Michael Palace1, Michael Keller1,2, Greg Asner3, Bobby Braswell1, Stephen Hagen1

1
Michael Palace1, Michael Keller1,2, Greg Asner3,
Bobby Braswell1, Stephen Hagen1
1 Complex Systems Research Center, Morse Hall,
University of New Hampshire, Durham, NH 03824 USA
2 International Institute of Tropical Forestry,
USDA Forest Service, Rio Piedras, PR 00928-5000
USA 3Department of Global Ecology, Carnegie
Institution of Washington, Stanford University,
Stanford, CA, USA
Abstract_
Forest canopy gaps resulting from natural tree
mortality and logging increase light in the
understory, release nutrients, and create
structural habitat for some species of flora,
fauna, and fungi. The measurement of gap
formation using remotely sensed data over broad
areas would allow foresters and ecologists to
study forest dynamics over greater areas than
those available from plot level surveys.
Previously, we developed a crown detection
algorithm that used high resolution (1 meter)
satellite image data. We have extended the
algorithm to examine logged forests and the
disturbances of such forests. Log decks and
canopy gaps have spectral signatures that can be
differentiated from surrounding trees in
multi-spectral 4 m resolution IKONOS images. By
combination of multi-spectral data with the
higher resolution panchromatic data from IKONOS,
our refined algorithm estimated gap size and
frequency and spatial patterning. Remote sensing
estimates of gap frequency and size will be
useful for understanding carbon budgets and fire
susceptibility in logged forests.
FIGURE 2. Cumulative frequency distribution for
field observed canopy diameters and automated
crown estimate at Cauaxi.
FIGURE 1. Digital number data used for
termination of an ordinate. The crown edge is
estimated to be 8 pixels from the local maxima.
Tree number 3
Algorithm_
FIGURE 3. Example of analysis of a 240 by 240
meter section of undisturbed tropical forest.
White circles are delineated crowns, black
polygons trace the outline of transects, pink
pixels indicate area of polygons.
Currently, there are two primary algorithms used
in high-resolution canopy automated analysis.
These two techniques are local maximum filtering
and local minima value finding. Local maximum
filtering has proven accurate in estimating the
number of trees using the assumption that the
brightest local value represents the
characteristic of a single crown (Wulder et al.
2000). Local minima value finding has been used
to detect the area between two crowns, using the
assumption that the darker image values are
created by shadows between crowns (Pouliot et al.
2002 ). Our algorithm combines local maximum and
minima finding methods. It introduces three new
concepts in crown detection analysis. These are
the analysis of iterative local maximum analysis,
a derivative threshold that ends ordinal transect
analysis, and the removal of previously analyzed
pixels from further analysis. Our algorithm
simultaneously estimates multiple canopy
structural parameters, rather than just the
number of trees per hectare.
Tree number 50

• Preprocessing
• Modal, maximum, and minimum brightness values
  (DN) found
• Moving window 3x3 averaging filter
• Crown Detection
• Local Maxima Analysis
• Brightness value (highest to modal brightness
  values examined in an iterative step)
• Local maxima seeds ordinate analysis
• Ordinate Analysis
• Ordinate analysis (series of DN values in
  straight line) radiates out in multiple
  directions from the local maxima or seeded pixel.
  Number of ordinates defined by user (64 in this
  study).
• End ordinate when the next pixel DN value is 2
  greater then current pixel
• Ordinate may not proceed into previously
  determined crowns
• Crown Determination
• Two longest opposite ordinates determined (crown
  width)
• Crown drawn as circle using radius of one-half of
  the longest crown width
• DBH and Biomass estimate conducted based on crown
  width
• Once a crown is determined, no new local maxima
  in that area may be analyzed

Study Area and Data_
Study Areas and Satellite Observations IKONOS
panchromatic images were used from two Amazonian
forests in Para, Brazil the Tapajos National
Forest, (3.08 S, 54.94 W) and the Fazenda
Cauaxi, (3.75 S, 48.37 W). Analysis was
conducted on undisturbed forests from both sites.
  Stand Data and Biomass Data Tree geometry data
from Asner et al. 2002 was used in analysis of
crown width distribution. Stand data collected
at Cauaxi 2000 was used for comparison with the
sensitivity analysis and used in parameterization
of the automated crown detection algorithm.
Stand data from Rice et al. in press and Keller
et al. 2001 were used for validation of stand
characteristics. Allometric Equations The crown
width to DBH allometric equation used here was
developed in Asner et al. 2002. The DBH to
biomass equation used here is from Brown et al.
1997. Multi-site Analysis We ran our automated
algorithm on 52 IKONOS data subsets, each
consisting of 1 km2 tiles from seven areas in the
Brazilian Amazon that are associated with LBA
(Figure 3). These data were selected from the
entire database of IKONOS LBA data and were
manually screened for clouds, cloud shadows,
water and cleared areas. All selected data
subsets represent undisturbed forest areas.
FIGURE 4. Seven IKONOS scenes were used to test
the algorithm for undisturbed forest areas
distributed across the Amazon region. The crown
detection algorithm was conducted on 52 tiles of
1 km2 (1Cauaxi 14 (tiles), 2Caxiuana 1,
3Jaru 2, 4Manaus 9, 5Mato Grosso 11,
6Santarem 67 10, 7Santarem 83 5).
2
Results
indicates a difference - indicates no difference
Crown Width
Figure 5. Comparison of field data and automated
crown detection for stand density and biomass
estimates.
DBH
Density
Figure 6. Comparison of data estimated at 7
LBA-ECO field sites for four forest structural
properties derived from an automated crown
detection algorithm using high resolution
satellite imagery.
Biomass
Figure 9. Comparisons between different LBA-ECO
sites using ANOVAs that utilize the crown
detection algorithm results for each site. All
comparison pairs use Tukey-Kramer HSD with an
alpha value of 0.05.
Summary
Figure 7. Comparison of Average Crown Width and
Areal Density derived from an automated Crown
detection algorithm. Two areas from two IKONOS
images.
Figure 8. Comparison of DBH Distribution derived
from an automated Crown detection algorithm.
Two areas from two IKONOS images.
We developed an automated crown detection
algorithm that uses high resolution satellite
image data. This algorithm combines local
maximum and minima finding methods. It
introduces three new concepts in crown detection
analysis. These are the analysis of iterative
local maximum analysis, a derivative threshold
that ends ordinal transect analysis, and the
removal of previously analyzed pixels from
further analysis. Our automated algorithm
provides useful estimates of tropical forest
structure over the vast undisturbed areas of the
Brazilian Amazon using high resolution satellite
image data. Our automated algorithm provided
estimates of crown width distribution and density
of trees. DBH and biomass estimations need
refinement because they were based on allometric
equations. It appears that with both crown width
and DBH distribution, our algorithm was
under-estimating small and large crowns/trees and
over estimating medium sized crowns/trees. This
over and under estimation is likely due to the
merging of smaller crowns, division of larger
crowns, and inability of remotely sensed data to
view smaller understory trees. A previous
manual interpretation (Asner et al 2002)
underestimated small to medium sized crowns and
overestimated large crowns. For the distribution
of crown widths, the automated algorithm results
were more similar to the field data than the
manual interpretation results were to the
field. We examined two IKONOS images of the same
area containing undisturbed and RIL logged
forest. Differences were found between the
undisturbed and logged sites within each image.
Temporal differences were also found for the same
area between the two images. Logged areas had
smaller average crown sizes and more trees. The
removal of larger trees for timber exposed
smaller canopies previously shaded by the large
trees. These initial results suggest that our
automated crown detection algorithm may be useful
for identification of logged areas and diagnosis
of logging damage and regeneration. The use of
the multi-spectral bands and an NDVI filter are
future endeavors that will aid in logged area
studies using IKONOS imagery.

• (Future Work)
• New Steps with Multi-Spectral IKONOS Bands
• Calculate NDVI
• Use NDVI as filter in Local Maxima Analysis (No
  seeding allowed in areas filtered
• Compare pixels within a determined crown for
  spectral similarity
• Mean and standard deviation for each crown for
  pan, ndvi and four spectral bands recorded
• Kmeans cluster analysis on crown information
• Examination of four new areas (repeat images of
  Undisturbed forest and Selectively Logged
  Reduced-Impact Logging at Tapajos

REFERENCES Asner, G.P. M. Palace, M. Keller, R.
Pereira Jr., J.N.M. Silva, and J.C. Zweede.
(2002) Estimating Canopy Structure in an Amazon
Forest from Laser Rangefinder and IKONOS
Satellite Observations. Biotropica 34(4)
483492. Brown, S., 1997. Estimating biomass and
biomass change of tropical forests A primer.
United Nations Food and Agriculture Organization,
1997. Keller, M., M. Palace, and G.E. Hurtt. 
2001. Biomass in the Tapajos National Forest,
Brazil  Examination of Sampling and Allometric
Uncertainties.  Forest Ecology and Management
154  371-382. Pouliot, D.A., D.J. King, F.W.
Bell, and D.G. Pitt, Automated tree crown
detection and delineation in high-resolution
digital camera imagery of coniferous forest
regeneration. (in press, Remote Sensing of
Environment 5718 (2002) Rice, A.H., E. H. Pyle,
S. R. Saleska, L. Hutyra, P. B. de Camargo, K.
Portilho, D.F. Marques, M. Palace, M. Keller and
S.C. Wofsy. (2004) Carbon Balance and Vegetation
Dynamics in an Old-growth Amazonian Forest.
Ecological Applications 14(4) pp.
s55-s71. Wulder, M. et al, (2000), "Local
Maximum Filtering for the Extraction of Tree
Locations and Basal Area from High Spatial
Resolution Imagery", Remote Sensing of
Environment, Vol. 73, pp. 103 - 114.
ACKNOWLEDGEMENTS We thank Greg Asner, Johan
Zweede, Rodrigo Pereira Junior and the foresters
and technicians of FFT for their high quality
work. Support for this research was provided by
the USDA Forest Service and the NASA Terrestrial
Ecology Program (NCC5-225).