Neural network classification of convective airmasses for a flood forecasting system

Abstract

A multilayer perceptron neural network was trained in order to classify shallow and deep convective airmasses for trial operational use in the GANDOLF thunderstorm flood forecasting system, being developed by the U.K. Meteorological Office and the Environment Agency. The airmass classification determined whether to use a precipitation forecast model based on extrapolation techniques, or an object-oriented convective precipitation forecast model based on a conceptual model of convective cells. The network was trained on a subset of 2190 Meteosat image samples taken during 1994, each labelled by a meteorologist at the U.K. Meteorological Office with the aid of surface pressure analysis charts. The output classes represented by the network included clear land and sea, dynamic cloud, and shallow and deep convective cloud. A total of seven spectral and two temporal features were used to discriminate the classes, resulting in a network with a correct classification rate of 80.8 +/- 2.2 per cent.