Ionospheric variability and electron density profile studies with neural networks Original Research Article

Previously, the authors used Neural Networks (NNs) to predict the maximum electron density in the ionosphere over Grahamstown, South Africa (33.3°S, 26.5°E). Day number, hour, a 2 month running mean sunspot number (R2) and a 2-day running mean magnetic index (A16) were used as inputs to the NN. Two further applications of NNs are discussed in this paper. The first section of this paper discusses the use of Neural Networks (NNs) to quantify and describe the variability of ionospheric parameters. The parameter foF2 is used by way of illustration. It is suggested that variability can be well described in terms of a predictable, average response that depends on external variables such as latitude, longitude, day number, local time, sunspot number and magnetic activity, and a part that is unpredictable in the short term but nevertheless can be described in terms of a standard deviation. The need to be specific about the values of the external variables is stressed, and is illustrated with reference to the highly non-linear response of the ionosphere to combined seasonal and magnetic influences. The lack of a simple relationship between the average, predictable variability and the unpredictable part is illustrated. It is shown that NNs can provide an elegant method of quantifying variability with respect to any external parameter without the need to divide the data into categories for independent analysis. A preliminary investigation into the use of NNs to produce a bottomside electron density profile is discussed in the second section of this paper. Three years of hourly electron density profile data have been collected at the Grahamstown ionospheric station. This data was produced using a Lowell Digisonde and processed using the automatic scaling software, Artist-4. A NN was trained with the 12 noon local time data to produce a bottomside electron density profile for Grahamstown at midday. Day number and R2 were used as inputs to the NN. The output was the electron density at selected heights. Further work will include expanding the model for all hours. Since only three years of archived Grahamstown data are available, this initial NN model will need to be updated continuously until at least one half of a solar cycle of data, from the recent minimum to the next peak is included. There are two other South African digisonde sites, but data is not currently archived for either of them. Should data become available from these sites, the NN could be re-trained to include latitude and longitude.