Use of neural networks to short-term forecast of airborne pollen data

In the last two decades the allergic diseases inducted by allergenic pollen have dramatically increased. A better application of preventive allergic therapy is highly dependent on improvements of the methods for early forecasting of pollen data. Analytical models, which are based on differential equations, describe emission and dispersion of pollen in the atmosphere and, therefore, can be use to forecast airborne pollen. These models combine many parameters related to plant observations and weather conditions, but their application is difficult because of lack of data with regard to parameters. An alternative way to forecast pollen data is based on the statistical analysis of past pollen concentrations, by means of time series methods. The Box-Jenkins methodology, which have dominated the area of time series forecasting since 1970, especially with the ARIMA models, forecast a variable by a linear combination of the previous state of the variable and the previous forecast errors. The pre assumed linear form of the model is the major limitation of ARIMA methodologies, that are not be able to capture non linear patterns that affect many environmental phenomena. Artificial neural networks (ANN) have been applied in time series modelling and forecasting as tool for modelling the complex and non-linear relationships among time series values. Several studies have also reported the capabilities of ANN in forecasting environmental variables with different time steps (short and medium term). The aims of this study are (I) to develop a neural network model to short-term forecast airborne pollen concentration and (II) to analyse and compare the effect on the forecasted values of the different model parameters and (III) to improve the accuracy of airborne pollen forecasting for some of the main allergenic plants of the Mediterranean area. The study was carried out on the urban area of Sassari, Italy where aerobiological and meteorological data have been collected from 1986 to 2000; the sampling device was a Burkard volumetric spore trap. The meteorological data, collected by a Sardinian Agrometeorological Service weather station, were: air temperature, air relative humidity, wind speed and direction, and rain intensity. The data set of eleven years of data was used to calibrate the model, whereas a second data set of four years, was used to validate and test the model. The analysis was performed on Graminaceae and Oleaceae, two of the main allergenic families of Mediterranean area. ANN models were realised using two topology (three-layer feed-forward and Radial Basis Function) and two learning optimisation algorithm (backpropagation and Levenberg-Marquardt). All the ANNs use as input the calibration airborne pollen time series; the forecasting period ranged between 1 to 7 days. Subsequent data were forecasted using both one-lag and multi-lag methodologies. The time-lag used as input and the number of processing units were determined by optimization techniques. The study shows the capabilities of ANN as a tool for short-term forecasting of pollen concentration and as a support for application preventive allergic therapy.