RS, Vol. XXXVIII, Part 7B
In: Wagner W„ Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Voi. XXXVIII, Part 7B
29
> METHODS
2008) was carried out to
ta. The group from the
aboratory” (UNIMONTES
:en meter plots, along both
plots were at a distance of
x and are always oriented
l 10 m gap is always left
j area was too degraded to
ots 1-35 are located on the
right bank. A total of 7000
if tree height and stem
1H) were taken for all trees
isses were not considered,
aken in each plot for later
ss and leaf area index,
rements were produced: (i)
breast height (DBH), (iii)
nsity, (vi) canopy openness
were conducted in January
/, to obtain ground control
S was employed to acquire
lying the Ikonos scene and
processing
study was provided by the
i. It was obtained with their
i and blue = 4m) and
to a spatial resolution of 1
mber 2007 during the dry
vith no cloud cover (Figure
a UTM grid coordinate by
mean square error (RMSE)
c correction was applied to
ir steps: 1) cartographic
ation, 3) texture feature
fling.
onsisted in using spatial
i to areas having a strong
ian vegetation (Maillard et.
onfusion between vegetation
, such as: palm swamps and
iphic network was digitized,
. to build a buffer of 50 m
l width in the study site is
ir was used to mask parts of
. The Ikonos image and the
entation algorithm which is
riparian vegetation e non-
2.3.2 Image Segmentation
The riparian vegetation was first visually interpreted in order to
validate the results of the segmentation. The MAGIC program
(Clausi et al., 2009) was chosen to segment the image due to its
excellent results reported in several studies (Maillard et. al., 2008;
Barbosa et al., 2009; Alencar-Silva and Maillard, 2009). MAGIC
is an acronym that means “Map Guided Ice Classification”
because it was originally developed as a tool for classification of
ice sea types. The segmentation of MAGIC is unique in its
implementation and the principles it embodies. It is an hybrid
segmentation approaches that uses two different approaches to
segmentation: “watershed” and Markov Random Field (MRF).
The segmentation is started by applying a “watershed” algorithm
that produces a preliminary segmentation and generates segments
(areas) of 10-30 pixels depending on the noise level in the image.
The smaller segments are then arranged topologically, so all
contiguous segments can be determined through an adjacency
graph or RAG (Region Adjacency Graph). The second step is
based on the MRFs that will join or not contiguous segments if
the union produces a decrease in the total energy of the
neighbourhood defined by Equation 1.
E = E f +aE r (1)
where: E f is the global spectral energy, E r is the local spatial
energy, a is normally a floating constant.
The advantage of the MRF model is its inherent ability to describe
both the spatial context location (the local spatial interaction
between neighboring segments) and the overall distribution in
each segment (based on parameters of distribution of spectral
values for example). That new approach was entitled “Iterative
Region Growing Using Semantics” or IRGS and is described in
Yu and Clausi (2008).
MAGIC is able to segment each band image individually or as a
multivariate data. In this study, the spectral bands were used both
as a multivariate dataset and individually. Three parameters have
to be specified for the segmentations to take place: (i) the number
of classes, (ii) pi, and (iii) P2. The number of classes varies
depending on how the user wants to segment the image.
For our study two categories were desired: riparian and non
riparian. However, because there are several different elements in
the non-riparian group (i.e. water, herbaceous, bare soil, grass,
etc), tests were performed with 3, 4, 5, 6, 7 and 8 classes. The best
result obtained by the MAGIC was to be used as a mask in the
texture calculations.
2.3.3 Image Texture Calculations
of 1 lxl 1, 15x15, 20x20, 25x25 and 30x30 pixels were used.
The distances between analysis pixels vary between 3 and 7
and the four directions: 0°, 45°, 90° and 315°.
A special program was created to compute the texture feature
to account for the use of a mask. MACOOC (Philippe
Maillard ©2010) takes an image and a binary mask as input to
compute all five texture measurements in all four directions.
Because the mask can adopt just about any shape, regular
texture extraction programs would have to discard the texture
computation for many riparian pixel when the analysis
window overlaps the zeros areas of the mask. MACOOC
compensates the “incomplete” windows by adjusting the
number of co-occurrences in order to compute comparable
probabilities. The probabilities values are then rescaled
between 0 and 10000.
Finally, the 70 plots were overlaid in the image. The average
values of the four spectral channels (blue, green, red and
infrared) and 20 texture bands were computed for each plot
and organized in a matrix along with the allometric data.
Multiple Regression using Stepwise feature selection was used
to analyze the data.
3. RESULTS AND DISCUSSIONS
The results of this study are presented in two blocks: image
segmentation and biophysical riparian forest modelling.
3.1 Image Segmentation Results
The best MAGIC segmentation was obtained using the image
as a multivariate dataset with all four Ikonos’ bands (Table 5).
Spectral Band
Riparian %
Non-Riparian %
Total %
1 (blue)
89.19
80.71
84.16
2 (green)
-
-
-
3 (red)
88.82
75.14
80.71
4 (infrared)
-
-
-
1, 2 and 3
91.28
85.01
87.56
1, 2, 3 and 4
88.31
90.61
89.68
Table 5: MAGIC overall segmentation success (average
user’s and producer’s) result for riparian and non-riparian
vegetation.
The best results were obtained with five classes and an overall
accuracy of 89.68% when compared with the visual
interpretation. This result takes into account both omission
and commission errors (Figure 4). Results obtained with the
green and infrared bands had very low correlation with the
interpreted image.
The texture of an image can be defined as changes in spatial
patterns of gray levels in a set distance (Tso and Mather, 2001).
An approach widely used in texture parameters calculation is the
Gray Level Co-occurrence Matrix (GLCM) (Lillesand and Kiefer,
2000). This method proposes that each element of the matrix is a
probability measure of occurrence between two gray levels
separated by a certain distance and direction (Haralick, 1979). In
this paper five features were considered: contrast (CON), angular
second moment (ASM), entropy (ENT), inverse difference
moment (IDM) and correlation (COR). The Ikono’s red and infra
red bands were chosen in order to calculate the five texture
features. The blue and green bands were not used because they
have strong correlation with the red band. Analysis window sizes
