Finding optimal neural networks for land use classification

In this letter we present a fully automatic and computationally eecient algorithm based on the Minimum Description Length Principle (MDL) for optimizing multilayer per-ceptron classiiers. We demonstrate our method on the problem of multispectral Landsat image classiication. We compare our results with a hand designed multi-layer percep-tron and a Gaussian maximum likelihood classiier where our method produces better classiication accuracy with a smaller number of hidden units.