Neural networks have been successfully used to classify pixels in remotely sensed images. Especially backpropagation neural networks have been used for this purpose. As is the case with all classification methods, the obtained classification accuracy is dependent on the amount of spectral overlap between classes. In this paper we study the new idea of using hierarchical neural networks to improve the classification accuracy. The basic idea is to use a first level network to classify the easy pixels and then use one or more second level networks for the more difficult pixels. First a rather standard backpropagation neural network is trained using the training pixels of a ground truth set. Two ideas to select the difficult pixels are tested. The first one is to take those pixels for which the value of the winning neuron is below a threshold value. The second one is to select pixels from output classes, which get a high contribution from wrong input classes. Both ideas improve on the percentage correctly classified pixels and on the average percentage correctly classified pixels per class.
[1]
R.P.H.M. Schoenmakers,et al.
Results of a hybrid segmentation method
,
1994,
Remote Sensing.
[2]
I. Kanellopoulos,et al.
Strategies and best practice for neural network image classification
,
1997
.
[3]
Geoffrey E. Hinton,et al.
Learning internal representations by error propagation
,
1986
.
[4]
Giles M. Foody,et al.
Land Cover Mapping from Remotely Sensed Data with a Neural Network: Accommodating Fuzziness
,
1997
.
[5]
Z. K. Liu,et al.
Classification of remotely-sensed image data using artificial neural networks
,
1991
.
[6]
Theo E. Schouten.
Fuzzy classification of pixels using neural networks
,
1999,
Remote Sensing.