- Email: [email protected]
A new burn severity index based on land surface temperature and enhanced vegetation index
International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
Contents lists available at ScienceDirect
International Journal of Applied Earth Observation and Geoinformation journal homepage: www.elsevier.com/locate/jag
A new burn severity index based on land surface temperature and enhanced vegetation index Zhong Zheng a,b , Yongnian Zeng a,b,∗ , Songnian Li c , Wei Huang c a b c
School of Geoscience and Info-Physics, Central South University, Changsha 410083, Hunan, China Spatial Information Technology and Sustainable Development Research Center, Central South University, Changsha 410083, Hunan, China Department of Civil Engineering, Ryerson University, Toronto, Ontario M5B 2K3, Canada
a r t i c l e
i n f o
Article history: Received 4 September 2015 Received in revised form 3 November 2015 Accepted 3 November 2015 Available online 19 November 2015 Keywords: Burn severity Vegetation index Land surface temperature Forest fire
a b s t r a c t Remotely sensed data have already become one of the major resources for estimating the burn severity of forest fires. Recently, Land Surface Temperature (LST) calculated from remote sensing data has been considered as a potential indicator for estimating burn severity. However, using the LST-based index alone may not be sufficient for estimating burn severity in the areas that has unburned trees and vegetation. In this paper, a new index is proposed by considering LST and enhanced vegetation index (EVI) together. The accuracy of the proposed index was evaluated by using 264 composite burn index (CBI) field sample data of the five fires across different regional eco-type areas in the Western United States. Results show that the proposed index performed equally well for post-fire areas covered with both sparse vegetation and dense vegetation and relatively better than some commonly-used burn severity indices. This index also has high potential of estimating burn severity if more accurate surface temperatures can be obtained in the future. © 2015 Elsevier B.V. All rights reserved.
1. Introduction Forest fire, as a major disturbance agent in ecological community, yearly affects and even totally removes millions of hectares of forest land around the world (Quintano et al., 2013). It is considered as a major cause of biodiversity reduction, soil fertility loss, gaseous pollutants emission, and other environmental impacts (Vasconcelos et al., 2013). Measurements of the post-fire damage levels over burned areas are critical to quantifying fire’s impact on landscapes (van Wagtendonk et al., 2004) and improving postdisaster management (Veraverbeke et al., 2012a), which have been widely used by environmental scientists, forest fire researchers, and policy makers. The preferable term burn severity has become a standard measurement of environmental damage levels after forest fires (Keeley 2009; Key and Benson, 2006). It can be measured accurately by conducting the field investigations (Parks et al., 2014), and also estimated by using remote sensing technologies (Keeley, 2009). For widespread forest fires, the remote sensing technology has
∗ Corresponding author at: Central South University, No. 932, South Lushan Road, Yuelu District, Changsha City 410083, China. E-mail addresses: [email protected] (Z. Zheng), [email protected] (Y. Zeng). http://dx.doi.org/10.1016/j.jag.2015.11.002 0303-2434/© 2015 Elsevier B.V. All rights reserved.
been considered as the most appropriate method to estimate the burn severity, due to the fact that it does not need a great amount of time, money and resources (Loboda et al., 2013). The remote sensing technology for burn severity estimation is primarily based on the quantitative relationships between field-observed data and severity-related indicators derived from remote sensing images, e.g., spectral index (i.e., calculated using spectral band reflectance, see also (Veraverbeke et al., 2012a)) and LST. Among these indicators, the spectral index has been used more widely, since it can be easily calculated and straightforwardly applied (Quintano et al., 2015). There are several typical spectral indices: Normalized Difference Vegetation Index (NDVI), Normalized Burn Ratio (NBR) and Enhanced Vegetation Index (EVI) (Chen et al., 2011; De Santis et al., 2010; Escuin et al., 2008; Key and Benson, 1999), which have been used to estimate burn severity using remote sensing images. Further, their differenced versions (e.g., deltaNBR) have been introduced to indicate the change level of forest community before and after fires (i.e., burn severity) (Key and Benson, 2006). As an effective and popular spectral index, the assessment ofi performance has been reported in many studies (Epting et al., 2005; Hall et al., 2008; Parks et al., 2014; Soverel et al., 2010). Following that, some researchers have taken into consideration the effects of the pre-fire vegetation on estimating burn severity and introduced the relative versions of differenced spectral indices (e.g., RdNBR and RNBR) (Miller and Thode, 2007; Parks et al., 2014). However, it has been
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
argued that many spectral indices like dNBR have limitations on extracting characteristics related to burn severity (De Santis and Chuvieco, 2007; Quintano et al., 2013) and the performances of these spectral indices in estimating burn severity still are actively debated (Soverel et al., 2010; Veraverbeke et al., 2012a). More recently, several studies have indicated that the land surface temperature (LST), as a surface biophysical parameter describing the balance of water, energy and CO2 on the forest land surface, may be another attractive indicator for estimating burn severity of forest fires (Quintano et al., 2015; Veraverbeke et al., 2012b; Vlassova et al., 2014). For examples, Veraverbeke et al., (2012b) evaluated the LST change in forest before and after fire and assessed its potential as an indicator for burn severity using MODIS images. Subsequently, Vlassova et al. (2014) assessed the spatio-temporal patterns of LST and burn severity, and found that the magnitude of LST differences was directly related to the burn severity of burned areas. Recently, Quintano et al. (2015) evaluated the usefulness of post-fire LST for mapping the spatial distribution of burn severity in Mediterranean forests. However, the forest fire not only altered the spatial distribution of land surface temperature, but also the vegetation condition. It is imperfect to only use LST to estimate burn severity. Especially, in areas that have undergone less severe forest fire, trees and vegetation might still cover some burned areas after fire. Therefore, it is certainly needed to combine LST with vegetation index to assess burn severity of forest fire. In order to address limitations of current methods for more accurately estimating burn severity, we try to build a new index of burn severity by combing LST and vegetation index, which can fully use multi-spectral remote sensing data. This paper is structured as follows. This section introduces background of burn severity estimation. The methods used in this study are described in Section 2, followed by a brief description of study area, field data, and remotely sensed data in Section 3. Results and discussion are presented in Section 4 and Section 5, respectively. Finally, Section 6 gives the conclusion of this study.
≈
85
2 Tat−sensor
(3)
b Lat−sensor
ı ≈ Tat−sensor −
2 Tat−sensor
(4)
b
where Lat−sensor means the at-sensor radiance of thermal infrared band (band 6 for TM and band 6L for ETM+); Tat−sensor indicates the at-sensor brightness temperature of thermal infrared band; for TM image band 6, K1 = 607.76 W m−2 sr−1 m−1 , K2 = 1260.56 K, and b = 1256 K; for ETM+ image band 6, K1 = 666.09 W m−2 sr−1 m−1 , K2 = 1282.71 K, and b = 1277 K. For the three atmospheric functions (i.e.,ϕ1 , ϕ2 , and ϕ3 ), we calculated them by using the atmospheric water vapor content (w), which was obtained from Total Precipitable Water product of MODIS (MOD 05) because it was required at the time near to TM and ETM+ images. The calculation of atmospheric functions are shown as follows:
⎡
ϕ1
⎤
⎡
⎢ ⎥ ⎣ ϕ2 ⎦ = Cij
3×3
w2
⎤
⎣w ⎦
(5)
1
ϕ3
where the Cij
3×3
is the coefficients matrix. Its elements were
obtained from TIRG 61 atmospheric sounding databases in this study. Finally, the land surface emissivity () was calculated by using vegetation cover (Pv ) based on the following formula: NDVI =
Pv =
nir − red nir + red
⎧ ⎪ ⎪ ⎪ ⎨
0
NDVI − NDVIs 2
NDVIv − NDVIs ⎪ ⎪ ⎪ ⎩ 1
(6) NDVI < NDVIs NDVIs ≤ NDVI ≤ NDVIv NDVI > NDVIv
= s (1 − Pv ) + v Pv 2. Methodology In order to build a new index for estimating burn severity, we firstly calculated Land surface temperature (LST) and enhanced vegetation index (EVI) in pre- and post-forest fires. Then, a new index for estimating burn severity (i.e., deltaLST/EVI) was proposed by combining LST and EVI together. To test the effectiveness of the proposed approach, an independent validation between deltaLST/EVI and CBI field data was conducted and compared with several other indices. Finally, we classified the burn severity levels by using the deltaLST/EVI-based nonlinear regression model and assessed their accuracy using the confusion matrix. The primary operations of the methodology are shown in Fig. 1.
2.1. Calculation of land surface temperature and vegetation index In this study, land surface temperature (LST) was calculated based on the thermal infrared band of TM/ETM+ images and the generalized single-channel method, described by Quintano et al. ˜ and Sobrino (2003) and on (2015) in turn based on Jiménez-Munoz ˜ et al. (2009). The general equation is as follows: Jiménez-Munoz LST = ␥
1
Tat−sensor =
(ϕ1 Lat−sensor + ϕ2 ) + ϕ3 + ı
ln
K2 K1
Lat−sensor
+1
(1)
(2)
(7)
(8)
where nir and red refer to the reflectance of near-infrared and red spectral bands; s and v refer to soil and vegetation emissivity, which are assumed to be of 0.97 and 0.99, respectively; NDVIs and NDVIv are the NDVI of soil and vegetation, which were visually extracted from the NDVI histogram of each study area. Since the EVI is suitable for monitoring the vegetation characteristics across a variety of vegetation types (Rocha and Shaver, 2009), we selected it to indicate the vegetation characteristics and calculated the pre-fire EVI (EVIpre−fire ) and post-fire EVI (EVIpost−fire ) values of burned areas using pre- and post-fire TM/ETM+ images. The EVI was calculated as follows: EVI = 2.5 ×
nir − red nir + 6red − 7.5bule + 1
(9)
where nir , red , bule , and swir refer to the reflectance of nearinfrared-, red spectral-, and blue spectral bands in TM/ETM+ images, respectively. 2.2. A new burn severity index based on LST and EVI (deltaLST/EVI) After fire event, land surface biophysical parameters, such as land surface temperature and vegetation cover, were changed obviously and had been individually used to indicate burn severity of forest fires. After considering the variation characteristics of biophysical parameters in burned areas, a new burn severity index (deltaLST/EVI) was proposed by combining pre- and post-land surface temperature and enhanced vegetation index, which would
86
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
Fig. 1. Flowchart of methodology in this study.
enhance the differences in burn severity levels. The deltaLST/EVI was calculated by using the following general equation:
delta
LSTpre−fire LSTpost−fire LST − = EVI EVIpost−fire EVIpre−fire
deltaEVI, post-LST, deltaLST, and RatioLST/EVI ). Detailed calculations of these indices are shown as follows:
deltaNBR = NBRpre−fire − NBRpost−fire × 1000 (10)
RdNBR =
deltaNBR
abs NBRpre−fire In this formula, LST value has already been transformed into interval value ranging from 0 to 1, to make sure that it has the same magnitude order as vegetation index.
2.3. Validation procedure 2.3.1. The effectiveness of deltaLST/EVI for predicting continuous burn severity In order to test the effectiveness of the deltaLST/EVI for predicting the continuous burn severity, an independent validation experiment was conducted in this study. To be specific, the field measured CBI data for each fire were randomly divided into two groups: 70% for training the deltaLST/EVI-based nonlinear regression model and 30% for testing its prediction of continuous CBI values. For testing group of each fire, the Pearson Correlation Coefficient (R) and the Root Mean Squared Error (RMSE) were determined and compared with those of other burn severity indices-based nonlinear regression model (i.e., deltaNBR, RdNBR, RNBR, post-EVI,
RNBR =
deltaNBR
NBRpre−fire + 1.001
(11) (12)
(13)
post − EVI = EVIpost−fire
(14)
deltaEVI = EVIpre−fire − EVIpost−fire
(15)
post − LST = LSTpost−fire
(16)
deltaLST = LSTpost−fire − LSTpre−fire
(17)
RatioLST/EVI =
LSTpost−fire EVIpost−fire
(18)
2.3.2. The accuracy assessment of burn severity classification In this study, the Confusion Matrix was used to evaluate the deltaLST/EVI’s performance in differentiating the burn severity levels. Specifically, with deltaLST/EVI data as the input for above deltaLST/EVI-based nonlinear regression model, CBI values over the burned areas of each fire could be computed. Based on common CBI breakpoints (i.e., 0.1, 1.25, and 2.25), the burn severity can thus
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
87
Fig. 2. Map of the study area showing locations of the fire used in this study.
been classified into four distinct categories, such as: Unchanged: CBI ≤ 0.1, Low: 0.1 < CBI ≤ 1.25, Moderate, 1.25 < CBI ≤ 2.25, and High: 2.25 > CBI. Then, using all field measured CBI data and their burn severity classification results (Miller and Thode 2007; Parks et al., 2014), the Confusion Matrix was calculated and the related statistics (e.g., Overall Accuracy, Kappa coefficient, and the user’s and producer’s accuracy of each category) were obtained. And, these statistics were compared with those of the Monitoring Trends in Burn Severity (MTBS) classification. 3. Materials 3.1. Study area Forest fires occurred more frequently in the Western United States, which is a typical region for forest fire studies (Dillon et al., 2011; Parks et al., 2014). This region includes four typical regional eco-type areas, including Southwest, Central, Northern Rockies, and California of North American continent (Zhu et al., 2006). Five fires occurred between 08/24/2000 and 10/03/2002 in this region were selected for analysis in this study (Fig. 2), since they are representative of areas of different eco-types and each of them has sufficient field-observed burn severity data. The Bear Fire started on 06/27/2002 and ended on 07/04/02. The burned area was about 18.62 km2 . It is located at the Dinosaur National Monument in the Southwestern United States which
belongs to the Highland (alpine) climate. The annual precipitation is lower than 30 cm. The elevations range from 1700 to 2740 m (Perryman et al., 2002), which plays important roles in the distribution of vegetation. At the higher elevation regions, the predominant vegetation are ponderosa pine and Douglas-fir (Perryman et al., 2002). In the lower elevation regions, the major vegetation include willows, boxelder, narrowleaf cottonwood, big sagebrush, rubber rabbit brush, service berry, black grease wood, and so on (Perryman et al., 2002). The pinyon pine, juniper wood-land and Mormon tea dominate at the rocky canyon slopes (Perryman et al., 2002). The Jasper Fire selected as the second fire for analysis, started on 24 August 2000 and got under control a month later on 25 September 2000 (Chen et al., 2011). It was located in the Black Hills National Forest in the western South Dakota and the northeastern Wyoming, which belongs to the Semiarid Steppe climate. The average precipitation is between 44 and 67.3 cm from northwest to southeast (Brown and Sieg, 1999). The elevations range from 1050 to 2207 m. The ponderosa pine covers about 95% of Jasper Fire, following by the white spruce, Limber pine, lodgepole pine, and Rocky Mountain juniper (Brown and Sieg, 1999; Smith et al., 2007). The Mule Fire started by a lightning strike on 07/11/2002. It is situated in the Bridger-Teton National Forest of the Northern Rockies region. The climate belongs to the Highland (alpine). The mean annual precipitation is 75–115 cm. The forest areas are located between 1981 and 3353 m above sea level (Berg et al., 2012). The main forests species are subalpine fir, Engelmann spruce, lodgepole
88
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
Table 1 Summary of remote sensing images and field data collection for fires used. Regional eco-types Fire name Southwest Central Northern Rockies California
Bear Jasper Mule Pw03-Wolf
Alarm date
Pre-fire image date Post-fire image date Landsat path/row Sensors
CBI collection date (MM/DD/YYYY)
06/27/2002 08/24/2000 07/11/2002 10/03/200207/11/2002
06/13/2002 05/01/2000 09/21/2001 10/18/2001
05/28-05/30/2003 05/14-05/26/2002 09/08-09/11/2003 08/15-10/30/2003
05/31/2003 05/31/2002 09/27/2003 10/16/2003
36/32 33/30 37/30 42/34
TM TM and ETM+ TM ETM+ and TM
Table 2 Comparisons of LST calculation results over burned area: Landsat and MODIS. Fire name
Pre-fire
Post-fire
Landsat
Bear Jasper Mule Pw03-Wolf
MODIS
Landsat
MODIS
Min
Max
Mean
Min
Max
Mean
Min
Max
Mean
Min
Max
Mean
24.98 14.88 21.45 14.17
51.28 34.20 23.85 32.82
39.50 23.71 22.46 19.48
34.55 18.47 14.63 18.01
40.03 24.93 35.40 21.68
37.46 21.48 23.69 19.35
29.75 23.19 15.12 14.62
57.32 60.83 39.94 31.83
42.89 38.18 24.87 18.76
36.71 28.73 22.15 18.89
42.63 40.93 23.87 21.21
40.74 36.15 23.10 19.97
pine, whitebark pine, Douglas-fir and limber pineand aspen (Berg et al., 2012). The subalpine fir, Engelmann spruce and aspen are dominant ones in moist areas, while lodgepole pine forests dominate the dry areas (Berg et al., 2012). The whitebark pine forests can be found to form pure stands in the highest elevation regions (Berg et al., 2012). The remaining two fires, named the Pw03 Fire and Wolf Fire, are all situated in the Yosemite National Park of California region. The elevation ranges from 657 to 3997 m. The climate is Mediterranean, where mean monthly temperatures range from −3 to 32 ◦ C (Guarín and Taylor, 2005) and annual precipitation is between 80.4 cm (in lower elevations) and 172.2 cm (in higher elevations) (Lutz et al., 2009). This area is mainly covered by the Chaparral and Sierra Mixed Conifer (Zhu et al., 2006), which is mainly comprised of six conifer species: ponderosa pine, sugar pine, incense cedar, white fir, Jeffrey pine and Douglas-fir (Guarín and Taylor, 2005). In general, the P. ponderosa and the C. decurrens dominate the south-facing stands. At the north-facing sites, the predominant species are A. concolor, C. decurrens, and P. lambertiana (Guarín and Taylor, 2005). The sub-canopy hardwoods black oak and canyon live oak appears at the understory of most stands (Guarín and Taylor, 2005). Since the locations of these two fires are close to each other (i.e., 07/11-10/03/2002) and the timing of field data collection and fire broke out are near (i.e., 08/15-10/30/2003), we analyzed them together in our study (i.e., Pw03–Wolf Fire). 3.2. Field data In order to evaluate the burn severity of forest fires in the field, 262 field plots in five burned areas were surveyed through a Joint Fire Sciences Program (JFSP) project undertaken by the National Park Service and the US Geological Survey (Zhu et al., 2006). In each field, a standard metric, namely Composite Burn Index (CBI), was recorded and calculated independently during the following growing season after fire for extended assessment (Key and Benson, 2006). A more detailed description of approach and methodology for CBI collecting can be found in (Key and Benson, 2006). For Jasper fire, some of CBI field plots (8 field plots) were removed due to misregistration errors, logging influences and creek channel issues (Chen et al., 2011). In addition, we randomly selected 10CBI data outside the burn perimeter of the Jasper fire and assumed their CBI values were 0s, since there is no unchanged level CBI field data for experimental analysis in this fire. As a result, there were 55 field plots for the Bear Fire, 76 field plots for the Jasper Fire, 55 field plots for the Mule Fire, and 78 field plots for the Pw03–Wolf Fire, respectively, which were analyzed in this study.
3.3. Remotely sensed data The remotely sensed data used in this study area were acquired by three types of satellite sensors: Landsat 5 TM, Landsat 7 ETM+, and Terra MODIS (see Table 1 for more details). In all study areas, the post-fire TM and ETM+ images were selected to ensure that their acquisition time approximately matches the time of field data collection. Further, pre-fire TM and ETM+ images were selected to reduce the difference of different moisture contents and phenology between pre- and post-fire images (Key and Benson, 2006). All images were downloaded from the U.S. Geological Survey EarthExplorer (USGS, 2013a). These images have Level L1G of processing and their thermal infrared bands have already been resampled to 30-meter pixels (USGS, 2013b). Consequently, other processing procedures for reflective bands of these images were needed according to suggestions in Key and Benson (2006). These processing procedures, including the radiance transformation, the atmospheric correction, the reflectance transformation and the spectral index calculation, were conducted using ENVI 5.0 software. Then, the CBI field data was co-registered to Landsat images and the spectral attributions of remote sensing images were extracted by using 3 × 3 pixel matrix in order to reduce CBI plots to image registration error caused by GPS positioning (Chuvieco et al., 2004). In this study, we did not need to consider the scan line corrector (SLC)’s effect on ETM+ images processing, since these images were acquired before May 2003. Additionally, two MODIS products, including the Total Precipitable Water (MOD 05) and the Daily Land Surface Temperature/Emissivity (MOD11A1), were taken from the NASA’s Level 1 and Atmosphere Archive and Distribution System (LAADS Web) on the same day when TM and ETM+ images were acquired (NASA). They are used to represent the atmospheric water vapor content for LST calculation (Jiménez˜ ˜ et al., 2009; Jiménez-Munoz and Sobrino, 2003) and to Munoz validate the LST calculation results of Landsat data, respectively. Finally, each fire’s perimeter layer used as burned mask and its official burn severity classification used for comparative analysis were obtained from the Monitoring Trends in Burn Severity (MTBS, 2013). 4. Results 4.1. Pre- and post-fire LST and EVI Table 2 indicates that, for burned areas of most fires, the LST mean value derived from Landsat data shows a good agreement with that of MOD11A1 product (all mean value differences are close
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
89
Fig. 3. The spatial distributions of pre-EVI, post-EVI, pre-LST (◦ C), and post-LST (◦ C): (a) Bear fire; (b) Jasper fire; (c) Mule fire and (d) Pw03-Wolf fire.
to ±2 ◦ C) (Srivastava et al., 2009). Also, we have checked that the LST histogram differences between the Landsat calculation and MODIS product are negligible. Thus, the accuracy of LST calculations in this study is acceptable. Fig. 3 displays the spatial distributions of pre-LST, post-LST, pre-EVI, and post-EVI across four typical regional eco-type areas of the Western United States. These spectral indices reveal that forest fire has caused a great change to the coverage of vegetation and to the distribution of LST. Among these study areas, vegetation in the Mule and Pw03-Wolf fires experienced the lower degree of damage while that in the Bear and Jasper fires were seriously destroyed. On the contrary, the max., min., and mean values of deltaLST (see Table 3) in the Bear and Jasper fires are higher and those in the Mule and Pw03-Wolf fires are lower, which agreed with official sources (MTBS, 2013) and statistics in Table 3. Furthermore, statistics in Table 3 further confirm that the Mule and Pw03-Wolf fires were rich with vegetation after the fire happened (post-EVI mean value are 0.25 and 0.15 respectively). On the con-
trary, Table 3 illustrates that the post-fire areas of the Bear and Jasper fires were sparsely covered with vegetation (post-EVI mean value are 0.12 and 0.09, respectively). 4.2. The independent validation results of deltaLST/EVI Table 4 shows the results of the independent validation using the observed CBI and different burn severity indices of testing groups. The R and RMSE values in this table indicate that deltaLST/EVI performed relatively better than some commonlyused burn severity indices (e.g., dNBR, RdNBR, and RNBR). Further, it can be observed in Table 4 that the post-EVI and deltaEVI would have the better performance if the forest fire’s post-fire areas were rich with vegetation (i.e., Mule and Pw03–Wolf fires). In contrast, for forest fire whose post-fire areas were sparsely covered with vegetation (i.e., Bear and Jasper fires), the post-LST and deltaLST would performed better. Unlike them, Table 4 indicate the deltaLST/EVI
90
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
Table 3 Statistics of burn severity indices over the burned area of each fire. Fire name
Burn severity indices
Min
Max
Mean
Bear
post-EVI deltaEVI post-LST deltaLST RatioLST/EVI deltaLST/EVI post-EVI deltaEVI post-LST deltaLST RatioLST/EVI deltaLST/EVI post-EVI deltaEVI post-LST deltaLST RatioLST/EVI deltaLST/EVI post-EVI deltaEVI post-LST deltaLST RatioLST/EVI deltaLST/EVI
−0.03 −0.12 29.75 0.00 −63.74 −52.82 0.02 −0.47 23.19 2.94 0.27 −48.21 0.02 -0.19 15.12 −3.61 0.44 −5.24 0.03 −0.13 14.63 −4.81 0.31 −1.31
0.35 0.21 57.32 20.91 68.30 1416.84 0.29 0.32 60.83 40.08 16.61 15.78 0.37 0.26 39.94 15.65 29.04 27.92 0.69 0.45 31.83 6.66 14.64 13.83
0.12 0.05 42.89 9.39 6.33 3.17 0.09 0.06 38.18 18.47 2.92 0.96 0.15 0.09 24.87 5.18 4.61 2.75 0.25 0.08 18.76 −0.72 1.39 0.38
Jasper
Mule
Pw03-Wolf
performed stably and steadily in predicting the continuous burn severity across regional forest areas of different eco-types after fire. 4.3. Accuracy of mapping burn severity classification The deltaLST/EVI spatial distribution of each fire is presented in Fig. 4(A-1), (B-1), (C-1), and (D-1). And, the burn severity levels over burned areas are classified and displayed in Fig. 4(A-2), (B-2), (C-2), and (D-2). We can visually notice in these figures that the spatial distribution pattern of the deltaLST/EVI, to some extent, was similar to that of its burn severity classification results. Further, burn severity classification statistics in Table 5 show that the Overall Accuracy (OA) and Kappa coefficient are 66.29% and 0.521, respectively, which are much better than those of the MTBS classification (see Table 6: the OA is 43.56% and the Kappa Coefficient is 0.249) and are consistent with those of other study areas (e.g., OA for four classes: 52.9–71.1% (Epting et al., 2005), OA: 58.7–59.9% (Miller and Thode, 2007), OA: 59.7–70.2% (Soverel et al., 2010), and OA: 64.2–66.2% (Parks et al., 2014). We therefore confirm a fact that deltaLST/EVI performs well at classifying the burn severity levels of forest fires. 5. Discussion This study proposed a new index (i.e., deltaLST/EVI) for estimating burn severity of forest fires based on surface biophysical
parameters (i.e., EVI and LST). The results indicated that the deltaLST/EVI performed relatively better than some commonlyused burn severity indices, and its performance was stable and steady across regional forest areas of different eco-type after fires. Our findings have further developed the LST-based index and provided a new insight for estimating burn severity of forest fires. The LST has already been shown to be a useful indicator for burn severity estimation in forest Mediterranean ecosystems (Quintano et al., 2015). In order to confirm its usefulness in the forest ecosystems of western United States, the two LST-based indices (i.e., post-LST and deltaLST) were applied in this study to estimate the burn severity of forest fires. Unlike previous study (Quintano et al., 2015), the remote sensing images for post-fire LST calculation were selected to ensure that their acquisition time approximately matches the time of field data collection. Although this might not be optimal for LST-based indices performance, the experimental results in Table 4 show that the post-LST and deltaLST’s accuracies of burn severity estimation are acceptable in most cases. In particular, for severe forest fire whose post-fire areas were sparsely covered with vegetation (i.e., Bear and Jasper fires), the post-LST and deltaLST performed better even after more than 1 year. The reason is that the severe damage and the low vegetation regeneration can ensure that the obvious LST variation characteristics remained. On the contrary, for Mule and Pw03-Wolf fires, the less severe fire damage and the good vegetation regeneration after fire leads to the fading of LST variation, and thus both post-LST and deltaLST have poorer performance. This limitation of LST-based indices might limit their widespread applicability. As another surface biophysical parameter, the EVI can also be used to effectively indicate the environmental damage levels caused by fires. Analysis results in Table 4 indicate that post-EVI and deltaEVI have greater accuracy in estimating burn severity for forest fires whose post-fire areas are rich with plant life (i.e., Mule and Pw03–Wolf fires). These results echo the previous study results (Chen et al., 2011) and can be explained by the fact that EVI-based indices are sensitive to monitoring the vegetation variation characteristics across a variety of vegetation types (Rocha and Shaver, 2009). This merit is very useful for addressing the above limitation of the LST-based index. As expected, Table 4 and Fig. 5 show that the combination of LST- and EVI-based indices (i.e., RatioLST/EVI and deltaLST/EVI) have a better performance and they can make full use of burn severity information contained in surface biophysical parameters (i.e., LST and EVI). To be specific, taking the deltaLST/EVI for the Wolf fire as an example, we can find in Fig. 5(F) that the Plot WM12 location was burned severely and its true CBI value is 2.77. Since this location is covered with sparse and open vegetation after fires, it would be incorrectly classified as the low-level burn severity using the lower deltaEVI value in Fig. 5(A). Taking advantage of deltaLST, the deltaLST/EVI value in Fig. 5(C) will enable this burned location (i.e., Plot WM12) to be correctly classified as high-level severity, as
Table 4 Results of independent validation using the observed CBI and different burn severity indices. Burn severity indices
deltaNBR RdNBR RNBR post-EVI deltaEVI post-LST deltaLST RatioLST/EVI deltaLST/EVI
Bear
Jasper
Mule
Pw03–Wolf
R
RMSE
R
RMSE
R
RMSE
R
RMSE
0.71 0.63 0.72 0.71 0.59 0.77 0.66 0.72 0.81
0.69 0.77 0.68 0.74 0.80 0.83 0.76 0.56 0.53
0.70 0.74 0.70 0.70 0.72 0.73 0.61 0.72 0.79
0.69 0.66 0.69 0.67 0.65 0.67 0.74 0.65 0.59
0.86 0.87 0.88 0.82 0.82 0.64 0.69 0.87 0.91
0.57 0.56 0.54 0.64 0.63 0.88 0.80 0.57 0.39
0.69 0.70 0.73 0.44 0.73 0.29 0.60 0.58 0.74
0.64 0.59 0.57 0.74 0.55 0.79 0.67 0.67 0.56
Note: The nonlinear regression model is CBI = a × (x)2 + b × x + c and the accuracy is measured by the R and RMSE of testing group (all P values <0.05).
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
91
Fig. 4. The spatial distributions of continuous deltaLST/EVI (1) and burn severity mapping results (2), including: (a) Bear fire, (b) Jasper fire, (c) Mule fire, and (d) Pw03-Wolf fire.
92
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
Table 5 Overall classification statistics of deltaLST/EVI when classifying all field plots into distinct categories. Class name
Unchanged
Low
Moderate
High
Total
User’s accuracy (%)
Unchanged Low Moderate High Total Producer’s accuracy(%) Overall accuracy:
5 12 14 0 31 16.13 66.29%
0 38 18 0 56 67.86
0 17 56 9 82 68.29 Kappa coefficient:
1 4 14 76 95 80.00 0.521
6 71 102 85 264
83.33 53.52 54.90 89.41
Table 6 The validation results of MTBS classification. Class name
Unchanged
Low
Moderate
High
Total
User’s accuracy (%)
Unchanged Low Moderate High Total Producer’s accuracy(%) Overall accuracy:
14 11 6 0 31 45.16 43.56%
25 28 3 0 56 50.00
15 42 17 8 82 20.73 Kappa coefficient:
3 6 30 56 95 58.95 0.249
57 87 56 64 264
24.56 32.18 30.36 87.50
illustrated in Fig. 5(D). On the other hand, for the severely burned location of Plot PH09 which was covered with rich and dense vegetation (see Fig. 5(E)), it might be underestimated as low-level
severity by using the lower deltaLST value in Fig. 5(B). However, as shown in Fig. 5(C), the deltaLST/EVI can correctly estimate Plot PH09, since the burn severity information contained in deltaEVI
Fig. 5. Pw03-Wolf fire: (a) continuous deltaEVI, (b) continuous deltaLST (◦ C), (c) continuous deltaEVI/EVI, (d) burn severity classification results of deltaEVI/EVI, (e) post-firescene photo of Plot PH09 and (f) post-fire-scene photo of Plot WM12 (Key et al., 2013).
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
have been comprehensively utilized. This fact is evident in other places of the Pw03–Wolf fire and other fires in this study. We therefore can conclude that these two indicators (i.e., EVI and LST) not only relate to one another, but also complement to each other. When we combine them together, the comprehensive spatial distributions of burn severity can be estimated, and a new metric across regional areas of different eco-types, can be established and proved useful. Furthermore, the independent validation results in Table 4 indicate that the deltaLST/EVI can accurately predict the continuous burn severity with high application adaptability. And, its performance is relatively better than that of some commonly-used burn severity indices (e.g., deltaNBR, RdNBR and RNBR), which does not mean that there is no need to further develop such LST-based index. Instead, we believe that it has high potential of burn severity estimation, if more accurate surface temperature can be obtained in the future. Therefore, further research should be conducted to test the effectiveness of deltaLST/EVI calculated from the new remotely sensed data (e.g., the Landsat 8 thermal infrared data (TIRS)). Although the overall classification results in Tables 5 and 6 have indicated that deltaLST/EVI performs well at classifying most burn severity levels of forest fires, we acknowledge that the producer accuracy of unchanged level is lower (its Producer’s Accuracy is 16.13%). The reason which might explain this is that the limited number of unchanged level field data is not enough for the regression model training and the true evaluation of deltaLST/EVI performance (more than 50 verification points for per class were commonly used (Epting et al., 2005; Underwood et al., 2003)). Since the merit of LST is more noticeable in post-fire areas covered with sparse and open vegetation, the LST-based index has been developed (Quintano et al., 2015). However, this merit might be diminished if the post-fire land surface is covered with rich and dense vegetation, which might restrict the application of LST-based index. Our results in this study present that, the combination of LST and EVI does improve the burn severity estimation performance of LST-based index. 6. Conclusion Compared to commonly used indicators for estimating the burn severity of forest fires, other potential indicators (i.e., LST) are also attracting more and more attention of scientists. In this study, a new metric based on EVI and LST was constructed (i.e., deltaLST/EVI), which was validated by CBI field data of five forest fires located in the Western United States. The results show that deltaLST/EVI can take the advantages of surface biophysical parameters (i.e., EVI and LST) and can fully reveal the spatial distributions of burn severity, when the composition and characteristics of vegetation are different in the study areas. And, its performance was relatively better than that of some commonly-used burn severity indices. It also has high potential of estimating burn severity if more accurate surface temperatures can be obtained in the future. This index will be useful to environmental scientists, forest fire researchers, and policy makers. Acknowledgments The research work was supported by the National Natural Science Foundation of China (No. 41171326, 40771198) and the Fundamental Research Funds for the Central Universities of Central South University (No. 2015zzts070). The authors would like to acknowledge the data made available by the U.S. Geological Survey, the U.S. Geological Survey’s National Center for Earth Resources Observation and Science, the USDA Forest Service Remote Sensing Applications Center for making remote sensing data available, and
93
the Joint Fire Sciences Program (JFSP) project undertaken by the National Park Service and the US Geological Survey. We wish to thank the anonymous reviewers for their thoughtful and helpful comments.
References Berg, N.D., Gese, E.M., Squires, J.R., Aubry, L.M., 2012. Influence of forest structure on the abundance of snowshoe hares in western Wyoming. J. Wildl. Manage. 76, 1480–1488. Brown, P.M., Sieg, C.H., 1999. Historical variability in fire at the ponderosa pine-Northern Great Plains prairie ecotone, southeastern Black Hills, South Dakota. Ecoscience 6, 539–547. Chen, X., Vogelmann, J.E., Rollins, M., Ohlen, D., Key, C.H., Yang, L., Huang, C., Shi, H., 2011. Detecting post-fire burn severity and vegetation recovery using multitemporal remote sensing spectral indices and field-collected composite burn index data in a ponderosa pine forest. Int. J. Remote Sens. 32, 7905–7927. Chuvieco, E., Cocero, D., Riano, D., Martin, P., Martınez-Vega, J., de la Riva, J., Pérez, F., 2004. Combining NDVI and surface temperature for the estimation of live fuel moisture content in forest fire danger rating. Remote Sens. Environ. 92, 322–331. De Santis, A., Asner, G.P., Vaughan, P.J., Knapp, D.E., 2010. Mapping burn severity and burning efficiency in California using simulation models and Landsat imagery. Remote Sens. Environ. 114, 1535–1545. De Santis, A., Chuvieco, E., 2007. Burn severity estimation from remotely sensed data: performance of simulation versus empirical models. Remote Sens. Environ. 108, 422–435. Dillon, G.K., Holden, Z.A., Morgan, P., Crimmins, M.A., Heyerdahl, E.K., Luce, C.H., 2011. Both topography and climate affected forest and woodland burn severity in two regions of the western US, 1984 to 2006. Ecosphere 2, art130. 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 Sens. Environ. 96, 328–339. Escuin, S., Navarro, R., Fernandez, P., 2008. Fire severity assessment by using NBR (Normalized Burn Ratio) and NDVI (Normalized Difference Vegetation Index) derived from LANDSAT TM/ETM images. Int. J. Remote Sens. 29, 1053–1073. Guarín, A., Taylor, A.H., 2005. Drought triggered tree mortality in mixed conifer forests in Yosemite National Park, California, USA. For. Ecol. Manage. 218, 229–244. Hall, R.J., Freeburn, J., De Groot, W., Pritchard, J., Lynham, T., Landry, R., 2008. Remote sensing of burn severity: experience from western Canada boreal fires. Int. J. Wildland Fire 17, 476–489. ˜ Jiménez-Munoz, J.C., Cristóbal, J., Sobrino, J.A., Sòria, G., Ninyerola, M., Pons, X., 2009. Revision of the single-channel algorithm for land surface temperature retrieval from Landsat thermal-infrared data. IEEE Trans. Geosci. Remote Sens. 47, 339–349. ˜ Jiménez-Munoz, J.C., Sobrino, J.A., 2003. A generalized single(c)\channel method for retrieving land surface temperature from remote sensing data. J. Geophys. Res. 108, 4688–4695. Keeley, J.E., 2009. Fire intensity, fire severity and burn severity: a brief review and suggested usage. Int. J. Wildland Fire 18, 116–126. Key, C.H., Benson, N.C., 1999. The Normalized Burn Ratio (NBR): A Landsat TM radiometric measure of burn severity. http://www.nrmsc.usgs.gov/research/ ndbr.htm. Key, C.H., Benson, N.C., 2006. Landscape Assessment (LA): Sampling and Analysis Methods. 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 (pp. LA-1-LA-51). Fort Collins, CO, USA. Rocky Mountain Research Station,USDA, Forest Service, Gen. Tech. Rep. RMRS-GTR-164-CD. Key, C.H., Benson, N.C., Suzanna, S., 2013. CBI Plot Data and Photos. USGS-NOROCK and NPS. http://www.nrmsc.usgs.gov/science/fire/cbi/plotdata (20.05.14). Loboda, T.V., French, N.H.F., Hight-Harf, C., Jenkins, L., Miller, M.E., 2013. Mapping fire extent and burn severity in Alaskan tussock tundra: an analysis of the spectral response of tundra vegetation to wildland fire. Remote Sens. Environ. 134, 194–209. Lutz, J.A., Van Wagtendonk, J.W., Franklin, J.F., 2009. Twentieth-century decline of large-diameter trees in Yosemite National Park, California, USA. For. Ecol. Manage. 257, 2296–2307. Miller, J.D., Thode, A.E., 2007. Quantifying burn severity in a heterogeneous landscape with a relative version of the delta Normalized Burn Ratio (dNBR). Remote Sens. Environ. 109, 66–80. MTBS, 2013. Monitoring trends in burn severity. http://mtbs.gov/index.html. NASA, Level 1 and Atmosphere Archive and Distribution System (LAADS). http:// earthexplorer.usgs.gov/ (20.10.14). Parks, S., Dillon, G., Miller, C., 2014. A New Metric for Quantifying Burn Severity: The Relativized Burn Ratio. Remote Sens. 6, 1827–1844. Perryman, B.L., Olson, R.A., Petersburg, S., Naumann, T., 2002. Vegetation response to prescribed fire in Dinosaur National Monument. West. North Am. Nat., 414–422. Quintano, C., Fernández-Manso, A., Calvo, L., Marcos, E., Valbuena, L., 2015. Land surface temperature as potential indicator of burn severity in forest Mediterranean ecosystems. Int. J. Appl. Earth Obs. Geoinf. 36, 1–12.
94
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
Quintano, C., Fernández-Manso, A., Roberts, D.A., 2013. Multiple Endmember Spectral Mixture Analysis (MESMA) to map burn severity levels from Landsat images in Mediterranean countries. Remote Sens. Environ 136, 76–88. Rocha, A.V., Shaver, G.R., 2009. Advantages of a two band EVI calculated from solar and photosynthetically active radiation fluxes. Agric. For. Meteorol. 149, 1560–1563. Smith, A.M.S., Lentile, L.B., Hudak, A.T., Morgan, P., 2007. Evaluation of linear spectral unmixing and (NBR for predicting post(c)\fire recovery in a North American ponderosa pine forest. Int. J. Remote Sens. 28, 5159–5166. Soverel, N.O., Perrakis, D.D.B., Coops, N.C., 2010. Estimating burn severity from Landsat dNBR and RdNBR indices across western Canada. Remote Sens. Environ. 114, 1896–1909. Srivastava, P.K., Majumdar, T.J., Bhattacharya, A.K., 2009. Surface temperature estimation in Singhbhum Shear Zone of India using Landsat-7 ETM+ thermal infrared data. Adv. Space Res 43, 1563–1574. Underwood, E., Ustin, S., DiPietro, D., 2003. Mapping nonnative plants using hyperspectral imagery. Remote Sens. Environ. 86, 150–161. USGS, 2013a. Earthexplorer. http://earthexplorer.usgs.gov/ (20.12.13). USGS, 2013b. Frequently asked questions about the Landsat Missions. http:// landsat.usgs.gov/band designations landsat satellites. php (12.06.14).
van Wagtendonk, J.W., Root, R.R., Key, C.H., 2004. Comparison of AVIRIS and Landsat ETM+ detection capabilities for burn severity. Remote Sens. Environ. 92, 397–408. Vasconcelos, S.S.d., Fearnside, P.M., Grac¸a, P.M.L.d.A., Dias, D.V., Correia, F.W.S., 2013. Variability of vegetation fires with rain and deforestation in Brazil’s state of Amazonas. Remote Sens. Environ. 136, 199–209. Veraverbeke, S., Hook, S., Hulley, G., 2012a. An alternative spectral index for rapid fire severity assessments. Remote Sens. Environ. 123, 72–80. Veraverbeke, S., Verstraeten, W.W., Lhermitte, S., Van De Kerchove, R., Goossens, R., 2012b. Assessment of post-fire changes in land surface temperature and surface albedo, and their relation with fire-burn severity using multitemporal MODIS imagery. Int. J. Wildland Fire 21, 243–256. Vlassova, L., Pérez-Cabello, F., Mimbrero, M., Llovería, R., García-Martín, A., 2014. Analysis of the relationship between Land Surface Temperature and Wildfire Severity in a series of Landsat Images. Remote Sens 6, 6136–6162. Zhu, Z., Key, C., Ohlen, D., Benson, N.C., 2006. Evaluate sensitivities of burn-severity mapping algorithms for different ecosystems and fire histories in the United States. In. Joint Fire Science Program: Boise, IA, USA, 2006. Final Report to the Joint Fire Science Program, JFSP 01-1-4-12.
Contents lists available at ScienceDirect
International Journal of Applied Earth Observation and Geoinformation journal homepage: www.elsevier.com/locate/jag
A new burn severity index based on land surface temperature and enhanced vegetation index Zhong Zheng a,b , Yongnian Zeng a,b,∗ , Songnian Li c , Wei Huang c a b c
School of Geoscience and Info-Physics, Central South University, Changsha 410083, Hunan, China Spatial Information Technology and Sustainable Development Research Center, Central South University, Changsha 410083, Hunan, China Department of Civil Engineering, Ryerson University, Toronto, Ontario M5B 2K3, Canada
a r t i c l e
i n f o
Article history: Received 4 September 2015 Received in revised form 3 November 2015 Accepted 3 November 2015 Available online 19 November 2015 Keywords: Burn severity Vegetation index Land surface temperature Forest fire
a b s t r a c t Remotely sensed data have already become one of the major resources for estimating the burn severity of forest fires. Recently, Land Surface Temperature (LST) calculated from remote sensing data has been considered as a potential indicator for estimating burn severity. However, using the LST-based index alone may not be sufficient for estimating burn severity in the areas that has unburned trees and vegetation. In this paper, a new index is proposed by considering LST and enhanced vegetation index (EVI) together. The accuracy of the proposed index was evaluated by using 264 composite burn index (CBI) field sample data of the five fires across different regional eco-type areas in the Western United States. Results show that the proposed index performed equally well for post-fire areas covered with both sparse vegetation and dense vegetation and relatively better than some commonly-used burn severity indices. This index also has high potential of estimating burn severity if more accurate surface temperatures can be obtained in the future. © 2015 Elsevier B.V. All rights reserved.
1. Introduction Forest fire, as a major disturbance agent in ecological community, yearly affects and even totally removes millions of hectares of forest land around the world (Quintano et al., 2013). It is considered as a major cause of biodiversity reduction, soil fertility loss, gaseous pollutants emission, and other environmental impacts (Vasconcelos et al., 2013). Measurements of the post-fire damage levels over burned areas are critical to quantifying fire’s impact on landscapes (van Wagtendonk et al., 2004) and improving postdisaster management (Veraverbeke et al., 2012a), which have been widely used by environmental scientists, forest fire researchers, and policy makers. The preferable term burn severity has become a standard measurement of environmental damage levels after forest fires (Keeley 2009; Key and Benson, 2006). It can be measured accurately by conducting the field investigations (Parks et al., 2014), and also estimated by using remote sensing technologies (Keeley, 2009). For widespread forest fires, the remote sensing technology has
∗ Corresponding author at: Central South University, No. 932, South Lushan Road, Yuelu District, Changsha City 410083, China. E-mail addresses: [email protected] (Z. Zheng), [email protected] (Y. Zeng). http://dx.doi.org/10.1016/j.jag.2015.11.002 0303-2434/© 2015 Elsevier B.V. All rights reserved.
been considered as the most appropriate method to estimate the burn severity, due to the fact that it does not need a great amount of time, money and resources (Loboda et al., 2013). The remote sensing technology for burn severity estimation is primarily based on the quantitative relationships between field-observed data and severity-related indicators derived from remote sensing images, e.g., spectral index (i.e., calculated using spectral band reflectance, see also (Veraverbeke et al., 2012a)) and LST. Among these indicators, the spectral index has been used more widely, since it can be easily calculated and straightforwardly applied (Quintano et al., 2015). There are several typical spectral indices: Normalized Difference Vegetation Index (NDVI), Normalized Burn Ratio (NBR) and Enhanced Vegetation Index (EVI) (Chen et al., 2011; De Santis et al., 2010; Escuin et al., 2008; Key and Benson, 1999), which have been used to estimate burn severity using remote sensing images. Further, their differenced versions (e.g., deltaNBR) have been introduced to indicate the change level of forest community before and after fires (i.e., burn severity) (Key and Benson, 2006). As an effective and popular spectral index, the assessment ofi performance has been reported in many studies (Epting et al., 2005; Hall et al., 2008; Parks et al., 2014; Soverel et al., 2010). Following that, some researchers have taken into consideration the effects of the pre-fire vegetation on estimating burn severity and introduced the relative versions of differenced spectral indices (e.g., RdNBR and RNBR) (Miller and Thode, 2007; Parks et al., 2014). However, it has been
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
argued that many spectral indices like dNBR have limitations on extracting characteristics related to burn severity (De Santis and Chuvieco, 2007; Quintano et al., 2013) and the performances of these spectral indices in estimating burn severity still are actively debated (Soverel et al., 2010; Veraverbeke et al., 2012a). More recently, several studies have indicated that the land surface temperature (LST), as a surface biophysical parameter describing the balance of water, energy and CO2 on the forest land surface, may be another attractive indicator for estimating burn severity of forest fires (Quintano et al., 2015; Veraverbeke et al., 2012b; Vlassova et al., 2014). For examples, Veraverbeke et al., (2012b) evaluated the LST change in forest before and after fire and assessed its potential as an indicator for burn severity using MODIS images. Subsequently, Vlassova et al. (2014) assessed the spatio-temporal patterns of LST and burn severity, and found that the magnitude of LST differences was directly related to the burn severity of burned areas. Recently, Quintano et al. (2015) evaluated the usefulness of post-fire LST for mapping the spatial distribution of burn severity in Mediterranean forests. However, the forest fire not only altered the spatial distribution of land surface temperature, but also the vegetation condition. It is imperfect to only use LST to estimate burn severity. Especially, in areas that have undergone less severe forest fire, trees and vegetation might still cover some burned areas after fire. Therefore, it is certainly needed to combine LST with vegetation index to assess burn severity of forest fire. In order to address limitations of current methods for more accurately estimating burn severity, we try to build a new index of burn severity by combing LST and vegetation index, which can fully use multi-spectral remote sensing data. This paper is structured as follows. This section introduces background of burn severity estimation. The methods used in this study are described in Section 2, followed by a brief description of study area, field data, and remotely sensed data in Section 3. Results and discussion are presented in Section 4 and Section 5, respectively. Finally, Section 6 gives the conclusion of this study.
≈
85
2 Tat−sensor
(3)
b Lat−sensor
ı ≈ Tat−sensor −
2 Tat−sensor
(4)
b
where Lat−sensor means the at-sensor radiance of thermal infrared band (band 6 for TM and band 6L for ETM+); Tat−sensor indicates the at-sensor brightness temperature of thermal infrared band; for TM image band 6, K1 = 607.76 W m−2 sr−1 m−1 , K2 = 1260.56 K, and b = 1256 K; for ETM+ image band 6, K1 = 666.09 W m−2 sr−1 m−1 , K2 = 1282.71 K, and b = 1277 K. For the three atmospheric functions (i.e.,ϕ1 , ϕ2 , and ϕ3 ), we calculated them by using the atmospheric water vapor content (w), which was obtained from Total Precipitable Water product of MODIS (MOD 05) because it was required at the time near to TM and ETM+ images. The calculation of atmospheric functions are shown as follows:
⎡
ϕ1
⎤
⎡
⎢ ⎥ ⎣ ϕ2 ⎦ = Cij
3×3
w2
⎤
⎣w ⎦
(5)
1
ϕ3
where the Cij
3×3
is the coefficients matrix. Its elements were
obtained from TIRG 61 atmospheric sounding databases in this study. Finally, the land surface emissivity () was calculated by using vegetation cover (Pv ) based on the following formula: NDVI =
Pv =
nir − red nir + red
⎧ ⎪ ⎪ ⎪ ⎨
0
NDVI − NDVIs 2
NDVIv − NDVIs ⎪ ⎪ ⎪ ⎩ 1
(6) NDVI < NDVIs NDVIs ≤ NDVI ≤ NDVIv NDVI > NDVIv
= s (1 − Pv ) + v Pv 2. Methodology In order to build a new index for estimating burn severity, we firstly calculated Land surface temperature (LST) and enhanced vegetation index (EVI) in pre- and post-forest fires. Then, a new index for estimating burn severity (i.e., deltaLST/EVI) was proposed by combining LST and EVI together. To test the effectiveness of the proposed approach, an independent validation between deltaLST/EVI and CBI field data was conducted and compared with several other indices. Finally, we classified the burn severity levels by using the deltaLST/EVI-based nonlinear regression model and assessed their accuracy using the confusion matrix. The primary operations of the methodology are shown in Fig. 1.
2.1. Calculation of land surface temperature and vegetation index In this study, land surface temperature (LST) was calculated based on the thermal infrared band of TM/ETM+ images and the generalized single-channel method, described by Quintano et al. ˜ and Sobrino (2003) and on (2015) in turn based on Jiménez-Munoz ˜ et al. (2009). The general equation is as follows: Jiménez-Munoz LST = ␥
1
Tat−sensor =
(ϕ1 Lat−sensor + ϕ2 ) + ϕ3 + ı
ln
K2 K1
Lat−sensor
+1
(1)
(2)
(7)
(8)
where nir and red refer to the reflectance of near-infrared and red spectral bands; s and v refer to soil and vegetation emissivity, which are assumed to be of 0.97 and 0.99, respectively; NDVIs and NDVIv are the NDVI of soil and vegetation, which were visually extracted from the NDVI histogram of each study area. Since the EVI is suitable for monitoring the vegetation characteristics across a variety of vegetation types (Rocha and Shaver, 2009), we selected it to indicate the vegetation characteristics and calculated the pre-fire EVI (EVIpre−fire ) and post-fire EVI (EVIpost−fire ) values of burned areas using pre- and post-fire TM/ETM+ images. The EVI was calculated as follows: EVI = 2.5 ×
nir − red nir + 6red − 7.5bule + 1
(9)
where nir , red , bule , and swir refer to the reflectance of nearinfrared-, red spectral-, and blue spectral bands in TM/ETM+ images, respectively. 2.2. A new burn severity index based on LST and EVI (deltaLST/EVI) After fire event, land surface biophysical parameters, such as land surface temperature and vegetation cover, were changed obviously and had been individually used to indicate burn severity of forest fires. After considering the variation characteristics of biophysical parameters in burned areas, a new burn severity index (deltaLST/EVI) was proposed by combining pre- and post-land surface temperature and enhanced vegetation index, which would
86
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
Fig. 1. Flowchart of methodology in this study.
enhance the differences in burn severity levels. The deltaLST/EVI was calculated by using the following general equation:
delta
LSTpre−fire LSTpost−fire LST − = EVI EVIpost−fire EVIpre−fire
deltaEVI, post-LST, deltaLST, and RatioLST/EVI ). Detailed calculations of these indices are shown as follows:
deltaNBR = NBRpre−fire − NBRpost−fire × 1000 (10)
RdNBR =
deltaNBR
abs NBRpre−fire In this formula, LST value has already been transformed into interval value ranging from 0 to 1, to make sure that it has the same magnitude order as vegetation index.
2.3. Validation procedure 2.3.1. The effectiveness of deltaLST/EVI for predicting continuous burn severity In order to test the effectiveness of the deltaLST/EVI for predicting the continuous burn severity, an independent validation experiment was conducted in this study. To be specific, the field measured CBI data for each fire were randomly divided into two groups: 70% for training the deltaLST/EVI-based nonlinear regression model and 30% for testing its prediction of continuous CBI values. For testing group of each fire, the Pearson Correlation Coefficient (R) and the Root Mean Squared Error (RMSE) were determined and compared with those of other burn severity indices-based nonlinear regression model (i.e., deltaNBR, RdNBR, RNBR, post-EVI,
RNBR =
deltaNBR
NBRpre−fire + 1.001
(11) (12)
(13)
post − EVI = EVIpost−fire
(14)
deltaEVI = EVIpre−fire − EVIpost−fire
(15)
post − LST = LSTpost−fire
(16)
deltaLST = LSTpost−fire − LSTpre−fire
(17)
RatioLST/EVI =
LSTpost−fire EVIpost−fire
(18)
2.3.2. The accuracy assessment of burn severity classification In this study, the Confusion Matrix was used to evaluate the deltaLST/EVI’s performance in differentiating the burn severity levels. Specifically, with deltaLST/EVI data as the input for above deltaLST/EVI-based nonlinear regression model, CBI values over the burned areas of each fire could be computed. Based on common CBI breakpoints (i.e., 0.1, 1.25, and 2.25), the burn severity can thus
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
87
Fig. 2. Map of the study area showing locations of the fire used in this study.
been classified into four distinct categories, such as: Unchanged: CBI ≤ 0.1, Low: 0.1 < CBI ≤ 1.25, Moderate, 1.25 < CBI ≤ 2.25, and High: 2.25 > CBI. Then, using all field measured CBI data and their burn severity classification results (Miller and Thode 2007; Parks et al., 2014), the Confusion Matrix was calculated and the related statistics (e.g., Overall Accuracy, Kappa coefficient, and the user’s and producer’s accuracy of each category) were obtained. And, these statistics were compared with those of the Monitoring Trends in Burn Severity (MTBS) classification. 3. Materials 3.1. Study area Forest fires occurred more frequently in the Western United States, which is a typical region for forest fire studies (Dillon et al., 2011; Parks et al., 2014). This region includes four typical regional eco-type areas, including Southwest, Central, Northern Rockies, and California of North American continent (Zhu et al., 2006). Five fires occurred between 08/24/2000 and 10/03/2002 in this region were selected for analysis in this study (Fig. 2), since they are representative of areas of different eco-types and each of them has sufficient field-observed burn severity data. The Bear Fire started on 06/27/2002 and ended on 07/04/02. The burned area was about 18.62 km2 . It is located at the Dinosaur National Monument in the Southwestern United States which
belongs to the Highland (alpine) climate. The annual precipitation is lower than 30 cm. The elevations range from 1700 to 2740 m (Perryman et al., 2002), which plays important roles in the distribution of vegetation. At the higher elevation regions, the predominant vegetation are ponderosa pine and Douglas-fir (Perryman et al., 2002). In the lower elevation regions, the major vegetation include willows, boxelder, narrowleaf cottonwood, big sagebrush, rubber rabbit brush, service berry, black grease wood, and so on (Perryman et al., 2002). The pinyon pine, juniper wood-land and Mormon tea dominate at the rocky canyon slopes (Perryman et al., 2002). The Jasper Fire selected as the second fire for analysis, started on 24 August 2000 and got under control a month later on 25 September 2000 (Chen et al., 2011). It was located in the Black Hills National Forest in the western South Dakota and the northeastern Wyoming, which belongs to the Semiarid Steppe climate. The average precipitation is between 44 and 67.3 cm from northwest to southeast (Brown and Sieg, 1999). The elevations range from 1050 to 2207 m. The ponderosa pine covers about 95% of Jasper Fire, following by the white spruce, Limber pine, lodgepole pine, and Rocky Mountain juniper (Brown and Sieg, 1999; Smith et al., 2007). The Mule Fire started by a lightning strike on 07/11/2002. It is situated in the Bridger-Teton National Forest of the Northern Rockies region. The climate belongs to the Highland (alpine). The mean annual precipitation is 75–115 cm. The forest areas are located between 1981 and 3353 m above sea level (Berg et al., 2012). The main forests species are subalpine fir, Engelmann spruce, lodgepole
88
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
Table 1 Summary of remote sensing images and field data collection for fires used. Regional eco-types Fire name Southwest Central Northern Rockies California
Bear Jasper Mule Pw03-Wolf
Alarm date
Pre-fire image date Post-fire image date Landsat path/row Sensors
CBI collection date (MM/DD/YYYY)
06/27/2002 08/24/2000 07/11/2002 10/03/200207/11/2002
06/13/2002 05/01/2000 09/21/2001 10/18/2001
05/28-05/30/2003 05/14-05/26/2002 09/08-09/11/2003 08/15-10/30/2003
05/31/2003 05/31/2002 09/27/2003 10/16/2003
36/32 33/30 37/30 42/34
TM TM and ETM+ TM ETM+ and TM
Table 2 Comparisons of LST calculation results over burned area: Landsat and MODIS. Fire name
Pre-fire
Post-fire
Landsat
Bear Jasper Mule Pw03-Wolf
MODIS
Landsat
MODIS
Min
Max
Mean
Min
Max
Mean
Min
Max
Mean
Min
Max
Mean
24.98 14.88 21.45 14.17
51.28 34.20 23.85 32.82
39.50 23.71 22.46 19.48
34.55 18.47 14.63 18.01
40.03 24.93 35.40 21.68
37.46 21.48 23.69 19.35
29.75 23.19 15.12 14.62
57.32 60.83 39.94 31.83
42.89 38.18 24.87 18.76
36.71 28.73 22.15 18.89
42.63 40.93 23.87 21.21
40.74 36.15 23.10 19.97
pine, whitebark pine, Douglas-fir and limber pineand aspen (Berg et al., 2012). The subalpine fir, Engelmann spruce and aspen are dominant ones in moist areas, while lodgepole pine forests dominate the dry areas (Berg et al., 2012). The whitebark pine forests can be found to form pure stands in the highest elevation regions (Berg et al., 2012). The remaining two fires, named the Pw03 Fire and Wolf Fire, are all situated in the Yosemite National Park of California region. The elevation ranges from 657 to 3997 m. The climate is Mediterranean, where mean monthly temperatures range from −3 to 32 ◦ C (Guarín and Taylor, 2005) and annual precipitation is between 80.4 cm (in lower elevations) and 172.2 cm (in higher elevations) (Lutz et al., 2009). This area is mainly covered by the Chaparral and Sierra Mixed Conifer (Zhu et al., 2006), which is mainly comprised of six conifer species: ponderosa pine, sugar pine, incense cedar, white fir, Jeffrey pine and Douglas-fir (Guarín and Taylor, 2005). In general, the P. ponderosa and the C. decurrens dominate the south-facing stands. At the north-facing sites, the predominant species are A. concolor, C. decurrens, and P. lambertiana (Guarín and Taylor, 2005). The sub-canopy hardwoods black oak and canyon live oak appears at the understory of most stands (Guarín and Taylor, 2005). Since the locations of these two fires are close to each other (i.e., 07/11-10/03/2002) and the timing of field data collection and fire broke out are near (i.e., 08/15-10/30/2003), we analyzed them together in our study (i.e., Pw03–Wolf Fire). 3.2. Field data In order to evaluate the burn severity of forest fires in the field, 262 field plots in five burned areas were surveyed through a Joint Fire Sciences Program (JFSP) project undertaken by the National Park Service and the US Geological Survey (Zhu et al., 2006). In each field, a standard metric, namely Composite Burn Index (CBI), was recorded and calculated independently during the following growing season after fire for extended assessment (Key and Benson, 2006). A more detailed description of approach and methodology for CBI collecting can be found in (Key and Benson, 2006). For Jasper fire, some of CBI field plots (8 field plots) were removed due to misregistration errors, logging influences and creek channel issues (Chen et al., 2011). In addition, we randomly selected 10CBI data outside the burn perimeter of the Jasper fire and assumed their CBI values were 0s, since there is no unchanged level CBI field data for experimental analysis in this fire. As a result, there were 55 field plots for the Bear Fire, 76 field plots for the Jasper Fire, 55 field plots for the Mule Fire, and 78 field plots for the Pw03–Wolf Fire, respectively, which were analyzed in this study.
3.3. Remotely sensed data The remotely sensed data used in this study area were acquired by three types of satellite sensors: Landsat 5 TM, Landsat 7 ETM+, and Terra MODIS (see Table 1 for more details). In all study areas, the post-fire TM and ETM+ images were selected to ensure that their acquisition time approximately matches the time of field data collection. Further, pre-fire TM and ETM+ images were selected to reduce the difference of different moisture contents and phenology between pre- and post-fire images (Key and Benson, 2006). All images were downloaded from the U.S. Geological Survey EarthExplorer (USGS, 2013a). These images have Level L1G of processing and their thermal infrared bands have already been resampled to 30-meter pixels (USGS, 2013b). Consequently, other processing procedures for reflective bands of these images were needed according to suggestions in Key and Benson (2006). These processing procedures, including the radiance transformation, the atmospheric correction, the reflectance transformation and the spectral index calculation, were conducted using ENVI 5.0 software. Then, the CBI field data was co-registered to Landsat images and the spectral attributions of remote sensing images were extracted by using 3 × 3 pixel matrix in order to reduce CBI plots to image registration error caused by GPS positioning (Chuvieco et al., 2004). In this study, we did not need to consider the scan line corrector (SLC)’s effect on ETM+ images processing, since these images were acquired before May 2003. Additionally, two MODIS products, including the Total Precipitable Water (MOD 05) and the Daily Land Surface Temperature/Emissivity (MOD11A1), were taken from the NASA’s Level 1 and Atmosphere Archive and Distribution System (LAADS Web) on the same day when TM and ETM+ images were acquired (NASA). They are used to represent the atmospheric water vapor content for LST calculation (Jiménez˜ ˜ et al., 2009; Jiménez-Munoz and Sobrino, 2003) and to Munoz validate the LST calculation results of Landsat data, respectively. Finally, each fire’s perimeter layer used as burned mask and its official burn severity classification used for comparative analysis were obtained from the Monitoring Trends in Burn Severity (MTBS, 2013). 4. Results 4.1. Pre- and post-fire LST and EVI Table 2 indicates that, for burned areas of most fires, the LST mean value derived from Landsat data shows a good agreement with that of MOD11A1 product (all mean value differences are close
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
89
Fig. 3. The spatial distributions of pre-EVI, post-EVI, pre-LST (◦ C), and post-LST (◦ C): (a) Bear fire; (b) Jasper fire; (c) Mule fire and (d) Pw03-Wolf fire.
to ±2 ◦ C) (Srivastava et al., 2009). Also, we have checked that the LST histogram differences between the Landsat calculation and MODIS product are negligible. Thus, the accuracy of LST calculations in this study is acceptable. Fig. 3 displays the spatial distributions of pre-LST, post-LST, pre-EVI, and post-EVI across four typical regional eco-type areas of the Western United States. These spectral indices reveal that forest fire has caused a great change to the coverage of vegetation and to the distribution of LST. Among these study areas, vegetation in the Mule and Pw03-Wolf fires experienced the lower degree of damage while that in the Bear and Jasper fires were seriously destroyed. On the contrary, the max., min., and mean values of deltaLST (see Table 3) in the Bear and Jasper fires are higher and those in the Mule and Pw03-Wolf fires are lower, which agreed with official sources (MTBS, 2013) and statistics in Table 3. Furthermore, statistics in Table 3 further confirm that the Mule and Pw03-Wolf fires were rich with vegetation after the fire happened (post-EVI mean value are 0.25 and 0.15 respectively). On the con-
trary, Table 3 illustrates that the post-fire areas of the Bear and Jasper fires were sparsely covered with vegetation (post-EVI mean value are 0.12 and 0.09, respectively). 4.2. The independent validation results of deltaLST/EVI Table 4 shows the results of the independent validation using the observed CBI and different burn severity indices of testing groups. The R and RMSE values in this table indicate that deltaLST/EVI performed relatively better than some commonlyused burn severity indices (e.g., dNBR, RdNBR, and RNBR). Further, it can be observed in Table 4 that the post-EVI and deltaEVI would have the better performance if the forest fire’s post-fire areas were rich with vegetation (i.e., Mule and Pw03–Wolf fires). In contrast, for forest fire whose post-fire areas were sparsely covered with vegetation (i.e., Bear and Jasper fires), the post-LST and deltaLST would performed better. Unlike them, Table 4 indicate the deltaLST/EVI
90
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
Table 3 Statistics of burn severity indices over the burned area of each fire. Fire name
Burn severity indices
Min
Max
Mean
Bear
post-EVI deltaEVI post-LST deltaLST RatioLST/EVI deltaLST/EVI post-EVI deltaEVI post-LST deltaLST RatioLST/EVI deltaLST/EVI post-EVI deltaEVI post-LST deltaLST RatioLST/EVI deltaLST/EVI post-EVI deltaEVI post-LST deltaLST RatioLST/EVI deltaLST/EVI
−0.03 −0.12 29.75 0.00 −63.74 −52.82 0.02 −0.47 23.19 2.94 0.27 −48.21 0.02 -0.19 15.12 −3.61 0.44 −5.24 0.03 −0.13 14.63 −4.81 0.31 −1.31
0.35 0.21 57.32 20.91 68.30 1416.84 0.29 0.32 60.83 40.08 16.61 15.78 0.37 0.26 39.94 15.65 29.04 27.92 0.69 0.45 31.83 6.66 14.64 13.83
0.12 0.05 42.89 9.39 6.33 3.17 0.09 0.06 38.18 18.47 2.92 0.96 0.15 0.09 24.87 5.18 4.61 2.75 0.25 0.08 18.76 −0.72 1.39 0.38
Jasper
Mule
Pw03-Wolf
performed stably and steadily in predicting the continuous burn severity across regional forest areas of different eco-types after fire. 4.3. Accuracy of mapping burn severity classification The deltaLST/EVI spatial distribution of each fire is presented in Fig. 4(A-1), (B-1), (C-1), and (D-1). And, the burn severity levels over burned areas are classified and displayed in Fig. 4(A-2), (B-2), (C-2), and (D-2). We can visually notice in these figures that the spatial distribution pattern of the deltaLST/EVI, to some extent, was similar to that of its burn severity classification results. Further, burn severity classification statistics in Table 5 show that the Overall Accuracy (OA) and Kappa coefficient are 66.29% and 0.521, respectively, which are much better than those of the MTBS classification (see Table 6: the OA is 43.56% and the Kappa Coefficient is 0.249) and are consistent with those of other study areas (e.g., OA for four classes: 52.9–71.1% (Epting et al., 2005), OA: 58.7–59.9% (Miller and Thode, 2007), OA: 59.7–70.2% (Soverel et al., 2010), and OA: 64.2–66.2% (Parks et al., 2014). We therefore confirm a fact that deltaLST/EVI performs well at classifying the burn severity levels of forest fires. 5. Discussion This study proposed a new index (i.e., deltaLST/EVI) for estimating burn severity of forest fires based on surface biophysical
parameters (i.e., EVI and LST). The results indicated that the deltaLST/EVI performed relatively better than some commonlyused burn severity indices, and its performance was stable and steady across regional forest areas of different eco-type after fires. Our findings have further developed the LST-based index and provided a new insight for estimating burn severity of forest fires. The LST has already been shown to be a useful indicator for burn severity estimation in forest Mediterranean ecosystems (Quintano et al., 2015). In order to confirm its usefulness in the forest ecosystems of western United States, the two LST-based indices (i.e., post-LST and deltaLST) were applied in this study to estimate the burn severity of forest fires. Unlike previous study (Quintano et al., 2015), the remote sensing images for post-fire LST calculation were selected to ensure that their acquisition time approximately matches the time of field data collection. Although this might not be optimal for LST-based indices performance, the experimental results in Table 4 show that the post-LST and deltaLST’s accuracies of burn severity estimation are acceptable in most cases. In particular, for severe forest fire whose post-fire areas were sparsely covered with vegetation (i.e., Bear and Jasper fires), the post-LST and deltaLST performed better even after more than 1 year. The reason is that the severe damage and the low vegetation regeneration can ensure that the obvious LST variation characteristics remained. On the contrary, for Mule and Pw03-Wolf fires, the less severe fire damage and the good vegetation regeneration after fire leads to the fading of LST variation, and thus both post-LST and deltaLST have poorer performance. This limitation of LST-based indices might limit their widespread applicability. As another surface biophysical parameter, the EVI can also be used to effectively indicate the environmental damage levels caused by fires. Analysis results in Table 4 indicate that post-EVI and deltaEVI have greater accuracy in estimating burn severity for forest fires whose post-fire areas are rich with plant life (i.e., Mule and Pw03–Wolf fires). These results echo the previous study results (Chen et al., 2011) and can be explained by the fact that EVI-based indices are sensitive to monitoring the vegetation variation characteristics across a variety of vegetation types (Rocha and Shaver, 2009). This merit is very useful for addressing the above limitation of the LST-based index. As expected, Table 4 and Fig. 5 show that the combination of LST- and EVI-based indices (i.e., RatioLST/EVI and deltaLST/EVI) have a better performance and they can make full use of burn severity information contained in surface biophysical parameters (i.e., LST and EVI). To be specific, taking the deltaLST/EVI for the Wolf fire as an example, we can find in Fig. 5(F) that the Plot WM12 location was burned severely and its true CBI value is 2.77. Since this location is covered with sparse and open vegetation after fires, it would be incorrectly classified as the low-level burn severity using the lower deltaEVI value in Fig. 5(A). Taking advantage of deltaLST, the deltaLST/EVI value in Fig. 5(C) will enable this burned location (i.e., Plot WM12) to be correctly classified as high-level severity, as
Table 4 Results of independent validation using the observed CBI and different burn severity indices. Burn severity indices
deltaNBR RdNBR RNBR post-EVI deltaEVI post-LST deltaLST RatioLST/EVI deltaLST/EVI
Bear
Jasper
Mule
Pw03–Wolf
R
RMSE
R
RMSE
R
RMSE
R
RMSE
0.71 0.63 0.72 0.71 0.59 0.77 0.66 0.72 0.81
0.69 0.77 0.68 0.74 0.80 0.83 0.76 0.56 0.53
0.70 0.74 0.70 0.70 0.72 0.73 0.61 0.72 0.79
0.69 0.66 0.69 0.67 0.65 0.67 0.74 0.65 0.59
0.86 0.87 0.88 0.82 0.82 0.64 0.69 0.87 0.91
0.57 0.56 0.54 0.64 0.63 0.88 0.80 0.57 0.39
0.69 0.70 0.73 0.44 0.73 0.29 0.60 0.58 0.74
0.64 0.59 0.57 0.74 0.55 0.79 0.67 0.67 0.56
Note: The nonlinear regression model is CBI = a × (x)2 + b × x + c and the accuracy is measured by the R and RMSE of testing group (all P values <0.05).
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
91
Fig. 4. The spatial distributions of continuous deltaLST/EVI (1) and burn severity mapping results (2), including: (a) Bear fire, (b) Jasper fire, (c) Mule fire, and (d) Pw03-Wolf fire.
92
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
Table 5 Overall classification statistics of deltaLST/EVI when classifying all field plots into distinct categories. Class name
Unchanged
Low
Moderate
High
Total
User’s accuracy (%)
Unchanged Low Moderate High Total Producer’s accuracy(%) Overall accuracy:
5 12 14 0 31 16.13 66.29%
0 38 18 0 56 67.86
0 17 56 9 82 68.29 Kappa coefficient:
1 4 14 76 95 80.00 0.521
6 71 102 85 264
83.33 53.52 54.90 89.41
Table 6 The validation results of MTBS classification. Class name
Unchanged
Low
Moderate
High
Total
User’s accuracy (%)
Unchanged Low Moderate High Total Producer’s accuracy(%) Overall accuracy:
14 11 6 0 31 45.16 43.56%
25 28 3 0 56 50.00
15 42 17 8 82 20.73 Kappa coefficient:
3 6 30 56 95 58.95 0.249
57 87 56 64 264
24.56 32.18 30.36 87.50
illustrated in Fig. 5(D). On the other hand, for the severely burned location of Plot PH09 which was covered with rich and dense vegetation (see Fig. 5(E)), it might be underestimated as low-level
severity by using the lower deltaLST value in Fig. 5(B). However, as shown in Fig. 5(C), the deltaLST/EVI can correctly estimate Plot PH09, since the burn severity information contained in deltaEVI
Fig. 5. Pw03-Wolf fire: (a) continuous deltaEVI, (b) continuous deltaLST (◦ C), (c) continuous deltaEVI/EVI, (d) burn severity classification results of deltaEVI/EVI, (e) post-firescene photo of Plot PH09 and (f) post-fire-scene photo of Plot WM12 (Key et al., 2013).
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
have been comprehensively utilized. This fact is evident in other places of the Pw03–Wolf fire and other fires in this study. We therefore can conclude that these two indicators (i.e., EVI and LST) not only relate to one another, but also complement to each other. When we combine them together, the comprehensive spatial distributions of burn severity can be estimated, and a new metric across regional areas of different eco-types, can be established and proved useful. Furthermore, the independent validation results in Table 4 indicate that the deltaLST/EVI can accurately predict the continuous burn severity with high application adaptability. And, its performance is relatively better than that of some commonly-used burn severity indices (e.g., deltaNBR, RdNBR and RNBR), which does not mean that there is no need to further develop such LST-based index. Instead, we believe that it has high potential of burn severity estimation, if more accurate surface temperature can be obtained in the future. Therefore, further research should be conducted to test the effectiveness of deltaLST/EVI calculated from the new remotely sensed data (e.g., the Landsat 8 thermal infrared data (TIRS)). Although the overall classification results in Tables 5 and 6 have indicated that deltaLST/EVI performs well at classifying most burn severity levels of forest fires, we acknowledge that the producer accuracy of unchanged level is lower (its Producer’s Accuracy is 16.13%). The reason which might explain this is that the limited number of unchanged level field data is not enough for the regression model training and the true evaluation of deltaLST/EVI performance (more than 50 verification points for per class were commonly used (Epting et al., 2005; Underwood et al., 2003)). Since the merit of LST is more noticeable in post-fire areas covered with sparse and open vegetation, the LST-based index has been developed (Quintano et al., 2015). However, this merit might be diminished if the post-fire land surface is covered with rich and dense vegetation, which might restrict the application of LST-based index. Our results in this study present that, the combination of LST and EVI does improve the burn severity estimation performance of LST-based index. 6. Conclusion Compared to commonly used indicators for estimating the burn severity of forest fires, other potential indicators (i.e., LST) are also attracting more and more attention of scientists. In this study, a new metric based on EVI and LST was constructed (i.e., deltaLST/EVI), which was validated by CBI field data of five forest fires located in the Western United States. The results show that deltaLST/EVI can take the advantages of surface biophysical parameters (i.e., EVI and LST) and can fully reveal the spatial distributions of burn severity, when the composition and characteristics of vegetation are different in the study areas. And, its performance was relatively better than that of some commonly-used burn severity indices. It also has high potential of estimating burn severity if more accurate surface temperatures can be obtained in the future. This index will be useful to environmental scientists, forest fire researchers, and policy makers. Acknowledgments The research work was supported by the National Natural Science Foundation of China (No. 41171326, 40771198) and the Fundamental Research Funds for the Central Universities of Central South University (No. 2015zzts070). The authors would like to acknowledge the data made available by the U.S. Geological Survey, the U.S. Geological Survey’s National Center for Earth Resources Observation and Science, the USDA Forest Service Remote Sensing Applications Center for making remote sensing data available, and
93
the Joint Fire Sciences Program (JFSP) project undertaken by the National Park Service and the US Geological Survey. We wish to thank the anonymous reviewers for their thoughtful and helpful comments.
References Berg, N.D., Gese, E.M., Squires, J.R., Aubry, L.M., 2012. Influence of forest structure on the abundance of snowshoe hares in western Wyoming. J. Wildl. Manage. 76, 1480–1488. Brown, P.M., Sieg, C.H., 1999. Historical variability in fire at the ponderosa pine-Northern Great Plains prairie ecotone, southeastern Black Hills, South Dakota. Ecoscience 6, 539–547. Chen, X., Vogelmann, J.E., Rollins, M., Ohlen, D., Key, C.H., Yang, L., Huang, C., Shi, H., 2011. Detecting post-fire burn severity and vegetation recovery using multitemporal remote sensing spectral indices and field-collected composite burn index data in a ponderosa pine forest. Int. J. Remote Sens. 32, 7905–7927. Chuvieco, E., Cocero, D., Riano, D., Martin, P., Martınez-Vega, J., de la Riva, J., Pérez, F., 2004. Combining NDVI and surface temperature for the estimation of live fuel moisture content in forest fire danger rating. Remote Sens. Environ. 92, 322–331. De Santis, A., Asner, G.P., Vaughan, P.J., Knapp, D.E., 2010. Mapping burn severity and burning efficiency in California using simulation models and Landsat imagery. Remote Sens. Environ. 114, 1535–1545. De Santis, A., Chuvieco, E., 2007. Burn severity estimation from remotely sensed data: performance of simulation versus empirical models. Remote Sens. Environ. 108, 422–435. Dillon, G.K., Holden, Z.A., Morgan, P., Crimmins, M.A., Heyerdahl, E.K., Luce, C.H., 2011. Both topography and climate affected forest and woodland burn severity in two regions of the western US, 1984 to 2006. Ecosphere 2, art130. 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 Sens. Environ. 96, 328–339. Escuin, S., Navarro, R., Fernandez, P., 2008. Fire severity assessment by using NBR (Normalized Burn Ratio) and NDVI (Normalized Difference Vegetation Index) derived from LANDSAT TM/ETM images. Int. J. Remote Sens. 29, 1053–1073. Guarín, A., Taylor, A.H., 2005. Drought triggered tree mortality in mixed conifer forests in Yosemite National Park, California, USA. For. Ecol. Manage. 218, 229–244. Hall, R.J., Freeburn, J., De Groot, W., Pritchard, J., Lynham, T., Landry, R., 2008. Remote sensing of burn severity: experience from western Canada boreal fires. Int. J. Wildland Fire 17, 476–489. ˜ Jiménez-Munoz, J.C., Cristóbal, J., Sobrino, J.A., Sòria, G., Ninyerola, M., Pons, X., 2009. Revision of the single-channel algorithm for land surface temperature retrieval from Landsat thermal-infrared data. IEEE Trans. Geosci. Remote Sens. 47, 339–349. ˜ Jiménez-Munoz, J.C., Sobrino, J.A., 2003. A generalized single(c)\channel method for retrieving land surface temperature from remote sensing data. J. Geophys. Res. 108, 4688–4695. Keeley, J.E., 2009. Fire intensity, fire severity and burn severity: a brief review and suggested usage. Int. J. Wildland Fire 18, 116–126. Key, C.H., Benson, N.C., 1999. The Normalized Burn Ratio (NBR): A Landsat TM radiometric measure of burn severity. http://www.nrmsc.usgs.gov/research/ ndbr.htm. Key, C.H., Benson, N.C., 2006. Landscape Assessment (LA): Sampling and Analysis Methods. 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 (pp. LA-1-LA-51). Fort Collins, CO, USA. Rocky Mountain Research Station,USDA, Forest Service, Gen. Tech. Rep. RMRS-GTR-164-CD. Key, C.H., Benson, N.C., Suzanna, S., 2013. CBI Plot Data and Photos. USGS-NOROCK and NPS. http://www.nrmsc.usgs.gov/science/fire/cbi/plotdata (20.05.14). Loboda, T.V., French, N.H.F., Hight-Harf, C., Jenkins, L., Miller, M.E., 2013. Mapping fire extent and burn severity in Alaskan tussock tundra: an analysis of the spectral response of tundra vegetation to wildland fire. Remote Sens. Environ. 134, 194–209. Lutz, J.A., Van Wagtendonk, J.W., Franklin, J.F., 2009. Twentieth-century decline of large-diameter trees in Yosemite National Park, California, USA. For. Ecol. Manage. 257, 2296–2307. Miller, J.D., Thode, A.E., 2007. Quantifying burn severity in a heterogeneous landscape with a relative version of the delta Normalized Burn Ratio (dNBR). Remote Sens. Environ. 109, 66–80. MTBS, 2013. Monitoring trends in burn severity. http://mtbs.gov/index.html. NASA, Level 1 and Atmosphere Archive and Distribution System (LAADS). http:// earthexplorer.usgs.gov/ (20.10.14). Parks, S., Dillon, G., Miller, C., 2014. A New Metric for Quantifying Burn Severity: The Relativized Burn Ratio. Remote Sens. 6, 1827–1844. Perryman, B.L., Olson, R.A., Petersburg, S., Naumann, T., 2002. Vegetation response to prescribed fire in Dinosaur National Monument. West. North Am. Nat., 414–422. Quintano, C., Fernández-Manso, A., Calvo, L., Marcos, E., Valbuena, L., 2015. Land surface temperature as potential indicator of burn severity in forest Mediterranean ecosystems. Int. J. Appl. Earth Obs. Geoinf. 36, 1–12.
94
Z. Zheng et al. / International Journal of Applied Earth Observation and Geoinformation 45 (2016) 84–94
Quintano, C., Fernández-Manso, A., Roberts, D.A., 2013. Multiple Endmember Spectral Mixture Analysis (MESMA) to map burn severity levels from Landsat images in Mediterranean countries. Remote Sens. Environ 136, 76–88. Rocha, A.V., Shaver, G.R., 2009. Advantages of a two band EVI calculated from solar and photosynthetically active radiation fluxes. Agric. For. Meteorol. 149, 1560–1563. Smith, A.M.S., Lentile, L.B., Hudak, A.T., Morgan, P., 2007. Evaluation of linear spectral unmixing and (NBR for predicting post(c)\fire recovery in a North American ponderosa pine forest. Int. J. Remote Sens. 28, 5159–5166. Soverel, N.O., Perrakis, D.D.B., Coops, N.C., 2010. Estimating burn severity from Landsat dNBR and RdNBR indices across western Canada. Remote Sens. Environ. 114, 1896–1909. Srivastava, P.K., Majumdar, T.J., Bhattacharya, A.K., 2009. Surface temperature estimation in Singhbhum Shear Zone of India using Landsat-7 ETM+ thermal infrared data. Adv. Space Res 43, 1563–1574. Underwood, E., Ustin, S., DiPietro, D., 2003. Mapping nonnative plants using hyperspectral imagery. Remote Sens. Environ. 86, 150–161. USGS, 2013a. Earthexplorer. http://earthexplorer.usgs.gov/ (20.12.13). USGS, 2013b. Frequently asked questions about the Landsat Missions. http:// landsat.usgs.gov/band designations landsat satellites. php (12.06.14).
van Wagtendonk, J.W., Root, R.R., Key, C.H., 2004. Comparison of AVIRIS and Landsat ETM+ detection capabilities for burn severity. Remote Sens. Environ. 92, 397–408. Vasconcelos, S.S.d., Fearnside, P.M., Grac¸a, P.M.L.d.A., Dias, D.V., Correia, F.W.S., 2013. Variability of vegetation fires with rain and deforestation in Brazil’s state of Amazonas. Remote Sens. Environ. 136, 199–209. Veraverbeke, S., Hook, S., Hulley, G., 2012a. An alternative spectral index for rapid fire severity assessments. Remote Sens. Environ. 123, 72–80. Veraverbeke, S., Verstraeten, W.W., Lhermitte, S., Van De Kerchove, R., Goossens, R., 2012b. Assessment of post-fire changes in land surface temperature and surface albedo, and their relation with fire-burn severity using multitemporal MODIS imagery. Int. J. Wildland Fire 21, 243–256. Vlassova, L., Pérez-Cabello, F., Mimbrero, M., Llovería, R., García-Martín, A., 2014. Analysis of the relationship between Land Surface Temperature and Wildfire Severity in a series of Landsat Images. Remote Sens 6, 6136–6162. Zhu, Z., Key, C., Ohlen, D., Benson, N.C., 2006. Evaluate sensitivities of burn-severity mapping algorithms for different ecosystems and fire histories in the United States. In. Joint Fire Science Program: Boise, IA, USA, 2006. Final Report to the Joint Fire Science Program, JFSP 01-1-4-12.