Study on the characteristics of the supervised classification of remotely sensed data using artificial neural networks

The characteristics of classification of remotely sensed data using artificial neural networks are investigated. The training method of neural networks consists of a generalized delta rule (GDR) and a conjugate gradient (CG). The GDR is divided into two methods, data adaptive and block adaptive. The effects of the number and order of input data and learning rate were analyzed in the training. Data adaptive and block adaptive methods showed similar trends of error convergence in the GDR. The CG especially with a small data set had faster error convergence than the GDR. The CG having low error in the training didn´t show good accuracy in the testing stage because of the overtraining effect