HR: 14:25h
AN: SA22B-04    [PDF]
TI: An Empirical Ionospheric Model for the High Latitude Lower Ionosphere Based on Neural Networks
AU: * McKinnell, L
EM: L.McKinnell@ru.ac.za
AF: Department of Communications and Wave Propagation, Technical University Graz Inffeldgasse 12, Graz, A-8010 Austria
AU: * McKinnell, L
EM: L.McKinnell@ru.ac.za
AF: Rhodes University, P O Box 94, Grahamstown, 6140 South Africa
AU: Friedrich, M
EM: friedrich@inw.tugraz.at
AF: Department of Communications and Wave Propagation, Technical University Graz Inffeldgasse 12, Graz, A-8010 Austria
AB: This paper discusses the development of a new empirical model for the lower ionosphere in the auroral zone. Available ionospheric data have been used to train neural networks (NNs) to predict the high latitude electron density profile. Data from the European Incoherent Scatter Radar (EISCAT), based in Tromso (69.58$\deg$N, 19.23$\deg$E), combined with rocket borne measurements make up the database of reliable D- and E- region data. NNs were trained with different combinations of the following input parameters: day number, time of day, total absorption, local magnetic K index, 10.7 cm solar radio flux, solar zenith angle and pressure surface. Initially the database was split into night and daytime data and optimum combinations of these inputs were determined for each dataset. The output that the NNs were trained to predict was the electron density for a given set of input parameters. The criteria for determining the optimum NN are a) the root mean square (RMS) error between the measured and predicted output values, and b) the ability to reproduce the absorption they are representative for. Results from the separate night and daytime models show this method to be successful. Comparisons were made between a conventional analytical approach and this new NN approach. However, a discontinuity showed up at the night-day boundaries when the models were combined to produce the electron densities over an entire 24-hour period. This was not surprising as information pertaining to this boundary was not implicitly included in the dataset with which the NN was trained. Therefore, as another approach NNs were also trained with the entire dataset, night and day time combined. Results from this approach will also be shown as well as comparisons with the conventional analytical method and with measured data. An essential requirement for the employment of the NN technique is a large reliable database that describes the history of the relationship between the input parameters and the output. NNs can still be designed and trained with a limited database as long as the end user is made aware of the limitations of the input space. It is well known that NNs interpolate well but do not extrapolate well. The advantages of the NN method include the ability to re-train a NN relatively easily should more data become available. This paper will show that a NN based model for the high latitude lower ionosphere has been developed and is successful within the limitations of the input space.
DE: 2400 IONOSPHERE
DE: 2407 Auroral ionosphere (2704)
DE: 2447 Modeling and forecasting
SC: SPA - Aeronomy [SA]
MN: 2003 Fall Meeting