International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
178 
The momentum term helps to update weights so that the error (a 
function of the weights) is global, not local and minimum 
(Fausett, 1994). In general, it is constrained to be in the range 
from 0 to 1, exclusive of the end points. This parameter changed 
from 0.00005 to 0.0005 depending on the training iterations in 
this study. The learning rate has a significant effect on an ANN 
performance. Increasing the learning rate as the ANN error 
decreases will often help to speed the ANN convergence so that 
the error quickly reaches a minimum. The learning rate is 
varying from 0.0005 to 0.005, also depending on the training 
iterations. 
000 
000 
000 
0.496 
000 
0.211 
Fig. 2. An ANN model with 7 input neurons, 10 neurons in the 
hidden layer and 9 output neurons in the output layer. 
Three ANNs were designed. ANN 1 was used for the textural 
indices extraction (see Section 2.2). It contains three neurons 
(three bands) in the input part, 18 neurons (eighteen spectral 
classes) in the output layer and 10 neurons in the hidden layer. 
ANN 2 was used for the landcover and landuse classification 
based only on spectral features. It is identical to ANN 1 with the 
exception of 9 output neurons. ANN 3 was used for the 
landcover and landuse classification combining spectral feature 
and textural features. It consists of 7 input neurons (three 
spectral bands and four textural measures), 10 neurons in the 
hidden layer, and 9 output neurons. 
2.4. Landcover and Landuse Classifications 
Three classification procedures were compared. Procedure 1 
uses only three spectral bands, procedure 2 is based on spectral 
bands and textural measures derived from matrix 1, and 
procedure 3 uses spectral bands and textural measures derived 
from matrix 2. Procedure 2 and procedure 3 were applied with 
textural window sizes of 3x3, 5x5, 7x7 and 9x9. So, the 
following nine classification results were produced: 
the classification based only on spectral features (termed 
cl), 
classifications using three spectral bands and four textural 
measures based on matrix 1 and window size 3x3 to 9x9 
(termed c2, c3, c4 and c5, respectively), 
classifications using three spectral bands and four textural 
measures based on matrix 2 and window size 3x3 to 9x9 
(termed c6, c7, c8 and c9, respectively). 
810 pixels (90 individual pixels per class) were chosen as 
training data, and 450 pixels (50 individual pixels per class) 
were chosen as test data. The pixels in the test data are non 
neighbouring and independent of the training data, and parts of 
them are mixed pixels (but dominated by one landcover or 
landuse) and isolated pixels. 
3. RESULTS AND ANALYSIS 
Fig. 3 (b), (c) and (d) illustrate classification results with 
procedure cl, procedure c3 and procedure c7, respectively. 
Table 2 shows confusion matrices derived for different selected 
classification procedures. Results of correctly classified pixels 
are summarised in Table 3. The overall classification accuracy 
with procedure 2 shows a small improvement compared to 
procedure 1. However, relatively significant improvements with 
procedure 2 are obtained for classes with complex spectral 
components or multimodal distribution, like parks, apartment- 
block areas, low density and high density urban areas and crop 
2. 
As illustrated in Table 3, a significant improvement in 
classification accuracy is obtained with procedure 3, e.g. from 
80.4% with procedure 1 to 94.2% with procedure 3 (see cl and 
c7 in Table 3). The most significant improvements are obtained 
in the classification of parks, apartment-block areas, low density 
and high density urban areas and crop 2. 
However, it can also be seen in Table 3 that an accuracy 
decrease occurs in the classifications of trees, grass and water, 
when textural measures are added to the three spectral bands. 
The reason may be that the randomness of textural measures at 
the boundary between different landcover and landuse classes 
results in the textural measures being noise. 
Table 3 shows that classification accuracy with the procedure 3 
is higher by about 10% than that with the procedure 2. The 
difference between procedures 2 and 3 is only the grey level co 
occurrence matrix for textural measure calculation. This 
indicates that the grey level co-occurrence matrix is an 
important factor affecting the classification accuracy. 
In addition, Table 3 also shows that an inappropriate window 
size can reduce the classification accuracy. This is because a too 
small window size can not effectively capture the textural 
features of landcover and landuse, while the too large ones may 
lead to inclusion of additional "noisy" pixels apart from the 
texture features.