Application of Neural Networks to Large-Scale Cloud Pattern Recognition

A Multi-layer Perceptron Neural Network methodology is used to classify eight types of large-scale cloud patterns. The data are taken from GOES-W visible images from Oct. 1-Dec. 31, 1983. Large-scale features are previously identified by a human expert to provide a data set for supervised learning. Discriminant Analysis is used to reduce the set of network inputs and as a comparison classification methodology. In three different tests, the neural network technique classifies the cases with consistently higher accuracy than Discriminant Analysis. The problem of image segmentation is addressed in a preliminary test of the Hierarchical Stepwise Optimization algorithm.