Operational use of a neural network cloud classifier for flood forecasting at the UK Met. Office

In 1996, a Meteosat neural network cloud classifier was developed for use in a flood forecasting scheme then being trialled at the UK Met. Office (GANDOLF: Generating Advanced Nowcasts for Deployment in Operational Land-based flood Forecasts). The aim was to deliver an automatic system that could discriminate stratiform (frontal) cloud from shallow and deep convective cloud, in order to determine which of two models should be used for forecasting precipitation rates in the timescale 0-6 hours. One model (Nimrod) is used in stratiform cloud situations, to advect precipitating cloud according to mesoscale model winds, whilst the other model (the Object Oriented Model) determines the current state of a convective cell in its life-cycle, and forecasts accordingly, whether that cell is growing, decaying or likely to produce daughter cells.

The scheme was trialled over 3 summer periods (when the convective scheme would have its largest impact), and results showed that from a number of potential cloud diagnoses, the Meteosat neural network cloud classifier used in conjunction with Convective Available Potential Energy diagnosed from the UK Met. Office's mesoscale model provided the best discrimination between stratiform and convective cloud.

On 1st June 1999, the GANDOLF system became an operational service to the UK Environment Agency, and an improved neural network cloud classifier was included. This paper describes the changes made to the classifier, in order to produce a robust operational system, with an overall classification accuracy of 93%. Rather than using two classifiers (day and night), one neural network was implemented with the ability to cope with the loss of visible data in such a way as to make dawn and dusk transitions as smooth as possible. Examples of day, night and terminator images are presented, together with the use of the cloud classifier in triggering the convective cell life-cycle model.