The effect of neural-network structure on a multispectral land-use/land-cover classification

While neural networks are now an accepted alternative to statistical multispectral classification techniques for remote sensing image classification, the network approach presents both unique challenges and abilities. The size of the hidden layer must be determined by trial and error, and the random initial weight settings result in different paths for the training procedure, making the network a non-deterministic classifier. For the sample classification presented here, it was found that there was a range of optimal hidden layer sizes below which the accuracy decreased and above which the training time increased. However, it was also found that, for a fairly wide range, the hidden layer size made little difference to the final classification accuracy. Initial weight randomization was as much of a factor as hidden layer size. Using 3 by 3 windows of data in each band was found, despite increased training time per iteration, to achieve similar accuracy with less overall training time, although with less consistency.