ADAR
ata. We
forest..
texture
at data,
nean of
JERS-1
texture
ifferent
easures
razilian
1). The
average
23,6°C
m late
rates in
1980).
ie local
o field
issance
os. On
y road,
mn. We
4. Land Cover Map
Using multi-temporal Landsat TM imagery from 1986 to 1996
we classified 69 test sites according to the legend found on
Table 1. Ground truth information, aerial photos and GPS
coordinates were used on this task. Figure 2 presents test sites
contour superimposed on Landsat TMS band for year 1996.
Table 1 also summarizes the number of sites and total area per
cover class.
Table 1: Survey fields summary for six studied cover classes
#of | pixels | Area in
these sites were classified into one of the following class: RD:
recent deforestation; PA: pasture or bare soil; Y2F: young
secondary forest; I2F: intermediate secondary forest; O2F: old
secondary forest and MF: mature forest. All pixels belonging to
each cover classes were exported to SAS - Statistical Analysis
System - for statistical treatment. We computed Mahalanobis
distances (D2) values for each SAR image, texture
measurements and windows size, taking cover class two-by-two
to a total of fifteen pairs (Table 3). The criteria for selecting the
window size for each SAR image was the maximization of the
Mahalanobis distances.
Table 2: List of herein used Haralick’s texture measures
Cover Class sites | per class | hectare
Recent Deforestation (RD) 5 6448 580.3 # Abbreviation Description
Pasture of bare soil (PA) 17 12883 1159.5 01 mean Mean
Young Secondary Forest (Y2F) 11 5109 459.8 02 vari Variance
Intermediate Second. Forest (I2F) 9 4872 438.5 03 ener Energy
Old Secondary Forest (O2F) 23 12927 1163.4 04 COIT Correlation
Mature Forest (MF) 4 30557 2750.1 05 entr Entropy
Total| 69 72796 6551.6 06 cont Contrast
07 homo Local homogeneity
08 diss Dissimilarity
5. JERS-1 and Radarsat Data Set 09 smea Mean of sum vector
10 svar Variance of sum vector
In this work we used one JERS-1 (LHH) and one Radarsat 1 Sent Entropy of sum vector
(CHH) amplitude images. The date of imaging were July 11 and 12 dmea Mean of difference vector
August 19/1996 for JERS-1 and Radarsat, respectively. These 13 dvar Variance of difference vector
images presents 12.5-meter pixel spacing. 14 dent Entropy of difference vector
6. Digital Image Processing
JERS-1 and Radarsat data were converted from 16 to 8 bits to
save disk space and to make computer processing easier. The
ENVI system was used for this task. Using Landsat 1995 data
as our base image, SAR data were resampled to 30 meters of
spatial resolution, using a linear transformation and nearest
neighbor sampling. A visual checking was made to assure a
minimum of border effect. After that a subset of 800x600 pixels
was selected for texture features computation.
7. Haralick Texture Computing
Haralick's texture measurements based on Grey-Level Co-
ocorrence Matrices (GLCM) have been the most used for
extraction of textural information from satellite (Haralick ef a/.,
1973). In this work fourteen Haralick texture measures were
tested for land cover mapping (Table 2). Soares et al. (1997)
made an extensive review of all herein used textures.
To compute all texture images we used five moving
windows sized 5x5, 7x7, 9x9, 11x11 and 15x15 pixels.
So, we generated 14 (textures) x 5 (windows size) images
for each SAR amplitude image, totalizing 70 texture
images for JERS-1 and 70 texture images for Radarsat.
This 140 texture images were normalized to have mean of
zero and unitary standard deviation.
8. Data Processing and Texture Features Extration
Our 69-test-site contours were digitized over the Landsat 1996
1mage, taking into account the border preservation. Each one of
Table3: List of fifteen paired cover class
Cover
Classes Description
1| RD PA | Recent Deforestation with PA
2 a" Y2F | Recent Deforestation with Y2F
3l = IF | Recent Deforestation with DF
4j.“ O2F | Recent Deforestation with O2F
$0" MF | Recent Deforestation with MF
6 | PA Y2F | Pasture with Y2F
Fos DF |Pasture with DF
8}. * O2F | Pasture with O2F
9j ^ MF | Pasture with MF
10| Y2F | DF | Young secondary forest with DF
Hi O2F | Young secondary forest with O2F
12 MF | Young secondary forest with MF
13| DF O2F | Intermediate secondary forest with O2F
141-7 MF | Intermediate secondary forest with MF
15| O2F | MF | Old secondary forest with MF
9. Results and discussion
Based on D2 (Mahalanobis Distance) values for paired cover
classes we selected 11x11 and 15x15 pixel window sizes for
JERS-1 and Radarsat data, respectively. Figure 3 presents D2
values for the best discriminated paired cover classes, for both
sensors data.
By analyzing D2 values for all 15 paired cover classes, we
selected these texture measurements that were good for
discriminating most paired cover class. For JERS-1 data the
best texture measurements were mean, variance, entropy,
contrast, local homogeneity, and mean of sum vector: for
Intemational Archives of Photogrammetry and Remote Sensing. Vol. XXXII, Part 7, Budapest, 1998 393
