Backpercolation Training of Neural Networks for Agricultural Land Use Classification with Landsat-tm Data

The recently published backpercolation algorithm for the training of neural networks will be compared with the backpropagation and quickpropagation algorithm by means of "artificial" classification problems (e.g. XOR, M-N-M decoder) and serveral others. Within all classification schemes the backpercolation algorithm is much more efficent and even successful where the other training schemes are without success. Several different neural networks are trained by backpercolation in order to classify agricultural land usewithin multitemporal LANDSAT-TM scenes and to show the efficency of various network configurations (e.g. number of neurons in the hidden layers, number of hidden layers) and the impact of training parameters. In the classification fourteen different classes are discriminated. A relaxation process after the classification is able to improve the accuracy of the result up to 15 % in various classes. For all classes the achieved classification result is better than 80 % up to 90 % or single classes on a pixel bases.