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
home aboutlibrarynews archivecontacts banner

Abstracts & Profiles
Research Sites
Synthesis Groups
Field Support
Find LBA Data
Investigator Checklist
Process & Policy
Documentation & Archive
Training & Education
Activities Summary
T&E Goals
Student Opportunities
  Folha Amazônica


ND-01 (Roberts / Barreto / Soares)

LBA Dataset ID:



4. BIGGS, T.W.
      7. DEWES, C.
10. SOUZA, C.M.

Point(s) of Contact:

ORNL DAAC User Services Office Oak Ridge National Laboratory Oak Ridge, Tennessee 37 (

Dataset Abstract:

We conducted a deep temporal analysis of Rondonia, southwestern Brazilian Amazonia, using seven Landsat path‐row scenes representing approximately 80% of the state and covering the 27 years between 1984 and 2010. Each image classifies land cover into 7 different classes allowing a long-term assessment of land-co.ver variation across the state.

Beginning Date:


Ending Date:


Metadata Last Updated on:


Data Status:


Access Constraints:


Data Center URL:

Distribution Contact(s):

ORNL DAAC User Services Office Oak Ridge National Laboratory Oak Ridge, Tennessee 37 (

Access Instructions:


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.

LBA-ECO ND-01 Landsat 28.5-m Land Cover Time Series, Rondonia, Brazil: 1984-2010:

Documentation/Other Supporting Documents:

LBA-ECO ND-01 Landsat 28.5-m Land Cover Time Series, Rondonia, Brazil: 1984-2010:

Citation Information - Other Details:

Roberts, D.A.,M. Toomey, I. Numata, T. Biggs, J. Caviglia-Harris, M. Cochrane, C. Dewes, K.W. Holmes, R.L. Powell, C, Souza and O.A. Chadwick. 2013. LBA-ECO ND-01 Landsat 28.5-m Land Cover Time Series, Rondonia, Brazil: 1984-2010. Data set. Available on-line ( from Oak Ridge National Laboratory Distributed Active Archive Center, Oak Ridge, Tennessee, USA

Keywords - Theme:

Parameter Topic Term Source Sensor


Keywords - Place (with associated coordinates):

(click to view profile)
(click to view profile)
North South East West
  RONDONIA -7.82720 -13.86200 -58.79540 -64.59500

Related Publication(s):

Holmes, K.W., D.A. Roberts, S. Sweeney, I. Numata, E. Matricardi, T.W. Biggs, G. Batista, and O.A. Chadwick. 2004. Soil databases and the problem of establishing regional biogeochemical trends. Global Change Biology 10(5):796-814.

Numata, I, M.A. Cochrane, D.A. Roberts and J.V. Soares. 2009. Determining dynamics of spatial and temporal structures of forest edges in South Western Amazonia. Forest Ecology and Management, 258, 2547-2555.

Numata, I., M. A. Cochrane, D. A. Roberts, J. V. Soares, C. M. Souza Jr., and M. H. Sales. 2010, Biomass collapse and carbon emissions from forest fragmentation in the Brazilian Amazon, J. Geophys. Res., 115, G03027, doi:10.1029/2009JG001198.

Roberts, D.A., I. Numata, K. Holmes, G. Batista, T. Krug, A. Monteiro, B. Powell, and O.A. Chadwick. (2002) Large area mapping of land-cover change in Rondonia using multitemporal spectral mixture analysis and decision tree classifiers. Journal of Geophysical Research-Atmospheres, Vol. 107, No. D20, p. 8073.

Data Characteristics (Entity and Attribute Overview):

Data Characteristics:

Data are presented as 27 ENVI files each with an associated header file identified by the hdr extension. File names follow the following scheme: rondonia_YYYY_landcover_pheno_bsq.gz with YYYY indicating the image year. Each image is a mosaic file covering about 80% of the state of Rondonia. ENVI header files include the class names, a color look up table and all geographic information associated with the scene (spatial resolution, scene coordinates and datum).

Data Application and Derivation:

This 27 year time series can be used to determine rates of deforestation, changes in forest area as well as forest edge as well as spatial and temporal patterns in land-cover for the area.

Quality Assessment (Data Quality Attribute Accuracy Report):

Quality Assessment:

Accuracy of the classified map was initially assessed using digital airborne videography acquired over Rondonia in June 1999 using methods similar to those used by Hess et al. [2002]. Initial accuracy assessment was applied to three of the scenes, Ariquemes, Ji-Parana and Luiza. Several seconds of video frames were accessed for each sample to construct a mosaic. Flight-logs, recording aircraft location and altitude, were used to overlay a 3 by 3 30 m grid on each mosaic and locate the mosaics on the Landsat data. Land cover, within the central box corresponding to a Landsat pixel, was described as percentages of herbaceous cover, short shrubs, tall shrubs, medium forest, tall forest, human construction (roads/buildings), water, soil and rock then translated into their corresponding image classes. Classified values for each sample site were extracted from the imagery and used to develop a confusion matrix [Richards, 1999] and calculate a kappa coefficient [Congalton and Mead, 1983]. Overall accuracy for the seven classes was calculated as 85.4% with a kappa of 0.761. Rock, savanna, water, soil and urban were rarely encountered in the videography and thus have insufficient samples to assess accuracy of these classes. Second growth slightly exceeded a minimum standard of 19 samples [Richards, 1999], while pasture and primary forest were well represented. Producers accuracy, defined as the proportion of reference points mapped correctly by the classifier ranged from highs of 90% for pasture and 89.7% for mature forest to a low of 55% in Second growth. Users accuracy, defined as a measure of the proportion of pixels that are correctly classified to the total number mapped by the classifier, was 98.1%, 85.7% and 39.3% for mature forest, pasture and second growth. Confusion was greatest between second growth and mature forest. Detailed analysis of these samples showed that confusion occurred primarily in areas of rugged topography on Sun facing slopes where decreased shade resulted in the Decision Tree Classifier interpreting these areas as second growth. The second most significant source of error was in discriminating second growth from pasture, accounting for the largest error in mapping second-growth reference sites. Overall, second growth was overmapped relative to the other classes, primarily due to errors on sunlit slopes.

Full mosaic Accuracy Assessment (1984 to 2010 product).

A second round of accuracy was assessed using Google Earth imagery. This accuracy was applied to the revised land-cover product, in which the sunlit slope error was corrected (see below). Representative samples of each class were identified in Google Earth imagery acquired closest to 2009, the reference year used for accuracy assessment (See Table below). A total of 267 polygons were identified, sampling a total of over 250,000 pixels. 500 pixels were selected randomly for each class, resulting in 2812 samples. Some of the pixels were thrown out because they were assessed to be poor samples based on the Google Earth imagery.

Table listing the number of polygons , pixels and randomly sampled pixels for each class.

Class Polygons Pixels Samples

Burn 62 7817 444

Urban/Soils 43 5176 469

Pasture 68 22479 499

Mature Forest 29 145919 500

Secondary Forest 53 8363 402

Water 12 67201 498

An error matrix was created comparing reference (Google Earth) to mapped (mosaic) land-cover. Samples ranged between 402 and 500 pixels per class. Using this approach, overall accuracy of the mosaic was assessed as 90.54% with a kappa coefficient of 0.866. The most significant confusion was observed between burn and pasture. However, it should be noted that burned areas are typically pastures, so these would normally be considered the same class.

It should also be noted, that the mosaic submitted has not been corrected for phenological errors, in which some early dry season pastures are misclassified as secondary forest. Thus, when interpreting an early dry season image, some of the transitions to secondary forest are likely to be incorrect.

Process Description:

Data Acquisition Materials and Methods:

Landsat TM scenes were assembled into a comprehensive time series mosaic including seven landsat scenes acquired between 1984 and 2010. In total 189 landsat scenes were processed to produce mosaics for all 27 years (see table above for specific image dates). Landsat data were initially coregistered to 1998 or 1999 georectified digital PRODES data supplied by the Instituto Nacional de Pesquisas Espaciais [INPE, 2000]. Landsat data were georectified using between 30 and 40 tie points and rubber sheet stretching. All images were resampled using nearest neighbor resampling. The spatial accuracy of this product was assessed in the field and found to have some minor errors. To correct this, all subsequent data sets were georectified to Geocover ( The general procedure was to use ERDAS Imagine AutoSync to warp the NIR band of the input image to the Geocover base map NIR band. We used a second order polynomial fitting. This would generally generate 1000+ points. These 1000 points were subsequently pared down to a 15x15 regular grid using a script that selected the GCP within the grid cell with the lowest RMSE (225 points in all). In rare instances where AutoSync failed to produce adequate points, ~ 50 GCPs were manually selected. To ensure that all data sets were warped to a common base map, the older 1990s PRODES-based products were rewarped to Geocover using the same general procedure.

Images were classified into 7 categories:

1. Primary upland forest, representing the dominant natural vegetation in the area, categorized as dense tropical forest [RADAMBRASIL, 1978].

2. Pasture and green pasture, dominated by several pasture grass species (Brachiaria brizantha and Panicum maximum) and ranging in quality from highly degraded to well-managed green pastures. Recent burn scars are classified as pasture.

3. Second growth, dominated by small trees and shrubs with low species diversity and biomass relative to primary forest. Second growth may follow pasture after abandonment, or after anthropogenic or natural disturbance of primary forest.

4. Soil/urban.

5. Rock/savanna. Rock is most abundant in areas of high topographic relief while Savanna is commonly located in close proximity to wetlands.

6. Water.

7. Cloud obscured, including smoke from burning, clouds and cloud shadows.

Land-cover change was mapped using a multistage process described by Roberts et al. [1998; 2002]

1. Using spectral mixture analysis, a spectrum consisting of radiance reflected off of multiple materials within the field of view is decomposed into fractions of several unmixed spectra, called end-members [Adams et al., 1993]. Initial candidates for green vegetation (GV), Non-photosynthetic vegetation (NPV, stems, branches and litter), soil and shade are selected from the image then evaluated based on the fit (measured by a Root Means Squared Error (RMS)) and fraction images. If necessary, image endmembers are revised to improve the fit and reduce fraction errors (physically unrealistic fractions) [Adams et al., 1993]. The final product is five images, one for each end-member and a RMS image.

2. To identify surface materials and compare satellite observations to laboratory or field measured spectra it is often necessary to convert encoded radiance to apparent surface reflectance either through absolute calibration [Kaufman, 1989] or relative reflectance retrievals [Elvidge and Portigal, 1990]. Because of the diversity of data sources and lack of consistency in radiometric calibration, we used the modified empirical line approach described by Smith et al. [1990]. Using this technique, encoded radiance was regressed against laboratory and field measured reflectance from soils, water and NPV collected in Rondonia and Manaus between 1991 and 1992. Candidate models for converting encoded radiance to reflectance were assessed based on the shape of the intercept term (which should resemble a path radiance spectrum) and retrieved surface reflectance, which must be physically reasonable (i.e., nonnegative) and match expected reflectance for known targets in the image, such as water. Spectra measured using the Airborne Visible Infrared Imaging Spectrometer of broadleaf deciduous and broadleaf evergreen forests in North America were used as spectral proxies for vegetation because no canopy level spectra were available for Brazilian forests covering the full spectral range of Landsat TM. The quality of retrieved surface reflectance for Rondonia was evaluated by comparison to surface reflectance measured over similar targets in Manaus, retrieved using multiple ground reflectance measured in the field during the summer the Manaus Landsat TM data were acquired.

3. Reference end-members are spectra of known materials [Adams et al., 1993]. Ideally, an image end-member can be represented as a mixture of one or more spectrally pure, identifiable reference end-members. When selecting candidate reference end-members for each image end-member, the objective is to locate library spectra that are more extreme than the image end-members, provide a good fit (as measured by RMS) and fractions that match expected values based on field measurements or aerial photo interpretation [Roberts et al., 1998]. Reference end-members were selected from the same spectral library used to retrieve surface reflectance.

4. Following step three, the remaining data sets are standardized to the reference scene using relative radiometric calibration techniques [Schott et al., 1988; Hall et al., 1991; Furby and Campbell, 2001]. In this research 20 to 30 candidate invariant targets were manually located within each scene then used to convert encoded radiance, as measured on that date, to the equivalent of encoded radiance measured in the reference scene. Candidate invariant targets, primarily water, forest, second growth and urban areas, were initially selected by comparing scenes acquired over the same location over several years. Encoded radiance was extracted for each target and regressed against encoded radiance measured within the reference scene. Candidates that are not temporally invariant are readily identified as outliers in the regression and removed from the analysis, typically resulting in the loss of only a few of the sites. Typical r2 values range from lows of 0.85 to 0.9 for TM bands 1 and 2, to values exceeding 0.99 for TM bands 4, 5 and 7. A similar approach can be used for adjacent Landsat scenes using spatially invariant targets and overlap between scenes.

4.a. Modification to Roberts et al., 2002. One modification incorporated in the mosaic is the use of the Carlotto (1999) haze correction technique. This technique is based on the assumption that spatially variable haze impacts only TM bands 1 to 3. Whole scene statistics are derived identifying unique combinations of Landsat TM bands 4 to 7 (6) and the associated TM band 1, 2, and 3 values associated with these combinations. Scene averages for TM bands 1 to 3 are calculated for each unique combination of TM bands 4 to 7, then used to replace the original TM values for the first three bands. In essence, this approach “homogenizes� spatially variable haze across the image. The Carlotto has correction was only applied to images with clearly identifiable haze (smoke). After haze correction, the images were then processed through step 4 applying a relative radiometric calibration to the scene.

5. Once the entire time series has been intercalibrated and reference end-members selected, the same model can applied to the entire data set. This generates four fraction images for each scene (GV, NPV, soil and shade) and a RMS error image. Because each data set is standardized to a common reference and spectral library, spectral fractions can be compared directly between different scenes and across the entire time series.

6. Spectral fractions and the RMS error images are used to train a decision tree classifier (DTC) [Hess et al., 1995; Friedl and Brodley, 1997; Roberts et al., 1998; 2002]. The DTC was trained by extracting at least 100 samples for 6 of the 7 classes mapped (Rock/savanna were excluded from the initial training). To capture the within class variability and temporal variability in training sets, spectral fractions and RMS values were extracted from several scenes in Ariquemes, Ji-Parana, Luiza and PortoVelho that spanned the range of conditions present across the scenes. Decision rules were determined using Splus [Clark and Pregibon, 1992]. The final set of rules was derived after several iterations, including initial classification, followed by several additional stages of training and classification designed to reduce classification error at each stage.

7. Disallowed transitions represent any transition that is not physically reasonable such as a transition from second growth or pasture to primary forest over a span of a few years [Roberts et al., 2002]. Disallowed transitions can be screened by identifying them through time series, then reassigning the class of one of the pixels. In this research, simple rules were used to remove disallowed transitions by comparing three temporally contiguous dates and reclassifying pixels using the temporal median. Disallowed transitions include: 1. Pasture, urban, or second growth to primary forest 2. Water to any other class but water.

This approach is also effective for cloud screening, treating all transitions as disallowed. In the event that two transitions occur in combination with a disallowed transition, the pixel is not reclassified. A transition from pasture to primary forest then second growth is an example of this type of transition, leading to a disallowed transition (pasture to primary forest) remaining in the time series. To reduce the effect of pixel misregistration, a spatial filter is applied to classified images prior to time series analysis. In this research, a median filter was applied along a 3 by 3 moving window and all pixels classified as having one or less neighbors within the same class were reassigned to the median class within the window.

8. Areas covered by rock and savanna were mapped and all areas outside of the overlap zone between all dates within a scene were masked. A rock/savanna mask became necessary because these two land-cover types are difficult to separate from pasture with sufficient accuracy. To map rock/ savanna, a mask was developed for each scene then applied to the full time series for each scene. Edges were masked by determining the minimum area overlap for each scene within the time series and assigning all areas outside of the area of overlap to no-class.

Steps 1 to 8 generate the product described in Roberts et al., 2002. Subsequently several more steps were added to the processing to reduce misclassification of mature forest as secondary forest, either due to sun lit slopes or emergents. This processing involved:

9. Generation of an image showing all instances in the time series where a pixel was mapped as secondary forest. This processing was done for each of the seven Landsat scenes indepedently and was used to generate an initial mask of secondary forest.

10. Manual editing of the secondary forest mask to remove all instances where mature forest was identified incorrectly as secondary forest. Using a combination of time series images and pattern analysis, it is possible to identify regions that are clearly secondary forest, most often due to a regular spatial pattern (i.e., rectangular shape of patches). Using this approach, the secondary forest mask was manually edited to remove instances where areas mapped as secondary forest, did not have the spatial characteristics of secondary forest (at any time), and thus were mature forest. One mask was generated for each landsat scene (seven total).

11. Once a mask was completed for each of the seven scenes, it was applied to each scene in the time series to reset areas misclassified as secondary forest to mature forest. This product was then mosaicked using the ENVI mosaicking tool to generate 27 mosaics for Rondonia, one for each year from 1984 to 2010 consisting of seven Landsat scenes each.


Adams, J.B., M.O. Smith and A.R. Gillespie. 1993. Imaging spectroscopy:Interpretation based on spectral mixture analysis, in Remote Geochemical Analysis: Elemental and Mineralogical Composition, edited by C. M.Pieters and P. Englert, pp. 145-166, Cambridge Univ. Press., New York.

Carlotto, M.J. (1999). Reducing the effects of space-varying, wavelength-dependent scattering in multispectral imagery. Int. J. Remote Sens., 20(17), 3333-3344.

Clark, L.A. and D. Pregibon. 1992. Tree-based models, in Statistical Models in S, edited by J. M. Chambers and T. J. Hastie, pp. 377–420, Wadsworth, Belmont, Calif.

Congalton, R.G. and R.A. Mead. 1983. A quantitative method to test for consistency and correctness in photointerpretation, Photogramm. Eng. Remote Sens., 49(1), 69-74.

Elvidge, C.D. and F.P. Portigal. 1990 Change detection in vegetation using 1989 AVIRIS data, in Proc. SPIE Imaging Spectroscopy of the Terrestrial Environment, Orlando, Fla., 16–17 April, edited by G. Vane, pp. 178-189, Int. Soc. for Opt. Eng., Bellingham, Wash.

Friedl, M.A. and C.E. Brodley. 1997. Decision tree classification of land cover from remotely sensed data, Remote Sens. Environ., 61, 399-409.

Furby, S. L., & Campbell, N. A. (2001), Calibrating images from different dates to “like-value� digital counts, Remote Sens. Environ., 77, 186–196

Hall, F.G., D.E. Strebel, J.E. Nickeson and S.J. Goetz. 1991. Radiometric rectification, toward a common radiometric response among multidate, multisensor images, Remote Sens. Environ., 35, 11-27.

Hess, L.L., J.M. Melack, S. Filoso and Y. Wang. 1995. Delineation of inundated area and vegetation along the Amazon floodplain with the SIR-C synthetic aperture radar, IEEE Trans. Geosci. Remote Sens., 33(4), 896-904.

Hess, L.L., et al. 2001. Geocoded digital videography for validation of land cover mapping in the Amazon Basin, Int. J. Remote Sens., 23(7), 1527- 1555.

INPE.2000. Monitoramento da floresta Amazonica Brasileira por satelite 1998-1999, 22 pp., Inst. Nac. de Pesqui. Espaciais, Sao Jose Dos Campos SP,Brazil.

Kaufman, Y.J. 1989. The atmospheric effect on remote sensing and its correction,in Theory and Applications of Optical Remote Sensing, edited by G. Asnar, pp. 336-428, John Wiley, New York.

Richards, J.A. 1999. Remote Sensing Digital Image Analysis: An Introduction, Springer-Verlag, New York.

Roberts, D.A., G. Batista, J. Pereira, E. Waller and B. Nelson. 1998. Change identification using multitemporal spectral mixture analysis: Applications in eastern Amazonia, in Remote Sensing Change Detection: Environmental Monitoring Applications and Methods, edited by C. Elvidge and R. Lunetta, pp. 137-161, Ann Arbor Press, Chelsea, Mich.

Roberts, D.A., I. Numata, K. Holmes, G. Batista, T. Krug, A. Monteiro, B. Powell, and O.A. Chadwick. (2002) Large area mapping of land-cover change in Rondonia using multitemporal spectral mixture analysis and decision tree classifiers. Journal of Geophysical Research-Atmospheres, Vol. 107, No. D20, p. 8073.

Schott, J., C. Salvaggio and W. Volchok. 1988. Radiometric scene normalization using pseudoinvariant features, Remote Sens. Environ., 26, 1-16.

Smith, M.O., S.L. Ustin, J.B. Adams and A.R. Gillespie. 1990. Vegetation in deserts, I, A regional measure of abundance from multispectral images,Remote Sens. Environ., 31, 1-26.


NASA logo
Get Acrobat Reader