Neural network models for the prediction of autumn migration of the cereal aphid Rhopalosiphum padi at Lincoln, Canterbury, New Zealand

Neural network models were developed to predict the number of R. padi caught during the autumn flight period, at Lincoln, Canterbury. The models were based on weather data and aphids caught in a suction trap over the period 1982-2000. The first neural network model was trained using weekly data over a time-series of 15 years. The network predicted the yearly bimodal flight accurately, but could not generalise well to new data because of over-fitting. To eliminate stochastic noise from the data, four main modifications were carried out, 1) the autumn flight period only was used for modelling; 2) the total number of aphids caught during that period was used as the dependent variable to be predicted; 3) years were arranged randomly to prevent the network models from learning major trends; and 4) the data was pre-processed, by identifying and selecting periods in which each weather variable showed a strong correlation with the total number of aphids caught during the autumn flight. The capability of the new models to generalise improved significantly when validated with 13 years of jack-knifed data. A very high r value between the observed and predicted number of aphids (0.9575) was obtained. The neural networks were compared with a multiple regression model developed using the same data. Despite the r value between observed and predicted number of aphids for the multiple regression model was high (0.8591), the neural network model performed better. The absolute mean error (ABSME) for the multiple regression model was 313.58 aphids, whereas for neural network model it was 121.51 aphids. A further development was the design of a calibration model. The calibration model was validated by jack-knifing each year and predicting the number of aphids in the autumn flight. Predicted and observed numbers of aphids were compared, giving a very high r value (0.91) and an ABSME of 94 aphids over the 19 years in the time series. Further development of the neural network models is discussed