In: Wagner W., Szflcely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
draw the landscape context for all the ground control points that 
were mainly acquired for thematic purposes (as opposed to 
geodetic). These points were essential for understanding the 
different environments and their respective context. It is during 
the February campaign that the test area presented here was 
selected for having the largest possible number of different 
natural environments within the smallest area. The selected area 
covers 1.44 km 2 and is illustrated in Figure 2. 
2.3 Image Aquisition and Pre-processing 
One fusionned Ikonos image 1 registered to a UTM 
projection of the central part of the PWS was provided by the 
Forest Institute of Minas Gerais (Figure 2). This image was 
acquired in September 2006 corresponding to the end of the dry 
season when the water level is at its lowest. Given the flatness 
of the relief and geodetic ground control points, registration 
errors were kept below 5 meters. 
2.4 Texture Pre - processing 
The Ikono’s red band was selected to create the texture features 
due to the high reflectance response of vegetation in this 
channel. This band was used as input data in the program 
MACOOC (Philippe Maillard ©2010) which is able to produce 
all five texture measures in four directions: 0°, 45°, 90° and 
315°. Although data was produced in four directions, we 
analysed this information grouped. 
2.5 Gray Level Co-occurrence Matrix (GLCM) Tests 
Initially, we selected a few parameters (pixel pair lag distance, 
window size and texture features) to conduct the first tests in 
order to stipulate the best value of distance. 
The following values were chosen for lag distance: 3, 4, 5, 6 
and 7. We did not use values 1 and 2 because of the high 
correlation between neighbouring pixels (especially considering 
the image resolution results from a fusion). A value of 7 was 
fixed as the maximum given that aquatic plants present 
homogeneous groups with little or no projection of shadows and 
the average object size is rarely larger than 7 meters. 
Window sizes of 15, 21, 25, 31, 35 and 41 were selected. A 
maximum window size of 41 was selected once most of the 
features present in the study do not exhibit homogeneous areas 
larger than 1.600 m 2 . We opted for the most commonly 
parameters used: Angular Second Moment (ASM), Entropy 
(Ent.), Constrast (Cont.), Inverse Difference Moment (IDM), 
and Correlation (Corr.) 
Supervised classification method, which demands ground truth 
data, were performed using the maximum likelihood decision 
rule (Biehl and Landgrebe, 2002). Areas selected as training 
and validation samples were chosen based on field data and 
image interpretation. A few tests were performed to evaluate 
the effect of window size on the variance of the class. We chose 
use windows of about 11 x 11 pixels or 121 m 2 in the 
classification process. The class of bare ground was the only 
exception and windows of 9 x 9 pixels were chosen to maintain 
the integrity of the samples considering the difficulty of having 
"pure" samples of bare soil. 
1 Fusionning involves resampling the 4 m multispectral to 1 m 
using the panchromatic channel. 
After apply the Supervised method to images compound by 
texture and spectral bands, a five steps Knock-out process (table 
01) was performed in order to reduce the features types and 
frequencies that play the most important role in the 
classification process. All the combinations that achieved equal 
or higher values than 80% of overall accuracy were ranked 
and analyzed. 
In a second step, tests combining the five variables of texture 
with 18 different sizes of window 11, 13, 15 .... 45) were 
conducted using only the value 3 for distance. We submit data 
to classification and Knock- out process as previously 
mentioned. 
The third step had consisted in apply the classification to each 
parameter of texture separately and combined with the spectral 
bands. All the results higher than 80% were used to find the 
best window size. 
wind + dist 
15 + 04 
Tex+S 
78,1 
Cont 
ASM 
IDM Ent 
Corr 
Spectra 
1 
79,4 
78,4 
79,9 76,1 
77,3 
59 
81,3 
79,8 
78,4 
77,8 
60,3 
78,9 
77,8 
79,8 
65 
82,5 
80,5 
40,1 
Table 1. Knock - out example. Legend: wind+dist (window + 
lag distance), Tex+Spec (Texture + Spectral bands), Constrast 
(Cont.), Angular Second Moment (ASM), Inverse Difference 
Moment (IDM), Entropy (Ent.), and Correlation (Corr.). 
2.6 Validation 
Access in wetland areas can be very difficult and a fully 
systematic or random sampling scheme was impossible. 
Additionally, although the PWS is a protected area, most of it is 
still privately owned and we were not always able to have 
permission of access from landowners. Still, we were able to 
visit a total of 72 sampling sites chosen from the interpretation 
of image data to serve as training and validation data. To 
overcome the access limitations, we also used a micro-light 
aircraft flown at low altitudes (< 500 m) to acquire over 700 
oblique photographs of the area using a digital camera (Nikon 
D40X) equipped with a zoom (Nikkor 18-200 mm 1:3.5-5.6). 
Data from the tracking log of a navigation GPS (Global 
Positioning System) set at a 50 m distance interval was coupled 
with the acquisition time of the photographs to account for the 
aircraft position at each shot (the camera and the GPS were 
previously synchronized). The level of detail on these 
photographs was such that the dominant plant families could 
easily be interpreted with the help of two botanists and the 
ground samples. 
3. RESULTS AND DISCUSSION 
This paper results are presented in three blocks: Distance 
definition, window definition and behaviour of textures 
features.