A neural network based electron density model for the E layer

Abstract

This paper describes the development of a neural network (NN) based model for the ionospheric E layer over Grahamstown, South Africa (33.32;S, 26.50;E). The development of this model involved the training of several different NNs to predict all aspects of the E layer. Five years of Lowell DPS data collected at Grahams were used for training and testing the NNs. The final model requires the day number, hour and 1 month running mean sunspot number as inputs. Various combinations of these inputs are used for each stage of determining the profile. NNs were trained to predict the peak height (hmE) and the maximum electron density (foE) of the layer for a particular set of inputs. Another NN, E profile NN, was then trained to predict the coefficients of a Chebyshev polynomial from which the real height could be determined as a function of frequency, between 0.2 MHz and foE. In addition of a fourth NN, E limits NN, was trained to determine the hours between which an E layer would be measurable by a ground based ionosonde, thereby preventing the foE, hmE and E profile NNs from being interrogated with input data with which they have not been trained. At times that fall outside of those predicted by the E limits NN existing empirical models are used. This NN based model is currently a single-station model, since the NNs were trained with only Grahamstown data. Future work will include adding latitude and longitude to the input space in order to expand the model to cover all of South Africa.