vegetati 
detectio 
Training Mode 
  
  
(b) 
: ; ; Figure 5. TerraSAR-X radar data (HH polarization) overlaid 
Figure 3. Block diagram of the slide detection algorithm with masks for training (a) and testing (b) 
Throughout the training phase, the errors are propagated from 
the output layer back to hidden layer using the delta rule such 4. RESULTS AND VALIDATION 
that the total squared error of output is minimized. 
  
Figure 5 shows the training and testing area. In (a) the training 
area which includes slide (blue) and healthy (green) pixels are 
shown. Figure 5 (b) depicts the testing area, from a nearby but 
different section of the levee system. Based on the algorithm 
explained in section 3, first the training area is used to train the 
BNN parameters. Around 300 pixels from each class (slide and 
non-slide) are used for training the BNN. The BNN has 36 
input neurons, one hidden layer with three neurons, and one 
output neuron. The transformation function is a hyperbolic 
tangent sigmoid, and the training function is Levenberg- CI: 
Marquardt back propagation. The maximum number of epochs 
for training is set to 100. For each pixel of testing area, the 
features are extracted and the BNN decides whether that that 
pixel is a slide or healthy pixel. The result of applying the 
algorithm to the testing area is depicted in Figure 6(a) showing 
the detected healthy (non-slide) and slide pixels. The healthy 
pixels are green and the slides are blue. Figure 6 (b) shows the 
ground truth and target data. In this figure, blue shows the slide 
and the green pixels are the healthy area. As we compare the 
two images we see that most of the slide area is correctly 
detected. However, a portion of the healthy area is classified as : 
slide. Table 1 shows the class confusion matrix for the 
Figure 4. Back propagation neural network with one hidden classification. As can be seen, slide pixels are classified with 
layer 50% accuracy. The overall accuracy is 67% for this testing area. 
Figu 
Terr 
  
  
  
He 
  
  
  
  
It appears that the area which was classified as slides has a high 
surface roughness, which is detected as slide pixels. Also, since 
the TerraSAR-X is sensitive to vegetation, any changes in the