edited in
ntaining
2s, Water
ed is the
t of map
. For the
t and 14
ees with
Lopes et
ferences.
nt of the
ving the
The six
Saic was
'es to be
iust refer
steps in
specially
ons. The
ry), map
Y andZ
oftware.
basis of
algorithm
anges in
a single
sed with
ormation
. For the
opriated.
> Brovey
6).
Sanden,
ults (van
yntextual
uated in
network
sification
accuracy
rmed the
Claudio Dupas
neural network algorithm (51.2%). However, the maximum likelihood gave the highest accuracy for the forest classes
using Landsat TM data. The Gaussian maximum likelihood is one of the most widely used algorithms for remote
sensing data classification. It was the one used is this study because of its reliability and for being easier to operate if
compared to the new classifiers described above. It assumes a normal distribution for the image values corresponding to
the training data, allowing the description of the category response pattern by the mean vector and the covariance matrix
(Lillesand and Kiefer, 1994).
The dendrometric (stand) parameters collected in the field, DBH and tree height, were used to stratify the forest training
and the forest test set in two.
The cloud cover (and the correspondent shadow) gap areas on the Landsat TM 95 were masked manually, pixel by
pixel. As a result, a bitmap layer delimiting the no-information areas was created (using PCI software) and added to
each image to be classified.
The errors in a classified image can be assessed by using an error matrix (also called confusion matrix) (Congalton and
Green, 1999). An error matrix compares the reference data (or ground truth), represented by the columns, with the
classified data, represented by the rows. The accuracy for a given class is represented by the errors of inclusion
(‘commission’ errors) and exclusion (‘omission’ error) of reference data units into the class (Congalton and Green,
1999). Similarly, the ‘producer’s’ accuracy is the ratio between the number of correctly classified units of a given class
and the total number of reference sampling units for this class, while the *user's' accuracy is obtained by dividing the
total number of correctly classified units by the total number of units (from all the classes) classified as this given class.
The matrix's major diagonal indicates the agreement between the reference data and the classified data. Thus, the
overall accuracy for a particular classified image is given by dividing the sum of the major diagonal's cells by the total
number of reference sampling units (Congalton and Green, 1999).
Supervised classification was performed for the Landsat TM 95 alone, the two SAR data alone, and the four different
fused Landsat TM 95 — SAR data using PCI ImageWorks. The maximum likelihood classifier was applied and the
bitmap mask was used to avoid the TM 95 no-information areas.
The Landsat TM 96 (with no clouds), and the two fused images that had previusly obtained the higher overall
classification accuracy for one forest class, TMJC and TMEC, were classified in PCI ImageWorks, using the bitmap
mask to classify only the area correspondent to the Landsat TM 95 no-information areas. The maximum likelihood
classifier was employed together with the training sets for one forest class. The classified TM 96 image was than used
as ‘ground truth’ information to generate tree test sets, using the stratified random sapling method.
Likewise the overall classification accuracy, the Kappa coefficient of agreement is a measure of accuracy. The Kappa
value, or “KHAT” is defined by the subtraction of the chance of agreement from the observed accuracy divided by 1
minus the chance of agreement. The equation can be found in Congalton and Green (1999).
4 RESULTS AND DISCUSSION
Figures 1a and 1c show a subset of the speckle filtered and rectified JERS-1 and ERS-2 images. For both images the
forest areas usually have a brighter tone than the non-forest. If the two images are compared to the Landsat TM 95
composite (Figure 1b), it can be noted that for the JERS-1 the contrast between forest and non-forest is higher than in
the ERS-2. The net of gallery forests which are visible on the Landsat TM is clearly identified on the JERS-1 but not on
the ERS-2 image.
The ERS-2 has short wavelength (C-band) and vertical (VV) polarisation results that the radar signal is dominantly
scattered by the forest canopy. The JERS-1 has comparatively longer wavelength (L-band) and horizontal polarisation
(HH). According to van der Sander (1997), the horizontally polarised radar waves penetrates deeper into a forest
formation than the vertical polarisation. Similarly, the longer the wavelength, the deeper it penetrates into the forest.
Therefore, it is expected that the JERS-1 signals can distinguish forest from non-forest (e.g. pasture) better than the
ERS-2. Recent studies using SAR imagery for tropical forest mapping give support to this idea. Using SIR-C SAR
imagery with different wavelenghs to assess deforestation in the Brazilian amazon forest, Rignot ef al. (1997) have
found that the L-band has higher capability to separate forest from clearings if compared with the C-band. Similarly,
Bijker and Hoekman (1996) observed that it’s difficult to discriminate forest from non-forest when single date ERS-1
(with the same sensor characteristics as the ERS-2) images are used.
Both images were affected in more or less extends by the complex landforms of the area. The ERS-2, having a
relatively low incidence angle (23 degrees), is very sensitive to terrain variations, which leads to shadowing (dark areas
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part Bl. Amsterdam 2000. 99