Artificial Neural Networks are becoming important tools in a wide variety of meteorological applications. However, the estimation of rainfall has continued to be a very difficult and complex problem to solve. A single multi-layered back propagation neural network used on complex problems involving different sub-tasks will often show strong inter subtask interference effects that lead to slow learning and poor generalisation. The modular neural network approach is to decompose the classification task into simpler sub-tasks, each one being handled by a separate module. The modules are then combined to produce an overall solution demonstrating a very natural way to solve a complex problem. One modular neural network strategy is an ensemble of neural networks where each network is a whole problem classifier. This paper investigates the development of an ensemble of neural networks for the rainfall estimation problem. The results demonstrate that an ensemble of neural networks has the potential to achieve an improvement in performance for rainfall estimations.

Genetic algorithms (GAs) have been utilised as tools in neural network development across a wide range of problem domains. However a potential disadvantage with combining these two techniques is the amount of processing time taken. Three important factors which significantly affect the time taken are the GA's population size, the number of GA generations, and the number of neural network passes. To investigate the tradeoffs between the three variables, an exhaustive set of experiments were carried out on four well known classification problems. In order to evaluate the results obtained, a comparison was made with previously published results using a number of other classification techniques including backpropagation neural networks, C4.5 and 1R. Results showed that a small GA population size was favoured for all problems investigated, while the best number of GA generations and neural network passes were problem dependent.

Various techniques for estimating meteorological parameters have been developed over the past few years that involve artificial neural networks. However, the estimation of rainfall has continued to be a very difficult and complex problem to solve. Data mining techniques are needed to extract the important information from the vast amount of meteorological data available. A single multi-layer backpropagation neural network used on complex problems involving different sub-tasks will often show strong inter sub-task interference effects that lead to slow learning and poor generalisation. Dividing the system up into several different "expert networks" each specialising in a different sub-task can reduce this interference at the cost of having to combine the outputs from each of the experts. This paper investigates the technique of dividing the rainfall estimation problem into a number of such experts each specialising in a particular rainfall band (i.e. low, medium or high rain). Results demonstrate that expert networks can be successfully developed which result in both improved individual classifications and improved overall classification accuracy.

This paper investigates the use of an Artificial Neural Network (ANN) to estimate the six hour rainfall over the south east coast of Tasmania. ANN's are becoming increasingly prominent in many areas of weather forecasting due to their potential to capture the complex relationships between the many factors that contribute to certain weather conditions. The estimations produced by the ANN's were compared to one estimation technique and one forecast technique used by the Bureau of Meteorology. The results confirm that ANN's have the potential for successful application to the problem of rainfall estimation.