Longterm monitoring of natural and anthropogenic change in a neotropical rainforest using remote sen

1
Long-term monitoring of natural and anthropogenic
change in a neotropical rainforest using remote
sensing imagery Jonathan Greenberg, Center for
Spatial Technologies and Remote Sensing (CSTARS),
University of California, Davis
Results Linear spectral unmixing of the LANDSAT
imagery resulted in predictable endmember
fractions. Primary rainforest had a much higher
and more variable shadow fraction than did all
other image classes (Figures 1a and 2a). Within
disturbed areas, the vegetation fraction was, as
expected, high in areas of successional forest
(Figure 2b) and agriculture, and low in cleared
fields and roads (Figure 2b). By examining the
shadow fraction from 1986 and 1996 (Figures 3a
and 3b), the expansion of the region around the
Vía Auca has increased dramatically in a short
period and seems to follow a pattern of spiral
colonization into untouched regions from the main
highways into the forest. Discussion Mechanistic
approaches to the analysis of image spectra like
linear spectral unmixing have many benefits over
statistical techniques. Even for initial
analyses, like this one, meaningful results can
be generated with little or no ground truthing.
As long as the images are properly calibrated and
the endmembers well chosen, realistic relative
fractions of shadow, green and non-photosynthetic
vegetation and soil can be determined. Further
refinement of the data is necessary, and true
classification using endmember thresholds remains
to be performed, as well as increasing the time
scale of the analysis to nearly 35 years of
change. Other mechanistic techniques, such as
geometric canopy modelling will also be employed
to provide a link between successional models,
which often output important structural
information about forests in different states,
and the remote sensing data which is directly
related to these variables. Image spectra can
then be placed in context of the ecological
dynamics which spawned them, instead of being
simply statistically lumped into what are often
somewhat arbitrary classifications. Landscape
scale examination of human disturbance (such as
Figures 3a and 3b) will hopefully reveal
predictable patterns of colonization and
community growth through tropical forests.
Finally, results from analysis of the ecological
and human dynamics will be used to generate a
model which will be used to test the effects of
current and alternative land use scenarios on the
future sustainability of this forest. References
Adams, J. B., D. E. Sabol, V. Kapos, R. Almeida
Filho, D. A. Roberts, M. O. Smith, and A. R.
Gillespie. 1995. Classification of multispectral
images based on fractions of endmembers
Application to land-cover change in the Brazilian
Amazon. Remote Sensing of Environment
52137-154. Elvidge, C. D. 1990. Visible and near
IR reflectance characteristics of dry plant
materials. International Journal of Remote
Sensing 111775-1796. Holmes, B. 1996. The low
impact road. New Scientist 15140. Pitman, N. C.
A. 2000. A large-scale inventory of two Amazonian
tree communities. Dissertation. Duke University,
Durham. Roberts, D. A., G. T. Batista, J. L. G.
Pereira, E. K. Waller, and B. W. Nelson. 1998.
Change identification using multitemporal
spectral mixture analysis applications in
Eastern Amazonia. Pages 137-161 in R. S. Lunetta
and C. D. Eldridge, editors. Remote sensing
change detection environmental monitoring
methods and applications. Ann Arbor Press,
Chelsea. Contact Information and
Acknowledgements Please feel free to email me any
questions or comments greenberg_at_ucdavis.edu, or
visit my labs website at http//www.cstars.ucdavi
s.edu. Special thanks to Jessica Gorin, Shawn
Kefauver, George Scheer, Hugh Stimson and Corey
Yeaton for assistance in preparing this poster.

• Introduction
• Tracking long-term changes rising from both
  natural and anthropogenic disturbances in a
  tropical rainforest can be both costly and
  difficult to quantify and apply across a
  landscape scale. Remote sensing provides a clear
  tool to examine these processes across large
  spatial and temporal scales. Here I present some
  initial results from a change detection project
  in the rainforests of eastern Ecuador which will
  ultimately span nearly 35 years of satellite data
  and provide the base for a mechanistic model of
  the ecological and economic dynamics at work in
  one of the most diverse forests on record (Pitman
  2000).
• Background
• Parque Nacional Yasuní is the largest national
  park in Ecuador, and along with the adjacent
  Huaorani Reserve, covers over 1.6 million
  hectares of largely undisturbed lowland
  rainforest, floodplain forest and swamp habitat.
  Home to a reported 2488 species of plants, this
  park is one of the most floristically diverse
  forests on record (Pitman 2000). To the west of
  the park, the now infamous Vía Auca oil access
  road was constructed in the early 1980s which
  allowed a massive influx of people from the
  surrounding degraded farmlands. In the early
  1990s, a second road was constructed through the
  national park, originally touted as the
  low-impact road (Holmes 1996) it has, in the
  past decade, seen a lesser but significant influx
  of people despite supposed safeguards to prevent
  what is rapidly becoming a major ecological
  disaster. More recently, a new oil access road
  is being proposed which will pass through another
  large and untouched region of the park.
• Methods
• One of the chief difficulties in analyzing remote
  sensing imagery of a rainforest lies in the huge
  variation in spectral signature which can arise
  from image pixels a researcher would label as
  primary rainforest. This variation makes most
  statistical approaches to sub-forest
  classification (for instance, successional state)
  extremely prone to error. For these initial
  results, I used a more mechanistic approach to
  analyzing spectra known as linear spectral
  unmixing (Adams 1995). This technique relies on
  the fact that a single pixel is likely to be
  composed of multiple, spectrally distinct
  materials (hereafter referred to as endmembers)
  and can determine the sub-pixel composition of
  these materials. For forests, the major spectral
  endmember components are green vegetation (GV),
  non-photosynthetic vegetation (NPV, including
  dead leaves, branches and bark), soil and shadow.
  
• For this initial analysis, I used two LANDSAT TM
  images dating from 1986 and 1996. The 1996 image
  was calibrated to reflectance using published
  gain and offsets, and the 1986 image was
  calibrated to the 1996 image using temporally
  invariant pixels and the empirical line technique
  (Adams 1995). The endmembers were taken from a
  published library (Elvidge 1990) and used to
  generate images of endmember fractions. Finally,
  the 1986 image was georegistered to the 1996
  image to within 0.5 pixel accuracy. For these
  preliminary results, I tested whether image
  classes would have predictable endmember
  fractions based on Adams et al. (1995). These
  classes include
• Primary rainforest large and variable shade
  fraction due to highly variable canopy structure,
  high GV, medium NPV and little or no soil.
• Successional forest (Cecropia dominated) low
  shade fraction due to low variability in the
  canopy heights, high GV.
• Clear cuts and roads low shade fraction, low GV,
  high NPV and soil components.
• Agriculture low shade fraction, high GV, NPV and
  soil.

River
a
a
b
Soil/road
Primary forest
Successional forest
Agriculture
b
Figure 1 Estación Científica Yasuní. (a) Shade
fraction, (b) Vegetation fraction. Successional
forest has about the same vegetation fraction,
but a much different shade fraction than primary
forest.
Figure 2 Small stretch of the Vía Auca. (a)
Shade fraction, (b) Vegetation fraction. Note
how human disturbed areas are uniformly
unshadowed, but agriculture has a much different
vegetation fraction than roads and cleared fields.