Improving prediction of aphid flights by temporal analysis of input data for an artificial neural network

Weather data in its raw form frequently contains irrelevant and noisy information. Often the hardest task in model development, regardless of the technique used, is translating independent variables from their raw form into data relevant to a particular model. A sequential or cascading temporal correlation analysis was used to identify weather sequences that were strongly correlated with aphid trap catches recorded at Lincoln, Canterbury, New Zealand, over 1982-2000. Trap catches in the previous year and 13 weather sequences associated with eight climate variables were identified as significant predictors of aphid trap catch during the autumn flight period. The variables were used to train artificial neural network (ANN) models to predict the size of autumn aphid migrations into cereal crops in Canterbury. Such models would assist cereal growers to make better informed and more timely pest management decisions. ANN predictive performance was compared with multiple regression predictions using jackknifed data. The ANN gave superior prediction compared with multiple regression over 13 jackknifed years.