizing the spectral-temporal response of burned savannah using in situ spectroradiometry and infrared thermometry. International Journal of Remote Sensing, 21. Ustin, S. L., Zarco-Tejada, P. J., Jacquemoud, S., & Asner, G. P. (2004). Chapter 13: Remote sensing of the environment: State of the science and new directions. In S. L. Ustin (Ed.), Remote sensing for natural resource management and environmental monitoring: Manual of remote sensing. Hoboken, New Jersey, USA: John Wiley & Sons, Inc. van der Werf, G. R., Randerson, J. T., Giglio, L., Collatz, G. J., Mu, M., Kasibhlatla, P. S., et al. (2010). Global fire emissions and the contribution of deforestation, savanna, forest, agricultural, and peat fires (1997–2009). Atmospheric Chemistry and Physics Discussions, 10, 16153–16230. Veraverbeke, S., Harris, S., & Hook, S. (2011). Evaluating spectral indices for burned area discrimination using MODIS/ASTER (MASTER) airborne simulator data. Remote Sensing of Environment, 115, 2702–2709. White, J. D., Ryan, K. C., Key, C. C., & Running, S. W. (1996). Remote sensing of forest fire severity and vegetation recovery. International Journal of Wildland Fire, 6(3), 125–136. Whight, S., & Bradstock, R. (1999). Indices of fire characteristics in sandstone heath near Sydney, Australia. International Journal of Wildland Fire, 9, 145–153. Williams, R. J., Gill, A. M., & Moore, P. H. R. (1998). Seasonal changes in fire behaviour in a tropical savanna in northern Australia. International Journal of Wildland Fire, 8, 227–239. Williams, R. J., Griffiths, A. D., & Allan, G. E. (2002). Fire regimes and biodiversity in the savannas of northern Australia. In R. A. Bradstock, J. E. Williams, & M. A. Gill (Eds.), Flammable Australia: The fire regimes and biodiversity of a continent (pp. 281–304). Cambridge: Cambridge University Press. Williams, R. J., Wahren, C. -H., Bradstock, R. A., & Müller, W. J. (2006). Does alpine grazing reduce blazing? A landscape test of a widely-held hypothesis. Austral Ecology, 31, 925–936. Woinarski, J. C. Z., Armstrong, M., Brennan, K., Fisher, A., Griffiths, A. D., Hill, B., et al. (2010). Monitoring indicates rapid and severe decline of native small mammals in Kakadu National Park, northern Australia. Wildlife Research, 37, 116–126. Yates, C. P., Edwards, A. C., & Russell-Smith, J. (2008). Big fires and their ecological impacts in Australian savannas: Size and frequency matters. International Journal of Wildland Fire, 17, 768–781. Yibarbuk, D., Whitehead, P. J., Russell-Smith, J., Jackson, D., Godjuwa, C., Fisher, A., et al. (2001). Fire ecology and Aboriginal land management in central Arnhem Land, northern Australia: A tradition of ecosystem management. Journal of Biogeography, 28, 325–343. Zarco-Tejada, P. J., Rueda, C. A., & Ustin, S. L. (2003). Water content estimation in vegetation with MODIS reflectance data and model inversion methods. Remote Sensing of Environment, 85, 109–124.