Pattern recognition with neural networks
Takushi Yoshida, S. Ōmatu · 2002
Remote sensing has become important in pattern classification from the view point of global environmental problems according to the progress of space technology. But the classification performance with remote sensing data depends on a training data set for supervised classification. However, it is difficult to select it from remote sensing images, since remote sensing data includes various kinds of categories and similar information, which depend on sensors. Therefore, one must take care of its selection. The authors investigated a training data set by independent component analysis (ICA) and proposed a pattern classification system for remote sensing data based on neural network theory. From independent component analysis, training data for each pattern are converted to independent data set regardless of observation sensors. Using the BP algorithm, a layered neural network is trained such that the training pattern can be classified within a level. The experiments on LANDSAT TM data show that this approach produces excellent classification results compared with conventional statistical approaches, which are Bayesian and distance method etc.