Abstract

Estimation of fire impacts and forest recovery using remote sensing is difficult because of the heterogeneity of fire history (frequency, severity, and time since last fire) across burned forest landscapes. The authors analyzed impacts of fire frequency and severity within recovering forests in the Amazon region using remote sensing. A multispectral Landsat time series dataset was used to reconstruct the fire history from 1990 to 2002 in a portion of Mato Grosso, Brazil. Five narrowband vegetation indices were then calculated from a hyperspectral Earth Observing One (EO-1) Hyperion image for spectral analysis of physiological characteristics of fire-disturbed forests and their recovery. A total of 30% of the forests burned during the study period, with 72% burned once, 24% burned twice, and less than 4% burned three times. In terms of severity, 70% of burned forest was lightly burned, 21.1% was moderately burned, and 9.1% was severely burned. Analyses of spectral indices [normalized difference vegetation index (NDVI), carotenoid reflectance index (CRI), and photochemical reflectance index (PRI)] showed that those related to canopy greenness and pigment contents can discriminate between burned forests and undisturbed forest for the first 3 years after forest fire, whereas the effectiveness of canopy water content indices [normalized difference water index (NDWI) and normalized difference infrared index (NDII)] varied from 1 to 3 years, depending on the fire severity. Despite the relatively low signal-to-noise ratios of Hyperion imagery, we show that narrowband-derived indices provide useful information for monitoring degraded forests beyond what is currently possible with Landsat. This illustrates the great potential for environmental monitoring using satellite-borne hyperspectral sensors, such as the Hyperspectral Infrared Imager (HyspIRI), which have better signal-to-noise ratios.

1. Introduction

Forest fire is one of the main forest disturbances that alter biomass, structure, and species composition in the Amazonian tropical forests. Most forest fire occurrences in this region are related to anthropogenic activities, such as escaped fires from forest clearing for agriculture/pasture conversion or associated with selective logging (Cochrane and Schulze 1999; Nepstad et al. 1999). The areas of burned forest expand when combined with prolonged dry weather conditions, such as ENSO events (Alencar et al. 2004; Alencar et al. 2006; Aragao et al. 2008). In some regions in the Brazilian Amazon, forest fire is a recurrent event and fire frequency tends to increase near fragment edges (Cochrane and Laurance 2002). Once a forest is disturbed by fire, the chances of recurrent fires are increased and forest damages in the next fire event are greater because of accumulated fuel loads, higher temperatures, and lower humidity. Different degrees of fire severity diversify forest recovery processes in terms of structure, the amount of remaining biomass, and species richness (Cochrane and Schulze 1999). The detection and characterization of forest disturbance and its recovery processes are key tasks for understanding actual structural damages, biomass, and species composition changes and for measuring changes in ecosystem functioning, especially carbon cycling (NRC 2007; GOFC-GOLD 2008). However, the estimation of forest fire impacts is hampered by the heterogeneity of burned forest landscapes related to the combination of factors such as fire frequency, severity, and time since last fire that change patterns of recovering forests (Frolking et al. 2009). To accomplish this task, we need to track landscape fire disturbance history.

Multispectral satellite measurements [e.g., Thematic Mapper (TM)/Landsat-5, Moderate Resolution Imaging Spectroradiometer (MODIS)/Terra, and Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER)/Terra data] have been widely used to monitor deforestation and forest degradation based on forest canopy cover, leaf area index, fractional covers of different land-cover types, and phenological or greenness changes (Koltunov et al. 2009; Asner et al. 2002; Souza et al. 2005a; Cochrane and Souza 1998). Detection of disturbed forests with current satellite data becomes challenging because of relatively small areas of disturbance, paucity of cloud-free imagery, and rapid forest recovery (Cochrane and Souza 1998; Souza et al. 2005a; Souza et al. 2005b; Matricardi et al. 2010). Furthermore, unlike forest conversion to other land-cover types, such as agriculture, subtle changes of disturbed forests in terms of physiology and structure may not be well characterized by multispectral sensors. Hyperspectral sensors, with several spectral bands between 400 and 2400 nm, provide detailed spectral information associated with biochemical and physiological properties of vegetation (Gamon et al. 1992; Ustin et al. 2004; Asner et al. 2005; Chambers et al. 2007; Kokaly et al. 2009) and have the potential to improve the characterization of burned forests but have never been tested for the Amazon region.

In this study, we use both a multispectral Landsat time series dataset and a hyperspectral Earth Observing One (EO-1) Hyperion image for spectral analysis of physiological characteristics of fire disturbed forests and their recovery through time. First, we reconstruct fire history, with focus on fire frequency, fire severity, and time since last fire, for a forested landscape located in the Brazilian state of Mato Grosso using the Landsat dataset [TM/Landsat-5 and Enhanced Thematic Mapper Plus (ETM+)/Landsat-7 data] for a 1990–2002 time period. Then we use the Hyperion data to characterize the impacts of fire frequency and severity on the recovery of disturbed forest.

2. Study area

The study area is located near the county of Sinop, Mato Grosso, in the southern Brazilian Amazon (Figure 1). The area used for analysis matched the footprint of the Hyperion image, covering 595 km2. It comprises a transitional forest between savanna and dense forest (RADAM BRASIL 1982). The average annual precipitation is 2000 mm. The state of Mato Grosso has experienced the highest deforestation rates in recent decades (INPE 2010; Morton et al. 2006).

Figure 1.

Study area. False color composite of a subset image of Landsat ETM (226/68) with the Hyperion image area highlighted in blue.

Figure 1.

Study area. False color composite of a subset image of Landsat ETM (226/68) with the Hyperion image area highlighted in blue.

Forest fire is frequent in this region, usually associated with selective logging (Souza et al. 2005a; Matricardi et al. 2010). The synergism of selective logging and forest fire increases the extent and severity of forest burning and results in extensive forest degradation (Cochrane and Laurance 2008; Monteiro et al. 2003).

3. Methodology

3.1. Reconstruction of forest fire history from Landsat data processing

A TM/Landsat-5 and ETM+/Landsat-7 time series dataset (path = 226, row = 68) was used to characterize forest fire history in terms of 1) fire frequency, 2) fire severity, and 3) postfire recovery age. Figure 2 shows the data processing stream for this study. The Landsat time series covered a period from 1990 to 2002 (Table 1). All images of the time series were georeferenced to the orthorectified Landsat GeoCover data at the Global Land Cover Facility (available online at http://glcf.umiacs.umd.edu/data/) using the second-degree polynomial algorithm and nearest-neighborhood resampling. The root-mean-square errors varied from 0.62 to 0.84. To make the interannual comparisons of the time series images and to track fire history, we used a relative radiometric calibration technique developed by Roberts et al. (Roberts et al. 1998). This technique intercalibrates images in digital numbers (DNs) to a reference reflectance image based on the same invariant targets and converts DN values of images to reflectance relative to the reference image. We chose a Landsat ETM+ image acquired on 2 August 2001, with no cloud and haze contamination, as our reference image for relative radiometric calibration. DNs of all bands of the reference image were converted into radiance with the gain and offset values from the metafile. Atmospheric correction was done for the radiance image using the fast line-of-sight atmospheric analysis of spectral hypercubes (FLAASH) algorithm (ITT) and a surface reflectance image was generated for the reference scene. Then, temporally invariant targets across the time series such as forest, water, rock, savanna, and bare soil found in both the reference reflectance image and each uncalibrated image were selected. A 5 × 5 pixel window was used and the mean values of these targets were extracted from both the reflectance and uncalibrated images. Linear models were estimated for each pair of reflectance–uncalibrated data. The coefficients of linear regressions were calculated and applied to the DN values of each band of the uncalibrated image. In this way, the DN values of each uncalibrated image were normalized to the 2001 reference image and converted to relative reflectance using the reference data.

Figure 2.

Flowchart of the data processing.

Figure 2.

Flowchart of the data processing.

Table 1.

Landsat time series dataset used in this study (path = 226, row = 68).

Landsat time series dataset used in this study (path = 226, row = 68).
Landsat time series dataset used in this study (path = 226, row = 68).

To detect burned area within each Landsat image, we used the relative delta normalized burned ratio (RdNBR) (Miller and Thode 2007),

 
formula

where NBR is the normalized burn ratio calculated as follows:

 
formula

The RdNBR highlights burned areas in a given year based on a pair of prefire and postfire NBR images (Miller and Thode 2007). Burned forest yields higher RdNBR values compared to unburned forest. However, we also found that those forests damaged by logging show high RdNBR values. To separate burned forests from logged forests, we determined thresholds by assessing the RdNBR values from burned forests and logged forests identified in the field through the forest inventories from Souza et al. (Souza et al. 2005a; Souza et al. 2005b). Souza et al. (Souza et al. 2005a; Souza et al. 2005b) conducted forest transect inventories between 2001 and 2004 within the same region of this study to identify and characterize forests degraded by selective logging and fire. They had 19 forest transects (10 m × 500 m) that included intact forests and different forest degradation classes. For each transect, they estimated ground cover, canopy cover, and above-ground biomass (for more details, see Souza et al. 2005b). There are three transects of forest burned in 1999 in Souza’s inventory data (Souza et al. 2005a). Because these burned forests are detectable only in 2000 by the Landsat imagery, we assessed the RdNBR values of these transects and surrounding burned forests, derived from the 2000 Landsat imagery, to determine the RdNBR range within these burned forests. The RdNBR values of those forests disturbed by logging activities, as identified by Souza et al. (Souza et al. 2005a), were also extracted from the RdNBR values derived from imagery corresponding to the years of the respective logging events. These values were used to determine thresholds for separating burned from logged forests. The RdNBR values for logged forests ranged between 28 and 120. To minimize forest fire misclassification, an RdNBR threshold of 150 was used. All pixels equal or greater than this value were classified as burned area. Log landings, which present high RdNBR values, are confused with burned forests. Because of their small size and characteristic nature, we were able to exclude them from the analysis manually. We focused our analysis on forest fires only; other fire types, such as deforestation fires and pasture maintenance fires were excluded. All nonforest areas, such as pasture/crop, river, rock, and roads, were also masked out from the analysis. The analyses were limited to the forest area existing as of 2002, the year of Hyperion data collection. After detecting burned forests for each year in the time series, a spatial filter was used to remove features with less than 5 pixels.

The classified time series of RdNBR, over the 1990–2002 time period, was used to generate fire frequency (the number of times the RdNBR value of a pixel exceeded the forest burning threshold in the time series) and postfire recovery age (years since last fire) as of 2002. To analyze the effects of fire severity on forest recovery, three classes of fire severity were created based on the RdNBR values of burned forests. Because there is no systematic way for determining fire severity classes from a continuous RdNBR range of values, the thresholds for the classes were defined based on a mean cluster analysis for the RdNBR ranges within burned forests. We defined the following fire severity classes: low severity (S1: threshold between 150 and 278); moderate severity (S2: threshold between 279 and 450); and high severity (S3: threshold between 451 and 778). The total area of burned forest and the areas of each fire severity class were calculated for each year of the time series.

3.2. Spaceborne Hyperion data processing

Launched in 2002, EO-1 Hyperion is the first spaceborne hyperspectral sensor. The sensor has 242 bands in the 400–2500-nm spectral range, but only 198 bands are calibrated. The spatial resolution is 30 m (Pearlman et al. 2003). A Hyperion image for the study area was acquired on 24 June 2002 (Figure 1). Erroneous vertical lines within the Hyperion data were replaced by the average values of adjacent columns to reduce the striping effects, as suggested by Goodenough et al. (Goodenough et al. 2003). Hyperion radiance values were converted to surface reflectance using the FLAASH algorithm (ITT), a Moderate Resolution Atmospheric Transmission (MODTRAN) 4–based approach. A correction for adjacent effects was applied using this algorithm. Hyperion bands covering the 1140-nm water vapor band were used to estimate precipitable water vapor on a per-pixel basis. To remove artifacts from atmospheric correction and to smooth data, inverse minimum noise fraction (MNF) transformations were applied to the Hyperion data. Because of the strong water vapor atmospheric absorption and noise from instrument artifacts, bands around the 1400- and 1900-nm wavelength regions were excluded from analysis and only 150 spectral bands were used in this study. Then, Hyperion data were registered to the same reference image used for registering the Landsat time series dataset by application of a second-degree polynomial algorithm and nearest-neighborhood resampling. The root-mean-square error was 0.78.

Five narrowband vegetation indices (VIs) were generated from the Hyperion data for the analysis of physiological properties of burned forests. Equations and references for the VIs are presented in Table 2. The normalized difference vegetation index (NDVI) is related to a composite of canopy cover, leaf area, and the fraction of photosynthetically active radiation (fPAR) (Myneni et al. 1997). To estimate changes in canopy water content of burned forests, the normalized difference infrared index (NDII) and the normalized difference water index (NDWI) were calculated. We also calculated the photochemical reflectance index (PRI), which is related to photosynthetic light-use efficiency (LUE; Gamon et al. 1992; Gamon et al. 1997) and is a fundamental determinant of net primary production (NPP). Similarly, we derived the carotenoid reflectance index (CRI) that is sensitive to carotenoid pigments in vegetation canopies and indicates plant stress (Gitelson et al. 2002).

Table 2.

Hyperspectral-derived measures for forest stress analysis.

Hyperspectral-derived measures for forest stress analysis.
Hyperspectral-derived measures for forest stress analysis.

3.3. Data analysis

To analyze the impacts of the combined effects of fire frequency and severity, we assessed temporal changes of each vegetation index through recovery ages after forest fire for different classes of severity and fire frequency. For statistical analysis, 200 pixels were randomly sampled from each class. Analysis of variance (ANOVA) was used to test the effects of each variable (fire frequency, severity, and forest recovery age) on the vegetation indices from burned forests. To determine the statistical mean, differences of the vegetation indices between undisturbed and burned forests of different severity classes and recovery ages were made using the Tukey test.

4. Results

4.1. Fire history

The total forested area within the 2002 Hyperion image was 295 km2. The fire history derived from the Landsat time series data of this same area showed that 197 km2 was undisturbed forest, but 98 km2, equivalent to 30% of the total forest area, was burned at least once during the 1991 and 2002 period. The area being burned each year showed an increasing trend over the study period (Figure 3a). In the first 5 years, the average burned forest area was 5 km2, whereas the average burned forest area in the last 5 years expanded to 13.5 km2, reaching 28 km2 in 2000. The overall annual average area burned during the 1991–2002 period was 9.1 km2.

Figure 3.

(a) Areas of burned forest over the study period. (b) Distribution of forest fire frequency. (c) Distribution of fire severity classes.

Figure 3.

(a) Areas of burned forest over the study period. (b) Distribution of forest fire frequency. (c) Distribution of fire severity classes.

With respect to fire frequency for burned forests during the study time period, 72% was burned once (F1), 24% was burned twice (F2), and less than 4% was burned three times (Figure 3b). In terms of fire severity, we identified three fire severity classes, low, moderate, and high, within burned forest for each year and calculated the area of each severity class. On average, 70% of burned forest was lightly burned (low severity), 21.1% was moderately burned (moderate severity), and 9.1% was severely burned (high severity) (Figure 3c). Because of the limited numbers of pixels available for higher fire frequencies, the data analysis was limited to those burned forests with fire frequencies up to two (F1 and F2) and recovery ages between 1 and 5 years.

4.2. Spectral changes of burned forests

The effects of all three variables (fire frequency, fire severity, and forest recovery age) were highly significant to explain changed vegetation indices values (Table 3). Figure 4 shows the changes in each of the vegetation indices from once burned forests (F1) for different fire severity classes over the postfire recovery ages. The hyperspectral vegetation indices change similarly over the recovery ages after fire, with high variability within each combination of fire severity and age classes. For NDVI, NDII, NDWI, and CRI, an increasing trend was observed as recovery age increased, suggesting the progress of forest recovery in burned forests over time. In terms of spectral distinction of burned forests from undisturbed forest, NDVI, CRI, and PRI, those related to canopy greenness and pigment contents, showed longer periods of being statistically distinct from undisturbed forest than NDWI and NDII (Table 4). These indices can discriminate burned from undisturbed forest for 3 years or longer after forest fires. The distinguishable periods of the canopy water content indices vary from less than 1 to 3 years, depending on the degrees of fire severity. This may indicate a rapid recovery of canopy water content as the vegetation reestablishes after a fire. The high values of standard deviation of NDWI may be related to its sensitivity to spatial variation of canopy structures (Asner et al. 2005).

Table 3.

ANOVA F values of variables of variance for vegetation indices from burned forest.

 ANOVA F values of variables of variance for vegetation indices from burned forest.
 ANOVA F values of variables of variance for vegetation indices from burned forest.
Figure 4.

Changes in the hyperspectral vegetation indices of forests burned once (F1) as a function of fire frequency over time as the forest recovers.

Figure 4.

Changes in the hyperspectral vegetation indices of forests burned once (F1) as a function of fire frequency over time as the forest recovers.

Table 4.

Statistical differences among undisturbed forest and burned forests with the combined effects of fire severity and frequency over the recovery ages. The uppercase letters (A, B, C, D, etc.) indicate statistical differences among fire severity classes for a given recovery year, whereas the lowercase letters indicate statistical differences among recovery ages at the P < 0.01 of confidence levels of the Tukey test. Measures that have different letters in a column or a row are statistically different; otherwise, the measures are statistically the same.

Statistical differences among undisturbed forest and burned forests with the combined effects of fire severity and frequency over the recovery ages. The uppercase letters (A, B, C, D, etc.) indicate statistical differences among fire severity classes for a given recovery year, whereas the lowercase letters indicate statistical differences among recovery ages at the P < 0.01 of confidence levels of the Tukey test. Measures that have different letters in a column or a row are statistically different; otherwise, the measures are statistically the same.
Statistical differences among undisturbed forest and burned forests with the combined effects of fire severity and frequency over the recovery ages. The uppercase letters (A, B, C, D, etc.) indicate statistical differences among fire severity classes for a given recovery year, whereas the lowercase letters indicate statistical differences among recovery ages at the P < 0.01 of confidence levels of the Tukey test. Measures that have different letters in a column or a row are statistically different; otherwise, the measures are statistically the same.

With respect to the effects of fire severity, low and moderate fire severity classes for all indices are very similar through the recovery ages in F1; however, the high severity class is distinct from the others in all indices, excepting NDWI, presenting the lowest values throughout all recovery ages. Most burned forests of the high severity class are statistically different at least 5 years after forest fires for NDVI, CRI, PRI, and NDII (Table 3). The effects of fire severity became more significant for those forests that were burned twice (F2) (Figure 5 and Table 4). Differences between low and moderate severity classes in the case of NDVI and CRI were greater in F2 than in F1. Furthermore, the detectable periods of burned forests from undisturbed forests became longer in F2 for some indices such as NDVI for moderate severity burned forests, until Y3 in F1 to >Y5 in F2, and NDII for moderate severity forest, Y2 in F1 to >Y5 in F2. As expected, the greatest impact of fire on forests is observed in year 1 after fire (Y1), indicated by the lowest values of indices. Rapid recovery is seen in the period from the first (Y1) to second (Y2) year after burning. In most indices, the values stabilize after Y2, excepting the low and moderate severity classes of CRI which continue to rise. Overall, the results suggest that the higher the fire severity and frequency of burning, the longer fire-disturbed forests remain detectable by remote sensing.

Figure 5.

Changes in the hyperspectral vegetation indices of forests burned twice (F2) as a function of fire severity over time as the forest recovers.

Figure 5.

Changes in the hyperspectral vegetation indices of forests burned twice (F2) as a function of fire severity over time as the forest recovers.

In terms of the performances of vegetation indices for detecting burned forests and tracking them over time, CRI was best. This index is a narrowband-derived vegetation index sensitive to carotenoid pigment content of the uppermost canopy. Plant stress was detectable for the longest period among the analyzed vegetation indices. The conventional indices such as NDVI and NDII, derivable from broadband sensors, also provided relatively long detectable periods.

5. Discussion

The trajectory of forest recovery after fire observed in this study is similar to previous studies. Souza et al. (Souza et al. 2005b) found that the combined effects of fire and logging on forests in Mato Grosso were detectable with Landsat-derived measures, with statistical differences between intact forest and burned forests up to 3 years after fire using several vegetation indices, including NDVI and NDII. However, in most cases, it has been indicated that burned forests recover rapidly and become difficult to separate from undisturbed forest more than one year after forest fire using remote sensing (Cochrane and Souza 1998). Our results showed that recovery processes vary as a function of both fire frequency and burn severity as indicated in changing and physiological properties (i.e., canopy water and pigment contents) detectable with the hyperspectrally derived vegetation indices. Burned forests with low fire frequency and burn severity were difficult to distinguish from intact forest beyond the early stages of recovery, but those highly affected with higher fire frequency remained distinctive over longer periods.

The degree of forest damage is largely a function of fire severity and frequency. A field study in the eastern Amazon showed that heavily burned forest had substantially less biomass (45%) compared to undisturbed forest, whereas lightly burned forest had only 8% less biomass compared to unburned forest (Cochrane and Schulze 1998). In the case of this study, because 90% of burned forests are low and moderate severity classes, most burned forests may become spectrally undistinguishable from undisturbed forest within a few years. Those forests burned more than once (about 30%), however, may recover more slowly and be detectable for longer periods. If spatial and temporal changes of forest fire history in the state of Mato Grosso are similar to our study area, the majority of burned forests, under low–moderate severity and low fire frequency, may become undetectable using remote sensing within 5 years after burning, especially if using only broadband-derived vegetation indices.

Our results indicate that hyperspectrally derived measures may provide useful information for estimating changes in ecosystem responses after fire. In the case of the CRI, pigment is related to canopy stress (Gitelson et al. 2002) and an increase of CRI as forest recovery progresses after fire suggests reduced canopy stress and rebounding carotenoid content. Although both NDVI and CRI are related to the amounts of chlorophyll, NDVI becomes saturated or insensitive to changes in vegetation with high leaf area index values (Franklin et al. 1991), whereas CRI may be able to be used to track canopy physiological properties after the NDVI saturates (Asner et al. 2005). Our results showed that CRI provided longer periods of burned forests being distinguishable from undisturbed forest, even as forest recovery takes place.

Despite the results discussed above, some points should be taken into consideration for future analyses. The gap between the time of image acquisition and the time of occurrence of degradation is an important issue (Souza et al. 2005b). Our method to detect burned area was based on Landsat data, which provide only a few cloud-free images per year that may not coincident with the best timing for burned area detection. For example, three burned forest areas used for deriving RdNBR, from Souza et al. (Souza et al. 2005a) inventory data, had burned in 1999. No postfire imagery was available after the fires that year so burned forests are only detectable in the 2000 Landsat image. Therefore, forest burned area may be underestimated because of this time gap. It may affect our ability to accurately identify forest burned area and detection of postfire forest age. Because of the lack of temporal data of Hyperion data, we were unable to track actual temporal changes of forest regrowth within the same burned forests, but instead we substituted space for time by identifying burned forests of different fire frequency, fire severity, and recovery ages in the Hyperion data and assessing the temporal changes of forest recovery metrics, using vegetation indices, over recovery periods as long as 5 years. Because these temporal results were not acquired from the same spatial locations, spatial heterogeneity may have affected our results.

6. Conclusions

This study assessed the effects of fire frequency and fire severity on physiological characteristics of burned forests and their changes during forest recovery in the southern Amazon using a Landsat time series dataset and EO-1 Hyperion data. NDVI, CRI, and PRI, related to canopy greenness and pigment contents, can distinguish burned forests from undisturbed forest for 3 or more years after forest fire, whereas canopy water content indices (NDWI and NDII) are less effective, detecting burned forests for less than 3 years after fire. Our results suggested that the higher fire severity and greater fire frequency, the longer fire-disturbed forests remain distinguishable from undisturbed forest and detectable by remote sensing. However, because 90% of burned forests are affected by the low to moderate degrees of forest fire impacts, monitoring the recovery process of most burned forests and its distinction from unburned forests for many years after fire event may be still challenging.

Hyperion data have the potential to improve our ability to estimate physiological changes within disturbed forests, despite its low signal-to-noise ratio. Hyperspectral imagery is needed because subtle changes in the physiological properties of disturbed forests, as captured by CRI, cannot be determined using multispectral sensors. The results of this study may contribute to better characterization of burned forest ecosystems in the tropics and monitoring of physiological responses to fire at landscape scales. Future analyses will be improved with sensors like the Hyperspectral Infrared Imager (HyspIRI), which will have better signal-to-noise ratios and imaging of the same area over time. This will allow for monitoring of more extensive areas of disturbed forests.

Acknowledgments

This work was supported by the Biological Diversity Program of the Earth Science Division of the NASA Science Mission Directorate (NNX07AF16G).

REFERENCES

REFERENCES
Alencar
,
A. A. C.
,
L. A.
Solorzano
, and
D. C.
Nepstad
,
2004
:
Modeling forest understory fires in an eastern Amazonian landscape
.
Ecol. Appl.
,
14
,
S139
S149
.
Alencar
,
A. A. C.
,
D.
Nepstad
, and
M. D. V.
Diaz
,
2006
:
Forest understory fire in the Brazilian Amazon in ENSO and non-ENSO years: Area burned and committed carbon emissions
.
Earth Interactions
,
10
.
[Available online at http://EarthInteractions.org.]
Aragao
,
L.
,
Y.
Malhi
,
N.
Barbier
,
A.
Lima
,
Y.
Shimabukuro
,
L.
Anderson
, and
S.
Saatchi
,
2008
:
Interactions between rainfall, deforestation and fires during recent years in the Brazilian Amazonia
.
Philos. Trans. Roy. Soc.
,
363B
,
1779
1785
.
Asner
,
G. P.
,
M.
Keller
,
R.
Pereira
, and
J. C.
Zweede
,
2002
:
Remote sensing of selective logging in Amazonia—Assessing limitations based on detailed field observations, Landsat ETM+, and textural analysis
.
Remote Sens. Environ.
,
80
,
483
496
.
Asner
,
G. P.
,
K. M.
Carlson
, and
R. E.
Martin
,
2005
:
Substrate age and precipitation effects on Hawaiian forest canopies from spaceborne imaging spectroscopy
.
Remote Sens. Environ.
,
98
,
457
467
.
Chambers
,
J. Q.
,
G. P.
Asner
,
D. C.
Morton
,
L. O.
Anderson
,
S. S.
Saatch
,
F. D. B.
Espirito-Santo
,
M.
Palace
, and
C.
Souza
,
2007
:
Regional ecosystem structure and function: Ecological insights from remote sensing of tropical forests
.
Trends Ecol. Evol.
,
22
,
414
423
.
Cochrane
,
M. A.
, and
M. D.
Schulze
,
1998
:
Forest fires in the Brazilian Amazon
.
Conserv. Biol.
,
12
,
948
950
.
Cochrane
,
M. A.
, and
C. M.
Souza
,
1998
:
Linear mixture model classification of burned forests in the eastern Amazon
.
Int. J. Remote Sens.
,
19
,
3433
3440
.
Cochrane
,
M. A.
, and
M. D.
Schulze
,
1999
:
Fire as a recurrent event in tropical forests of the eastern Amazon: Effects on forest structure, biomass, and species composition
.
Biotropica
,
31
,
2
16
.
Cochrane
,
M. A.
, and
W. F.
Laurance
,
2002
:
Fire as a large-scale edge effect in Amazonian forests
.
J. Trop. Ecol.
,
18
,
311
325
.
Cochrane
,
M. A.
, and
W. F.
Laurance
,
2008
:
Synergisms among fire, land use, and climate change in the Amazon
.
Ambio
,
37
,
522
527
.
Franklin
,
J.
,
S. D.
Prince
,
A. H.
Strahler
,
N. P.
Hanan
, and
D. S.
Simonett
,
1991
:
Reflectance and transmission properties of West African savanna trees from ground radiometer measurements
.
Int. J. Remote Sens.
,
12
,
1369
1385
.
Frolking
,
S.
,
M. W.
Palace
,
D. B.
Clark
,
J. Q.
Chambers
,
H. H.
Shugart
, and
G. C.
Hurtt
,
2009
:
Forest disturbance and recovery: A general review in the context of spaceborne remote sensing of impacts on aboveground biomass and canopy structure
.
J. Geophys. Res.
,
114
,
G00E02
,
doi:10.1029/2008JG000911
.
Gamon
,
J. A.
,
J.
Penuelas
, and
C. B.
Field
,
1992
:
A narrow-waveband spectral index that tracks diurnal changes in photosynthetic efficiency
.
Remote Sens. Environ.
,
41
,
35
44
.
Gamon
,
J. A.
,
L.
Serrano
, and
J. S.
Surfus
,
1997
:
The photochemical reflectance index: An optical indicator of photosynthetic radiation use efficiency across species, functional types, and nutrient levels
.
Oecologia
,
112
,
492
501
.
Gao
,
B. C.
,
1996
:
NDWI—A normalized difference water index for remote sensing of vegetation liquid water from space
.
Remote Sens. Environ.
,
58
,
257
266
.
Gitelson
,
A. A.
,
Y.
Zur
,
O. B.
Chivkunova
, and
M. N.
Merzlyak
,
2002
:
Assessing carotenoid content in plant leaves with reflectance spectroscopy
.
Photochem. Photobiol.
,
75
,
272
281
.
GOFC-GOLD
,
2008
:
Reducing greenhouse gas emissions from deforestation and degradation in developing countries: A sourcebook of methods and procedures for monitoring, measuring and reporting
.
GOFC-GOLD Rep. Version COP13-2, 108 pp
.
Goodenough
,
D. G.
,
A.
Dyk
,
O.
Niemann
,
J. S.
Pearlman
,
H.
Chen
,
T.
Han
,
M.
Murdoch
, and
C.
West
,
2003
:
Processing Hyperion and ALI for forest classification
.
IEEE Trans. Geosci. Remote Sens.
,
41
,
1321
1331
.
Hunt
,
E. R.
, and
B. N.
Rock
,
1989
:
Detection of changes in leaf-water content using nearinfrared and middle-infrared reflectances
.
Remote Sens. Environ.
,
30
,
43
54
.
INPE
,
2010
:
PRODES—Desflorestamento nos Municipios
.
Instituto National de Pesquisas Espaciais. [Available online at http://www.dpi.inpe.br/prodesdigital/prodesmunicipal.php.]
Kokaly
,
R. F.
,
G. P.
Asner
,
S. V.
Ollinger
,
M. E.
Martin
, and
C. A.
Wessman
,
2009
:
Characterizing canopy biochemistry from imaging spectroscopy and its application to ecosystem studies
.
Remote Sens. Environ.
,
113
,
78
91
.
Koltunov
,
A.
,
S. L.
Ustin
,
G. P.
Asner
, and
I.
Fung
,
2009
:
Selective logging changes forest phenology in the Brazilian Amazon: Evidence from MODIS image time series analysis
.
Remote Sens. Environ.
,
113
,
2431
2440
.
Matricardi
,
E. A. T.
,
D. L.
Skole
,
M. A.
Pedlowski
,
W.
Chomentowski
, and
L. C.
Fernandes
,
2010
:
Assessment of tropical forest degradation by selective logging and fire using Landsat imagery
.
Remote Sens. Environ.
,
114
,
1117
1129
.
Miller
,
J. D.
, and
A. E.
Thode
,
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
.
Monteiro
,
A. L.
,
C. M.
Souza
, and
P.
Barreto
,
2003
:
Detection of logging in Amazonian transition forests using spectral mixture models
.
Int. J. Remote Sens.
,
24
,
151
159
.
Morton
,
D. C.
,
R. S.
DeFries
,
Y. E.
Shimabukuro
,
L. O.
Anderson
,
E.
Arai
,
F. D.
Espirito-Santo
,
R.
Freitas
, and
J.
Morisette
,
2006
:
Cropland expansion changes deforestation dynamics in the southern Brazilian Amazon
.
Proc. Natl. Acad. Sci. USA
,
103
,
14 637
14 641
.
Myneni
,
R. B.
,
R. R.
Nemani
, and
S. W.
Running
,
1997
:
Estimation of global leaf area index and absorbed par using radiative transfer models
.
IEEE Trans. Geosci. Remote Sens.
,
35
,
1380
1393
.
Nepstad
,
D. C.
, and
Coauthors
,
1999
:
Large-scale impoverishment of Amazonian forests by logging and fire
.
Nature
,
398
,
505
508
.
NRC
,
2007
:
Earth Science and Applications From Space: National Imperatives for the Next Decade and Beyond
.
National Academic Press, 456 pp
.
Pearlman
,
J. S.
,
P. S.
Barry
,
C. C.
Segal
,
J.
Shepanski
,
D.
Beiso
, and
S. L.
Carman
,
2003
:
Hyperion, a space-based imaging spectrometer
.
IEEE Trans. Geosci. Remote Sens.
,
41
,
1160
1173
.
RADAM BRASIL
,
1982
:
Geologia, geomorfologia, pedologia, vegetação e uso potencial da terra
.
Ministerio de Minas e Energia Rep., 428 pp
.
Roberts
,
D. A.
,
G.
Batista
,
J. G.
Pereira
,
E.
Waller
, and
B.
Nelson
,
1998
:
Change identification using multitemporal spectral mixture analysis: Applications in eastern Amazonia
.
Remote Sensing Change Detection: Environmental Monitoring Applications and Methods, C. Elvidge and R. Lunetta, Eds., Ann Arbor Press, 137–161
.
Rouse
,
J. W.
,
R. H.
Haas
,
J. A.
Schell
, and
D. W.
Deering
,
1973
:
Monitoring vegetation systems in the Great Plains with ERTS
.
Proc. Third ERTS-1 Symp., Washington, DC, NASA, SP-351, Vol. 1, 309–317
.
Souza
,
C. M.
,
D. A.
Roberts
, and
M. A.
Cochrane
,
2005a
:
Combining spectral and spatial information to map canopy damage from selective logging and forest fires
.
Remote Sens. Environ.
,
98
,
329
343
.
Souza
,
C. M.
,
D. A.
Roberts
, and
A. L.
Monteiro
,
2005b
:
Multitemporal analysis of degraded forests in the southern Brazilian Amazon
.
Earth Interactions
,
9
.
[Available online at http://EarthInteractions.org.]
Ustin
,
S. L.
,
D. A.
Roberts
,
J. A.
Gamon
,
G. P.
Asner
, and
R. O.
Green
,
2004
:
Using imaging spectroscopy to study ecosystem processes and properties
.
Bioscience
,
54
,
523
534
.