Applications of
Artificial Neural Networks
to Ionograms

Markus Hagenbuchner

Supervised by
John A. Fulcher

University of Wollongong
June 26, 1996
Department of Computer Science


The aim of this project is to determine methods to extract main F-layer traces from either parametric or graphical inputs using Artificial Neural Network techniques.
Various networks, learning rules and preprocessing methods are compared by their performance and ability of being realistic applications.
The project is split into two parts:
  1. The usage of F-layer specific parameters (e.g. layer thickness and height) to train networks. This is discussed in section 3.

  2. Using the photographic record (ionogram) of an ionosonde as network input, which is presented in section 4.

An overview to simulators, network architectures and learning algorithms used in this project is given in section2.


In 1901, Marconi established the first radiowave transmision between Europe and North America. Later, in 1924, Kennely and Heaviside independendly suggested that this communication was only possible because of the reflection of radio signals by a conducting layer near 80km altitude. Radiowave methods later led to the first quantitative studies of this layer, through analysis of emitted signals reflected to the surface.
This layer plays an important role in the transmission of certain radio waves, which can be reflected or refracted if the frequency is below 30MHz (used for shortwave radio transmissions), or transmitted if their frequency is above this limit (used for television transmissions and satelite communications).


As first guessed by Heaviside and Kennely, the layers in the atmosphere that reflect radio waves indeed contain charged particles, namely ions and free electrons. In later experiments it was shown that radio waves travel faster within an ionized atmosphere than in air without ions. This change of speed is greater the smaller the frequency. It was also found that for higher frequencies more charged particles are needed to reflect the wave. What this means is that a high frequency wave will travel to higher altitudes in the atmosphere than radio waves with a lower frequency.
With this knowledge and the theory that the concentration of ions gradually changes within an ionized layer, the reflection of radio waves can be explained. Figure 1 shows that wave front A travels more rapidly than B, since the speed of a wave is faster the higher the concentration of ions.

---(Graphic missing)---

Figure 1: Reflection of radio waves in the ionosphere.

A definition was made by general agreement so that the part of the atmosphere that contains sufficient ions to affect the travel of radio waves was called ionosphere.

The ionosphere starts at an altitude of about 60km and is subdivided into 4 layers according to the occurance of peaks in the height distribution of electrons. Those peaks occur near heights of 70, 100, 170 and 200 km and are said to belong to the D-layer, E-layer, F1-layer and the F2-layer respectively.

The ionosphere changes its state throughout the solar cycle, as well as throughout the day and according to the seasons. The current state is investigated by a simple apparatus named {\bf ionosonde} that sends radio signals towards the sky and receives the reflected signal. It records the frequency of the wave and the time delay between sending and receiving. This delay can be used to calculate the apparent height of the reflectin ionized layer. The ionosonde produces a photographic record called an ionogram (Figure 2), in which the time delay is plotted against the frequency.

---(Graphic missing)---

Figure 2: Noisy, real world ionogram.

Vertical lines are caused by fixed-frequency radio transmitters. The horizontal line on the bottom represents the E-layer reflection. The nosy curve above are the F-layer reflections. The two weaker signals above show 2-hop and 3-hop F-layer reflections, where a reflected signal bounces back from the Earth's surface once or twice before it reaches the receiver.
For this project we look mainly at the F1 and F2-layer reflections, as these reflect shortwave (HF) signals, which is most useful for radio transmissions. As the F1 and F2 layers are not sharply separated and the F1-layer frequently (and always at night) disappears, the layers above the E-layer are often treated as a single F-layer.

There are two possible paths that radio waves can take when they are reflected in a thick ionized F layer (Figure 3).

---(Graphic missing)---

Figure 3: Two possible signal paths.

The high angle path is the longer path and has therefore a greater delay than the shorter low angle path. The result of this is that in transmitting radio waves the receiver obtains signals with an echo.

---(Graphic missing)---

Figure 4: The main trace of an ionogram (1-hop F-layer reflection).

The high- and low angle paths come close together as the frequency increases and join at the nose (also called the junction frequency), as shown in Figure 4. Beyond this point the reflection rapidly cuts off as there are insufficient charged particles to reflect these high frequency waves. The most powerful reflection occurs exactly at the nose frequency. It is therefore very important for radio transmissions to know this nose frequency, so that a receiver gets a clean and clear signal with maximum signal strength.

Unfortunately the ionosphere is already heavily used for radio transmissions. Also magnetic storms (caused by high sunspot activities) in the upper atmosphere often cause noise which distorts the shape of an ionogram. In addition, the high angle path of a radio wave is often very weak, so that the characteristic ionogram does not have a nosey shape. This makes it very difficult and sometimes impossible to identify the nose frequency.


Radio waves obey known physical laws. So a simulation of natural ionograms on a computer is possible and would provide a time and money saving tool. Jinoi, a simulater of this type was programmed by DSTO (Defense, Science and Technology Organization). This simulator produces the main trace of F-layer reflections. Unfortunately it was found that Jinoi has some minor bugs so that it sometimes produces ionograms that are obviously impossible in the real world, and even produces NaN errors or locks up. Undoubtedly the physical laws behind this approach are highly complex so that it is impossible for todays computers to handle all the cases that could happen in real life. In fact the production of 100% accurate ionograms would require a simulator that handles the behaviour of the huge number of ions that exist in the atmosphere, as well as airplanes, winds and sunspot activity, that could affect radio waves. From that point of view Jinoi works well although a postprocessing of its output is necessary to filter out 'bad' ionograms (Ionograms that had a negative frequency, more than one nose, a nose that shows to the left or had a high- and low angle path but no nose are found to be impossible in the real world and defined to be 'bad'.).

Early in the life of this project an upgrade version of Jinoi, called Jornoi, was provided by DSTO. Jornoi is much more stable than Jinoi, and produces mostly reasonable ionogramms. The output, however, still contains about 8% of ionograms which where found to be 'bad'.

One aim of this project is to produce a simulator that produces Jinoi (or Jornoi) like output by using Artificial Neural Network (ANN) techniques. A second aim is the extraction of the main traces out of real ionograms, again by using ANNs.

The University of Wollongong holds the copyright to this document. This is the reason why the main part had to be cut out here. If you are interested to get a copy of the full version of this document please contact John Fulcher at the University of Wollongong.


  1. J.A.Ratcliffe. Sun, Earth and Radio, World University Library, London 1970.
  2. John J. Hopfield. Neural networks and physical systems with emergent collective computational abilities 1982.
  3. D.O. Hebb. The Organization of Behavior, Wiley 1949.
  4. J.A.Ratcliffe. Physics of the upper atmosphere, Academic Press, 1960.
  5. Matlab Version 4.2c.1, MathWorks, Inc. 1992.
  6. M.Minsky, S.Papert. Perceptrons : an introduction to computational geometry, MIT Press, 1988, c1969.
  7. J.A.Ratcliffe.An introduction to the ionosphere and magnetosphere, Cambridge University Press, 1972.
  8. Guy Brasseur and Susan Soloman. Aeronomy of the Middle Atmosphere (Second Edition), D.Reidel Publishing Company, 1986.
  9. Hong Yan. Building a robust nearest Neighbour Classifier Containing only a Small Number of Prototypes, International Journal Of Neural Sysems V3 n4, pp 361-369, 1992.
  10. W.E.Simon, and J.R. Carter. Generalizing from a small set of taining exemplars for hand-written digit recognition, Applications of Artificial Neural Networks II, pp 592-601, 1991.
  11. D.E.Rummelhart, G.E. Hinton, and R.J. Williams. Learning Representation by Back-Propagating Errors., Nature vol. 323, pp 533-536, 1986.
  12. F.Rosenblatt. Principles of Neurodynamics, New York: Spartan, 1962.
  13. SNNS. User Manual, Version 4.1, University of Stuttgart, 1995.
  14. SNNS. The Stuttgart Network Simulator, Version 4.1, University of Stuttgart, 1995.
  15. J.Fulcher. Application of Neural Networks to the Analysis of Oblique-Incidence Ionograms, University of Wollongong, 1996.

  16. -- last update: Jan 17, 1998