Through analyzing the influence of the learning samples’ location in the spectral space on the accuracy of multi spectral remote sensing image classification using BP neural network, a method for learning samples selection based on x 2 distribution was presented and used in TM image classification.

Adopting the improved B P algorithm to establish network structure and thoroughly investigating the process, and the method constructing learning sample has been improved.

Firstly, divide the main class into subclasses which assume the similar index value, take the subclass center as the typical learning sample then make a new set of learning sample, and train the neural network using this sample set.

One of selecting rules of learning samples is proposed according to the principle of the pattern recognition model.

It shows that the number of learning samples is the key for a good p.

According to the variation characteristics of the operating condition of internalcombustion engines, an improved orthogonal design method which considers the changes on factor boundary is proposed to choose learning samples of neural networks.

Secondly,to extract learning samples from the MADM problem,an approach to estimate the utility functions for attributes is presented.

This method integrates the setting of initial parameters with the wavelet type,time-frequency parameters of the wavelet and the training samples.

In view of the original unstructured data existing in various tables, the sample learning algorithm based on artificial neural network is put forward to refine the diagnosis rules so as to realize automatic knowledge acquiring.