An artificial neural network predictor for tropospheric surface duct phenomena
Abstract. In this work, an artificial neural network (ANN) model is developed and used to predict the presence of ducting phenomena for a specific time, taking into account ground values of atmospheric pressure, relative humidity and temperature. A feed forward backpropagation ANN is implemented, which is trained, validated and tested using atmospheric radiosonde data from the Helliniko airport, for the period from 1991 to 2004. The network's quality and generality is assessed using the Area Under the Receiver Operating Characteristics (ROC) Curves (AUC), which resulted to a mean value of about 0.86 to 0.90, depending on the observation time. In order to validate the ANN results and to evaluate any further improvement options of the proposed method, the problem was additionally treated using Least Squares Support Vector Machine (LS-SVM) classifiers, trained and tested with identical data sets for direct performance comparison with the ANN. Furthermore, time series prediction and the effect of surface wind to the presence of tropospheric ducts appearance are discussed. The results show that the ANN model presented here performs efficiently and gives successful tropospheric ducts predictions.