IAPRS & SIS, Vol.34, Part 7, "Resource and Environmental Monitoring", Hyderabad, India,2002
Finally, for the sample A window (24025m?), Avicennia
schaueriana Stapf & Leechman coverage was estimated in
14992m? (62,496) and canopy gaps were estimated in 9033m?
(37.6%). Figure 1 shows spatial distribution of photointerpreted
features.
Figure 3 shows the sample A window final data fusion image
obtained by application of Kraus and Albertz technique. In this
figure the thematic differentiation improvement between
Avicennia schaueriana Stapf & Leechman mangle trees (dark
green areas) and canopy gap (light green areas) can be seen.
Figure 1: Field-truth representation of the sample A window.
Green areas are Avicennia schaueriana Stapf & Leechman trees
and yellow areas are canopy gap. Scale ~1:3700.
2.3.2. Supervised classifications: Main goals in this fourth
stage were determination of most accurate classification method
when compared field-truth and sample A window, and
automatic classification of Avicennia schaueriana Stapf &
Leechman and canopy gap with relative coverage determination
in sample B window.
Four methods were applied for supervised classification:
parallelepiped, minimum distance to means and maximum
likelihood classification. Six Avicennia schaueriana Stapf &
Leechman and twelve canopy gap training sites were used for
signature development both in sample windows A and B.
Software histogram tools allowed the determination of total
number of pixels per class.
Parallelepiped procedure used a z-score of 1.96. Raw distance
and no limits in search distance were used for minimum
distance to mean procedure. An equal 0.5 probability for each
feature and all pixels classification were used in maximum
likelihood routine.
3. RESULTS
Fused compositions shown improved outputs for an a priori
visual photointerpretation. Differentiation between canopy gap
and Avicennia schaueriana Stapf & Leechman trees was
optimised when conventional visual photointerpretation was
considered.
Figure 2 (left) shows visible high spatial resolution sample A
window from aerial photograph taken on November 2000 and
(right) 900nm infrared window from image obtained on
September 2001. These images were used as input data for data
fusion techniques. The 900nm infrared window was used as
brightness channel for the new data fused image while tone and
saturation channels were extracted from visible high spatial
resolution aerial photograph taken on November 2000.
Figure 2: Sample A window. Visible high spatial resolution
window from aerial photograph taken on November 2000 (left)
and 900nm infrared window from image obtained on September
2001 (right) used for data fusion. Scale ~1:3700.
Figure 3: Sample A window. Data fusion image obtained by
Kraus and Albertz technique. Dark green areas are Avicennia
schaueriana Stapf & Leechman mangle trees and light green
areas are canopy gaps. Scale ~1:3700.
Figure 4 shows results of parallelepiped, minimum distance and
maximum likelihood classifications based on data fusion image
for sample A window. Green areas are Avicennia schaueriana
Stapf & Leechman trees and yellow areas are canopy gap. As
can be seen, parallelepiped classification was the poorest
thematic classifier under similar given techniques conditions,
while minimum distance and maximum likelihood
classifications show homogeneous outputs.
Using data fusion image as analysis base, 33.3% of the pixels in
sample A window, and 14.3% of the pixels in sample B
window were no classified when parallelepiped procedure with
z-score of 1.96 was used. So parallelepiped classification
results were not considered due to expressive percentage of
non-classified pixels. As minimum distance to means is
commonly applied when the number of pixels used to define
signatures is very small better than maximum likelihood,