Neural networks for meteorological satellite image interpretation
Meteorological satellite images at visible and infra-red wavelengths are an invaluable source of information on cloud systems because of their extensive coverage of the whole of the Earth's surface, providing data in areas that are only sparsely monitored, if at all, by other means. Although this information has been used subjectively by forecasters for many years, the lack of automatic, quantitative analysis techniques largely prevents its assimilation into numerical weather prediction (NWP) models, which are the basis of all modern weather forecasting. This thesis investigates the use of neural network techniques for the analysis of the images in order to make fuller use of the available data. The recognition of a particular type of cloud is dependent on the determination of a set of features from the satellite image spectral bands that will give discriminating information. This feature extraction and selection process is dealt with in detail, and a feature selection process based on the radial basis function (RBF) neural network is presented. The high-dimensional feature space is visualized on a two-dimensional plane by the use of three techniques: the Kohonen map, the Sammon map, and a recently-developed technique known as the Generative Topographic Mapping (GTM). Classification results using a multi-layer perceptron (MLP) and an RBF neural network are presented. The results of independently classifying each pixel in an image are compared with a method that makes use of contextual information, the Markov Random Field (MRF) model. The limitations of the pixel-based approach are highlighted, and a region-based approach is presented that enables the definition of large-scale regions of uniform cloud type. Two segmentation methods are used, the active contour (or snake) model, and the more recentlydeveloped level set technique. The latter method was found to provide many benefits over the former. The region-based approach will facilitate the assimilation of cloud system information into NWP models in the future.