NOTICE -- The LBA-ECO Project website is no longer being supported.  This archive is a snapshot, as it existed in 2013, of the LBA-ECO website, maintained by NASA Goddard Space Flight Center, and now archived at the ORNL DAAC.  Links to external websites may be inactive. Final data products from the LBA project can be found at the ORNL DAAC.
banner
banner banner banner banner banner banner
banner banner banner banner banner banner banner
home aboutlibrarynews archivecontacts banner

spacer
banner
Investigations
Overview
Abstracts & Profiles
Publications
Research Sites
Meetings
Synthesis Groups
LBA-HYDROMET
LBA-Air-ECO
Logistics
Overview
Field Support
Travel
Visa
Shipping
Data
  Overview
Find LBA Data
Investigator Checklist
Process & Policy
Documentation & Archive
Training & Education
  Overview
Activities Summary
T&E Goals
Student Opportunities
  Folha Amazônica
 
spacer

Investigation:

LC-21 (Asner / Bustamante / Silva)

LBA Dataset ID:

LC21_SELECTIVE_LOGGING

Originator(s):

1. ASNER, G.P.
2. KNAPP, D.E.
3. BROADBENT, E.N.
      4. OLIVEIRA, P.J.C.
5. KELLER, M.M.
6. SILVA, J.N.M.

Point(s) of Contact:

ORNL DAAC User Services Office Oak Ridge National Laboratory Oak Ridge, Tennessee 37 (ornldaac@ornl.gov)

Dataset Abstract:

Amazon deforestation has been measured by remote sensing for three decades. In comparison, selective logging has been mostly invisible to satellites. In selective logging, a limited number of marketable tree species are cut, and logs are transported off site to sawmills. Unlike deforestation, which is readily observed from satellites, selective logging in the Brazilian Amazon causes a spatially diffuse thinning of large trees, which is hard to monitor by using satellite observations. Selective logging causes widespread collateral damage to remaining trees, subcanopy vegetation, and soils; with impacts on hydrological processes, erosion, fire, carbon storage, and plant and animal species Objective spatially explicit reporting on selective logging requires either labor-intensive field surveys in frontier and often violently contested areas or remote detection and monitoring approaches. Previous studies of small areas show the need for high-resolution observations via satellite. We developed a large-scale, high-resolution, automated remote-sensing analysis of selective logging in the top five timber-producing states of the Brazilian Amazon.

Beginning Date:

1999-01-01

Ending Date:

2002-12-31

Metadata Last Updated on:

2013-07-02

Data Status:

Archived

Access Constraints:

PUBLIC

Data Center URL:

http://daac.ornl.gov/

Distribution Contact(s):

ORNL DAAC User Services Office Oak Ridge National Laboratory Oak Ridge, Tennessee 37 (ornldaac@ornl.gov)

Access Instructions:

PUBLIC

Data Access:

IMPORTANT: The LBA-ECO Project website is no longer being supported. Links to external websites may be inactive. Final data products from the LBA project can be found at the ORNL DAAC. Please follow the fair use guidelines found in the dataset documentation when using or citing LBA data.
Datafile(s):

LBA-ECO LC-21 Selective Logging Activity in the Brazilian Amazon: 1999-2002:  http://daac.ornl.gov/cgi-bin/dsviewer.pl?ds_id=1172

Documentation/Other Supporting Documents:

LBA-ECO LC-21 Selective Logging Activity in the Brazilian Amazon: 1999-2002:  http://daac.ornl.gov/LBA/guides/LC21_Selective_Logging.html

Citation Information - Other Details:

Asner, G.P., D.E. Knapp, E.N. Broadbent, P.J.C. Oliveira, M.M. Keller, and J.N.M. Silva. 2013. LBA-ECO LC-21 Selective Logging Activity in the Brazilian Amazon: 1999-2002. Data set. Available on-line [http://daac.ornl.gov] from Oak Ridge National Laboratory Distributed Active Archive Center, Oak Ridge, Tennessee, USA http://dx.doi.org/10.3334/ORNLDAAC/1172

Keywords - Theme:

Parameter Topic Term Source Sensor
DEFORESTATION BIOSPHERE LAND USE/LAND COVER Landsat ETM+ (ENHANCED THEMATIC MAPPER PLUS)

Uncontrolled Theme Keyword(s):  AMAZON, LANDSAT, SELECTIVE LOGGING, TROPICAL FOREST

Keywords - Place (with associated coordinates):

Region
(click to view profile)
Site
(click to view profile)
North South East West
  BRAZILIAN AMAZON -7.05290 -19.86830 -60.42570 -74.79310

Related Publication(s):

Asner, G.P., D.E. Knapp, A.N. Cooper, M.M.C. Bustamante, and L.P. Olander. 2005. Ecosystem Structure throughout the Brazilian Amazon from Landsat Observations and Automated Spectral Unmixing. Earth Interactions 9(7):1-31.

Asner, G.P., D.E. Knapp, E.N. Broadbent, P.J.C. Oliveira, M. Keller, and J.N. Silva. 2005. Selective logging in the Brazilian Amazon. Science 310(5747):480-482.

Asner, G.P., E.N. Broadbent, P.J.C. Oliveira, M. Keller, D.E. Knapp, and J.N.M. Silva. 2006. Condition and fate of logged forests in the Brazilian Amazon. Proceedings of the National Academy of Sciences of the United States of America 103(34):12947-12950.

Data Characteristics (Entity and Attribute Overview):

Data Characteristics:

The data are digital images in ENVI Standard format.

Each image is a single band in which each pixel is represented by a single byte.

A zero value (0) indicates that no logging was detected while a value of one (1) indicates that damage from logging was detected.



Projection

-----------

Because the data cover such a large area, no single UTM zone can adequately represent the data at the state level. Goodes is an equal-area projection, making it ideal for areal determinations and comparisons over large regions. Thus, these data are provided in the Goodes Interrupted Projection. For most of South America, the parameters of this projection are:



Projection Type: SINUSOIDAL

Map units: Meters

Radius of Sphere: 6370997.0

Central Meridian: W 60 00 00

False Easting: -6671692.45708

False Northing: 0.0



A more detailed description of the Goode's projection can be found at:

http://gis.esri.com/library/userconf/proc98/PROCEED/TO850/PAP844/P844.HTM



The image files are compressed using gzip compression. ENVI can read these files directly since the corresponding header file for each image indicates that the file is compressed. Although ENVI can read the compressed files directly, some interactive functions performed on the imagery may be slow. This can be remedied by saving the image in ENVI to a new image file, making sure not to check the Compression box. The image file can also be explicitly uncompressed by using the gunzip command and simply removing the line that begins with file compression from the ENVI header file.



The gzip and gunzip commands are freely available at the following website:

http://www.gzip.org/





Files

-----------

The files in this dataset include:

15 image file (.img)

15 header files (.img.hdr)

and one csv file



CIW_image_dates_in_selective_logging_study.csv - ASCII comma-separated file listing the Landsat ETM+ Image pairs used in the study



Para_Logging_PVC_2000.img - Logging Image file of 1999-2000 for Para

Para_Logging_PVC_2001.img - Logging Image file of 2000-2001 for Para

Para_Logging_PVC_2002.img - Logging Image file of 2001-2002 for Para

MT_Logging_PVC_2000.img - Logging Image file of 1999-2000 for Mato Grosso

MT_Logging_PVC_2001.img - Logging Image file of 2000-2001 for Mato Grosso

MT_Logging_PVC_2002.img - Logging Image file of 2001-2002 for Mato Grosso

Rondonia_Logging_PVC_2000.img - Logging Image file of 1999-2000 for Rondonia

Rondonia_Logging_PVC_2001.img - Logging Image file of 2000-2001 for Rondonia

Rondonia_Logging_PVC_2002.img - Logging Image file of 2001-2002 for Rondonia

Roraima_Logging_PVC_2000.img - Logging Image file of 1999-2000 for Roraima

Roraima_Logging_PVC_2001.img - Logging Image file of 2000-2001 for Roraima

Roraima_Logging_PVC_2002.img - Logging Image file of 2001-2002 for Roraima

Acre_Logging_PVC_2000.img - Logging Image file of 1999-2000 for Acre

Acre_Logging_PVC_2001.img - Logging Image file of 2000-2001 for Acre

Acre_Logging_PVC_2002.img - Logging Image file of 2001-2002 for Acre

Para_Logging_PVC_2000.img.hdr - Logging Header file of 1999-2000 for Para

Para_Logging_PVC_2001.img.hdr - Logging Header file of 2000-2001 for Para

Para_Logging_PVC_2002.img.hdr - Logging Header file of 2001-2002 for Para

MT_Logging_PVC_2000.img.hdr - Logging Header file of 1999-2000 for Mato Grosso

MT_Logging_PVC_2001.img.hdr - Logging Header file of 2000-2001 for Mato Grosso

MT_Logging_PVC_2002.img.hdr - Logging Header file of 2001-2002 for Mato Grosso

Rondonia_Logging_PVC_2000.img.hdr - Logging Header file of 1999-2000 for Rondonia

Rondonia_Logging_PVC_2001.img.hdr - Logging Header file of 2000-2001 for Rondonia

Rondonia_Logging_PVC_2002.img.hdr - Logging Header file of 2001-2002 for Rondonia

Roraima_Logging_PVC_2000.img.hdr - Logging Header file of 1999-2000 for Roraima

Roraima_Logging_PVC_2001.img.hdr - Logging Header file of 2000-2001 for Roraima

Roraima_Logging_PVC_2002.img.hdr - Logging Header file of 2001-2002 for Roraima

Acre_Logging_PVC_2000.img.hdr - Logging Header file of 1999-2000 for Acre

Acre_Logging_PVC_2001.img.hdr - Logging Header file of 2000-2001 for Acre

Acre_Logging_PVC_2002.img.hdr - Logging Header file of 2001-2002 for Acre

Data Application and Derivation:

These data can be used to identify areas of logging activity. Since the image pairs that were used to create these mosaics have varying acquisition dates, care should be taken in determining the total amount of logging for a one-year period.

Quality Assessment (Data Quality Attribute Accuracy Report):

Quality Assessment:

We carefully quantified our uncertainty in four key areas: atmospheric correction (aerosol and water vapor and unobserved areas caused by persistent cloud cover, annualization and auditor related uncertainty.



Atmospheric Uncertainty



In the CLAS processing stream, Landsat ETM+ images are atmospherically corrected using the 6S atmospheric correction algorithm, with monthly averages of aerosol and water vapor inputs from the MODIS satellite sensor. The CLAS AutoMCU algorithm was found to be minimally sensitive to uncertainties in aerosol and water vapor from MODIS (Asner et al. 2005). To further understand the effect that the atmospheric correction has on the sensitivity of entire CLAS process, we atmospherically corrected five Landsat image pairs using randomly-selected, monthly aerosol and water vapor values from MODIS. The difference in the amount of automatically detected logging between the different atmospherically-corrected images was only 0.7%.



Unobserved Area Uncertainty



When cloud and cloud-shadow cover is greater than 50% in any 5,625 km squared area (2,500 x 2,500 pixels), the area of observed logging is used to estimate the amount of logging in the unobserved, cloudy areas. We assessed our sensitivity to this type of error by simply quantifying the fractional cover of clouds and cloud shadows in comparison to observed logging extent. The calculated absolute uncertainty caused by this step was approximately +5% over the five states.



Annualization Uncertainty



Although the rate of logging is assumed constant throughout the dry season, there is a level of uncertainty inherent in this assumption. Marengo et al. reported rainy season length for five regions of the Amazon (i.e., North Amazonia, Central Amazonia, Mouth of Amazon, Southeast Amazonia, and Southwest Amazonia) for the period 1979-1996. To determine the uncertainty in the logging estimate related to assumption of dry season length, a series of matched pairs of dry season length for two consecutive years (e.g., 1979-1980, 1980- 1981, 1995-1996) was compiled to calculate the standard error of the difference in dry season length for each region. This standard error (in days) was divided by the average length of the dry season for the respective region to express the uncertainty in percent of dry season. This percentage uncertainty was then applied to actual satellite image pairs or by averaging the uncertainty for states located between two regions. These uncertainties ranged from 2-9% as a

result of interannual variation in dry season length (see supporting documentation). To further assess the sensitivity of the logging area estimates to the annualization and timing of the dry season, the estimates were also annualized without the constraint that logging activity only occurs only during the dry season. These results are reported in the supporting documentation. It is clear that the differences between these two assumptions can be large in the smaller states, however, in the larger states, these uncertainties tend to balance out. In the majority of cases, the amounts of logging estimated

without the dry season constraint still falls within the minimum and maximum limits of estimated logged area caused by other sources of uncertainty.



Auditor Uncertainty



Each auditor reviewed a set of the same 25 image subsets (400 by 400 pixels) in which most images include some form of logging. A test was performed in which a novice and an experienced image analyst manually delineated areas containing logged forest. This comparison was used to calculate one standard error of the difference in logging assessments between auditors for each image subset. The standard error between auditors was 0.69 km2 of logging, which when scaled by the average amount of logging identified by the two analysts (5.4 km2), resulted in an uncertainty of 12.8%.



These different sources of uncertainty were compiled and used to estimate an overall uncertainty in the logging extent estimates of 11-14% for each Brazilian state in each year of analysis . These uncertainties were then propagated to the Basin scale for annual estimates of selective logging for the years 2000, 2001, and 2002 (supporting information).

Process Description:

Data Acquisition Materials and Methods:

Geographic Coverage



The Brazilian Amazon basin covers an area of approximately 4.1 million km squared. Analysis of the entire region with Landsat Enhanced Thematic Mapper-Plus (ETM+) imagery would require approximately 220 scenes per year or 880 images for the years 1999-2002, yet much of the northwestern Amazon still contains relatively little deforestation and logging (Nepstad et al. 1999). We therefore limited our study to the States of Acre, Para, Mato Grosso (northern 58% of the state containing most of the forested area), Rondonia, and Roraima. These five states contain ~90% of the deforestation reported by Brazil for all of the Legal Amazon (INPE 2005). This strategy reduced the number of required Landsat images to 480 scenes. Our study covered the period 1999 to 2002, which is prior to the failure of the Scan Line Corrector (SLC) in the ETM+ instrument onboard Landsat 7. Following the SLC failure, roughly 40% of each acquired Landsat image is missing data. To seek out alternatives to Landsat 7, we conducted a satellite inter-comparison of logging detection capability based on our network of low- and high-intensity logging sites in Amazonia (Asner et al 2004b). We compared the detection capabilities of 10 sensors hyperspectral and the only sensor to meet or exceed the performance of Landsat ETM+ was EO-1 Hyperion; all others failed to detect at least 80% of the logging damage in our field sites. However, the extremely limited spatial and temporal coverage of the EO-1 Hyperion makes its application to large-area analysis intractable.



Materials and Methods Processing Methodology The Carnegie Landsat Analysis System (CLAS) uses high spatial resolution satellite data for regional and global studies of forest disturbance. CLAS is an automated processing system that includes: (i) atmospheric correction of satellite data; (ii) deconvolution of spectral signatures into sub-pixel fractional cover of live forest canopy, forest debris and bare substrates; (iii) cloud, water, and deforestation masking; and (iv) pattern recognition algorithms for forest disturbance mapping.



Image Preparation and Atmospheric Correction The version of CLAS presented here ingests raw Landsat Enhanced Thematic Mapper Plus (ETM+) satellite imagery and applies sensor gains and offsets to convert from digital number (DN) to exo-atmospheric radiance. The radiance data are passed to a fully automated version of the 6S atmospheric radiative transfer model. The 6S program is integrated into the CLAS processing stream and uses monthly averages of aerosol optical thickness (AOT) and water vapor (WV) values from the Moderate Resolution Imaging Spectrometer (MODIS) sensor onboard the NASA Terra spacecraft. Time-stamping of MODIS AOT and WV data with Landsat data is done on an automated basis.



Sub-pixel Analysis The CLAS process relies upon the quantitative determination of fractional material cover at the sub-pixel scale (e.g., within each Landsat 30 x 30 m pixel). This core step employs a probabilistic spectral mixture sub-model. This process spectrally decomposes each image pixel into fractional cover estimates (0-100% cover) of photosynthetic vegetation (PV) canopy, non-photosynthetic vegetation (NPV), and bare substrate. This sub-model is based on an algorithm developed for forest, savanna, woodland and shrubland ecosystems (Asner and Lobel 2000 RSE, Asner et al. 2004a, Asner and Heidebrecht 2002). It is fully automated and uses a Monte Carlo Unmixing (AutoMCU) approach to derive uncertainty estimates of the sub-pixel cover fraction values. The method uses three spectral endmember bundles, derived from extensive field databases and satellite imagery, to decompose each image pixel using a linear equation (see supporting information for full equation)



Until recently, there were a limited number of spectral signatures of green and senescent vegetation and bare substrates for tropical regions. Our mixture modeling technique requires spectral reflectance bundles (rho pv(lambda), rho npv(lambda), and rho substrate(lambda)) that encompass the common variation in canopy and soil properties. Asner (1998) and Asner et al. (2004a, 2003) collected these spectral data using full optical range field spectroradiometers (Analytical Spectral Devices, Inc., Boulder, CO, USA) during field campaigns conducted from 1996 to 2000. The spectral endmember database encompasses the common variation in materials found throughout the Brazilian Amazon, with statistical variability well defined (Asner et al 2004a). The bare substrate spectra have been collected across a diverse range of soil types, surface organic matter levels, and moisture conditions. Spectral collections for NPV have included surface litter, senescent grasslands, and deforestation residues (slash) from a wide range of species and decomposition stages. In contrast to the NPV and bare substrate spectra that can be collected via ground-based spectroscopic measurements, the photosynthetic vegetation (PV) spectra of forest species require overhead viewing conditions. This is very difficult in forest canopies with heights typically ranging from 10-50 m. Spectral measurements of individual leaves, stacks of foliage, or partial canopies (e.g., branches) introduce major errors in spectral mixture models and cannot be used (Asner 1998). Therefore, we collected canopy spectra using the Earth Observing-1 (EO-1) Hyperion sensor, the first spaceborne hyperspectral sensor for environmental applications (Ungar et al 2003). The PV spectral bundle was derived from more than 40,000 spectral observations made at 30 m spatial resolution with Hyperion (images taken throughout 1999), atmospherically corrected to apparent top-of-canopy reflectance using the ACORN-4 atmospheric correction algorithm for hyperspectral data (ImSpec Inc., Palmdale, CA USA), and convolved to Landsat ETM+ optical channels (Asner at al 2005). These green vegetation spectra thus inherently included the variable effects of intra- and inter-crown shadowing, which are prevalent in tropical forests (Gastellu-Etchegorry et al 1999). In Amazonia, shade fractions average 25% cover in humid tropical forests, but the variance is high with standard deviations of 12% or more (Asner and Warner 2003). It is thus critically important to note that our PV results include shade, which varies substantially with forest structure. Using a separate shade endmember is attractive (Souza and Barreto 2000), but doing so with multi-spectral Landsat data and such high shadow fraction variability often results in an under-determined spectral and mathematical problem in linear mixture models. That is, there are many viable solutions to the mixture modeling problem in forests. Imaging spectroscopy (hyperspectral) data are needed to solve this problem (Roberts et al. 1993). We avoided this issue by accepting the limitations of incorporating variable shade directly into our PV bundle derived from the EO-1 Hyperion sampling of undisturbed forest canopies in Brazil. The PV bundle includes spectra from mature forest, late-stage forest regrowth, and logged forest of at least five years post-harvest. In the end, the total number of spectra retained in the endmember bundles for the AutoMCU sub-model was 252 for PV, 611 for NPV, and 434 for bare substrate. These spectra represent more than 130,000 field and spaceborne spectrometer observations collected over a five-year period of study (Asner et al 2005).



Non-forest Masking and Atmospheric Compensation A series of automated masks were designed to exclude clouds, water bodies, cloud shadows, non-image and non-forest areas (e.g., pasture, urban and agriculture) from the CLAS processing stream. Prior to execution of the AutoMCU sub-model, clouds are masked using the thermal channel (band 6) from the raw Landsat images. Asner et al. (2005) found that a thermal band threshold DN value of 125 can conservatively detect cloudy pixels over Amazonia. Water bodies are masked by finding pixels in the calibrated Landsat reflectance data in which bands 1-4 (blue, green, red, and near-infrared) have a negative slope. Only water displays such a negative reflectance slope with increasing wavelength. Non-image areas containing zero values are also masked. Cloud shadows are identified using the root mean square error (RMSE) image that results from the AutoMCU processing. Areas shadowed by clouds have large RMSE values and are masked by identifying pixels above a specific RMSE threshold (Asner et al 2005). To limit the logging analysis to forested areas, Landsat thermal band 6, combined with the AutoMCU results, is used to identify pixels containing primarily forest and non-forest areas. Forests have a lower brightness temperature and a higher PV fractional cover than deforested lands. We employ a conservative PV fractional cover threshold of 60% to delineate forest cover in the PV mask. The minimum and maximum thermal thresholds, which encompass forested areas in the thermal mask, are dynamically generated for each image by calculating the mean thermal value of all pixels having a PV fraction cover greater than 80% and then masking all pixels with values > 15 digital numbers (DN) from the mean thermal value. These final masking steps have the added feature of removing residual clouds and cloud shadows that were missed in the masks applied earlier in the CLAS process.



Although atmospheric correction was performed on the raw imagery before processing through the AutoMCU sub-model, residual atmospheric effects can persist (Asner et al 2005). These residual effects exist spatially within a scene and temporally between scenes. We greatly reduce these effects by calculating the average change in fractional forest cover in 55 km2 subsets of the imagery. These large geographic subsets are made at a spatial scale far greater than that of the most extensive logging activities, so temporal differences in the overall forest fractional cover at this scale are a result of atmospheric effects (e.g., haze) or forest phenology. These false fractional cover changes are normalized by adjusting the background forest temporal variation to zero. Since disturbances related to logging or other anthropogenic activities occur at a much smaller spatial scale than is considered in this processing step, normalization of the forest values across large areas does not affect the CLAS process in discriminating true disturbances from the surrounding forested areas.



Pattern Recognition:

The specific criteria used in this procedure were determined following a comprehensive analysis and review of the forest responses to logging at various intensities in the Brazilian states of Para, Mato Grosso and Acre where we conducted field studies. The mean and standard deviation fractional cover images from the AutoMCU step in CLAS provide quantitative data on canopy damage and forest disturbance intensity from which selectively logged areas can be determined. By identifying areas of canopy disturbance that are arranged in specific spatial patterns, it is possible to detect logged areas on an automated basis. The primary method by which logging is detected is image differencing, where pairs of AutoMCU sub-pixel fractional cover images, separated by approximately one year, are used to create images of PV (forest canopy) and NPV (surface woody and senescent vegetation material) change that indicate areas of relative canopy disturbance or recovery. Forest disturbances in these images always have reductions in PV, simultaneous with increases in

NPV fractional cover. Logging activity results in low intensity forest disturbances from tree felling gaps, moderate intensity linear features from skid trails along which felled trees are dragged by tractors or skidders, and high intensity points of damage called log decks where logs are loaded onto trucks for transportation. The log decks are connected by logging roads, seen as linear features causing large reductions in the fractional cover of PV, to local roads or rivers for transportation to markets. These patterns are unique to logging throughout most of the Amazon, and thus they serve as the basis upon which our method for logging detection functions. CLAS identifies points (e.g., treefall gaps and log decks) and linear features (e.g., skid trails and logging roads) of recent disturbance occurring in forested areas. As these features also exist at a lower frequency in intact forest regions, their spatial density and diversity (see definition in next section) are calculated to identify those areas having disturbances in patterns most indicative of logging activity. The procedure then identifies these areas for further analysis by creating point maps, termed logging nodes, indicating their locations.



Log decks are automatically detected by searching for pixels where PV decreases

significantly in a 30 m pixel centered on a 7 x 7 pixel kernel (4.41 ha). A positive detection is flagged when pixels with large PV reduction are surrounded by three concentric rings of incrementally greater PV cover surrounding the target pixel. This indicates an increase in canopy damage with greater proximity to the log deck, a pattern consistent with most logging

activities. The strategy for detecting decks works well in areas logged at higher intensities, as the decks tend to be abundant and equally spaced. However, in areas where the logging is more haphazard, where the forest damage is extremely high or low, or where the roads themselves also function as loading zones, individual log decks are not always distinguishable. Skids trails

are a typological feature of selective logging practices, and they are the single-most ubiquitous surface feature found in harvested areas (Pereira et al 2002, Asner et al 2004 GCB). The presence of skid trails is quantifiable

based on large decreases in PV fractional cover in linear or near-linear patterns (Asner et al 2004 EA). To detect the concentration of skid trails and auxiliary roads, a moving 6 x 6 pixel (3.24 ha) kernel is applied to the PV change image to enhance linear features in the N-S, E-W, NE-SW, and NWSE

directions. The number of directions in which the linear features are arranged

(which we define as their diversity), and their spatial density, in conjunction with the presence or absence of logging decks, is calculated for each location. With this information, it is possible to automatically distinguish probable logging events. In general, areas of greater logging intensity have a roughly equal proportion and higher density of linear features with the presence of logging decks. Lower intensity areas are normally dominated by one direction of linear feature and have few or no logging decks. An example of a typical logging detection is shown in the supporting information.



Final Integration:

After the linear and logging deck pattern recognition steps are completed, CLAS

automatically integrates the various results to identify contiguous pixel clusters of probable logging activity. This process starts by creating a list of the logging nodes that are identified in the previous steps. Logged areas are identified using a moving kernel approach. A base kernel of 7 x 7 pixels (4.41 ha) and four 3 x 3 pixel (0.81 ha) subset kernels, one located at each corner of the base kernel, are used. The base kernel begins at each logging node and tests the criteria described below. If the area in question tests positive, the analysis kernel is moved to its 7 x 7 pixel neighbors to the north, south, east, and west, which are then each tested against the criteria. This iterative process continues until all neighbors have been evaluated or the

maximum logged cluster size (maximum of 17 positive detections per logging node) has been reached. The input layers and specific criteria tested within the base and subset kernels are described below. For the criteria below, all units for PV and NPV are % fractional cover within a pixel; units for PV CI and NPV CI are % change in cover fractions between image dates.



Input layers to logged area detection procedure:

� Logging node map

� Thermal RMS mask (dynamically generated in earlier procedure) (T-mask)

� PV mask (> 60% fractional cover) (PV-mask)

� PV change difference image (PV CI)

� NPV change difference image (NPV CI)

� After image PV (AI PV)



Base kernel criteria:

� 75% good data pixels (not cloud, cloud shadow, or water)

� Non-forested area < 0.54 ha (12.2%); based on T- and PV-masks.

� 60% < Mean AI PV > 93%

� Mean PV CI > -9%

� Mean NPV CI < 2%

� Mean PV CI standard deviation > 33%

� Mean NPV CI standard deviation > 46%

� More than 6 pixels (0.54 ha) with PV CI values > 80%

� More than 6 pixels (0.54 ha) with NPV CI values < -85%

� Masked area < 0.18 ha



Subset kernel criteria:

� > 2 subsets with PV CI > 32% standard deviation

� > 2 subsets with mean PV CI < 3% and > 60%

� > 2 subsets having > 1 pixel (0.09 ha) with a PV CI value > 80%

� > 2 subsets with NPV CI > 46% standard deviation

� > 2 subsets with mean NPV CI < -5% and > -65%

� > 2 subsets having > 1 pixel (0.09 ha) with a NPV CI value < -85%



Manual Audit

Maps of probable logging events were visually audited to verify whether an area is being logged or not, in accordance with criteria established for identification of logged areas. In this process, false positives and negatives were manually removed and added. In this Amazon study, two analysts were employed during the audit, and their results and uncertainties were monitored and compared. The audit logging criteria are divided into high- and low-damage obvious and nonobvious categories. These categories encompassed all probable logging events in the study area and were identified after extensive review of logging events identified in the field. The criteria applied in each category are listed below.



High-damage obvious criteria

� Abundance of logging decks

� Obvious linear features (including primary-tertiary access roads and skid trails)

� Severe canopy damage visible in PV change difference image

� Areal extent normally > 1 ha

� Evidence of logging from previous years in close proximity (< -15% PV change

difference image)



High-damage non-obvious criteria

� Few to no logging decks

� Few to no linear features

� Severe canopy damage visible in PV change difference image (> 15%)

� Presence of access roads or rivers, if not adjacent to a anthropogenic non-forest areas

� Areal extent normally > 1 ha

� Evidence of logging from previous years in close proximity (< -15% in PV change difference image)



Low-damage obvious criteria

� Few to no logging decks

� Obvious linear features

� Presence of access roads or rivers

� Often tree-like in formation (graduating from higher to lower damage linear features)

� Not a linear feature connecting non-forest areas, or otherwise used for generaltransportation

� Often encompassing large extents

� Evidence of logging from previous years in close proximity (< -15% PV change

difference image)



Low-damage non-obvious criteria

� Few to no logging decks

� Few to no linear features

� Close proximity to access (i.e. roads, rivers or anthropogenic non-forest areas)

� Speckles of recent canopy damage (felling gaps; > 15%) in PV change difference image occurring at a density greater than in the surrounding forest areas

� Areal extent normally < 6.5 ha

� Evidence of logging from previous years in close proximity (< -15% in PV change difference image)

References:

Asner, G.P., 1998. Biophysical and biochemical sources of variability in canopy reflectance. Remote Sens. Environ., 64, 234-253.



Asner, G.P. and D.B. Lobell, 2000. A biogeophysical approach for automated SWIR unmixing of soils and vegetation. Remote Sens. Environ., 74, 99-112.



Asner, G.P. and K.B. Heidebrecht, 2002. Spectral unmixing of vegetation, soil and dry carbon cover in arid regions: Comparing multispectral and hyperspectral observations. Int. J. Remote Sens., 23, 3939-3958.



Asner, G.P. and A.S. Warner, 2003. Canopy shadow in IKONOS satellite observations of tropical forests and savannas. Remote Sens. Environ., 87, 521-533.



Asner, G.P., M.M.C. Bustamante, and A.R. Townsend, 2003. Scale dependence of biophysical structure in deforested areas bordering the Tapajos National Forest, Central Amazon. Remote Sens. Environ., 87, 507-520.



Asner, G.P., M. Keller, R. Pereira, J.C. Zweede, and J.N.M. Silva, 2004. Canopy damage and recovery after selective logging in Amazonia: Field and satellite studies. Ecol. Appl., 14, S280-S298.



Asner GP, Knapp DE, Cooper AN, Bustamante MMC, Olander LP. 2005a. Ecosystem Structure throughout the Brazilian Amazon from Landsat Observations and Automated Spectral Unmixing. Earth Interactions, 9, 1-31



Asner, G.P., M. Keller, J.N M. Silva. 2005b Global Change Biol. 10, 7652005b

Gastellu-Etchegorry, J.P., P. Guillevic, F. Zagloski, V. Demarez, V. Trichon, D. Deering, and M. Leroy, 1999. Modeling BRF and radiation regime of boreal and tropical forests: I. BRF. Remote Sens. Environ., 68, 281-316.



INPE (Instituto Nacional de Pesquisas Espaciais (INPE),2005. PRODES: Assessment of deforestation in Brazilian Amazonia (http://www.obt.inpe.br/prodes/index. html)



Marengo, J.A., B. Liebman, V. E. Kousky, N. P. Filizola, I. C. Wainer. 2001. J. Climate 14,833.



Nepstad DC, Verissimo A, Alencar A, Nobre C, Lima E, Lefebvre P, Schlesinger P, Potter C, Moutinho P, Mendoza E, Cochrane M, Brooks V. (1999) Large-scale impoverishment of Amazonian forests by logging and fire. Nature, 398, 505-508.





Roberts, D.A., M.O. Smith, and J.B. Adams, 1993. Green vegetation, nonphotosynthetic vegetation,and soils in Aviris data. Remote Sens. Environ., 44, 255-269.



Souza and Barreto 2000 Int\'l J. Rem. Sens. 21, 173



Ungar, S.G. J.S. Pearlman, J.A. Mendenhall, D. Reuter, IEEE Trans. Geosci. Rem. Sens. 41, 1149

Skip navigation linksHOME | ABOUT | LIBRARY | NEWS ARCHIVE | CONTACTS | INVESTIGATIONS | LOGISTICS | DATA |TRAINING & EDUCATION

NASA logo
ORNL DAAC
Get Acrobat Reader