>I met a problem in neural network when doing remote sensing image classification. >The problem is that the classification is not stable as to training samples. In the code, >I use simple random sampling. I can accept the fact that we could get different results >for different sampling. However, in many cases, some classes are not classified at all.
Try stratified random sampling to get uniform training priors.
How can you not get a classification? Aren't you using columns of the unit matrix as targets and vec2ind on your posterior probability estimates?