Retrieval of in situ electron density in the topside ionosphere from cosmic radio noise intensity by an artificial neural network



[1] An artificial neural network (ANN) is applied to retrieve electron density at satellite height from cosmic radio noise measurements made near the plasma cutoff frequency in the topside ionosphere. Observations were made using the topside sounder on board the Ionosphere Sounding Satellite launched in 1978. The automatic gain control voltage of the sounder receiver supplied information on the spectrum of cosmic radio noise intensity. Local electron densities, which are the target of the ANN, were scaled from the plasma resonance spikes and the echo traces in the ionogram for several passes. Each pass contained about 100 frames of ionograms; it is not a very easy task to scale the local plasma parameters by an eyeball judgment for all passes. With the aid of a neural network application, we determined the in situ electron density for about 150,000 ionograms. The results were applied to global ionospheric mapping.

1. Introduction

[2] In the topside ionosphere, the short-wave cosmic radio noise (CRN) intensity at frequencies lower than the ionospheric critical frequency is free from interference from man-made radio noises and transmissions on the ground. Because of this advantage, in the 1960s and 1970s, many rocket-borne and satellite-borne radio receivers were launched above the F layer ionization peak to measure the cosmic radio noise spectrum in the HF band [e.g., Hoang, 1971]. As the frequency lowers, however, the intensity is influenced by the surrounding ionospheric plasma. The extraterrestrial radio waves having a large incidence angle into the stratified ionospheric layer are reflected and never reach the antenna when they are at frequencies close to the local electron plasma frequency (fN). Using a ray-tracing technique, Hartz and Roger [1964] calculated a cone of directions within which a radio signal coming toward the Earth can reach an antenna in the topside ionosphere. They also compared the results with the experimental data from the Alouette satellite sweep-frequency receiver and thus confirmed the validity of the calculations. The effective aperture of the antenna was shown to reduce gradually first with decreasing frequency and then to reach rapidly toward zero near the plasma cutoff frequency [Hartz and Roger, 1964].

[3] Usually, electrically short dipole antennas are used on rockets and satellites to receive the HF cosmic radio noise in the topside ionosphere. At frequencies higher than the upper hybrid frequency (fT), received signals vary directly as the antenna-radiation resistance changes with the frequency due to the influence of the plasma; the resistance decreases gradually at first with decreasing frequency; then, it plunges rather rapidly to a cusp after which it again decreases slowly [Weil and Walsh, 1964]. The cusp corresponds to the extraordinary cutoff frequency (fX), at which the radiation resistance is about half the value in free-space.

[4] From the viewpoint of ionospheric interests, the radio noise spectrum close to the cutoff frequency contains information on the electron density at the point where the cosmic radio noise was received. In this paper, we will retrieve the in situ electron density parameter from a large amount of radio noise spectrum data obtained by the ISS-b satellite. However, as mentioned above, received radio noise spectra are a complicated function of the frequency spectrum of radio sources, variations in the effective antenna aperture due to the reflection, and variations in the radiation resistance due to the surrounding plasma. Furthermore, extraterrestrial radio sources in the MF and HF bands distribute nonuniformly over the sky [Kotaki et al., 1981]. For a quantitative treatment of the CRN intensity, detailed prelaunch and in-flight calibrations are required of a particular receiving system; unfortunately, in the case of ISS-b, such information is not quite sufficient. Thus, theoretically formalizing the relationship between the in situ electron density and the observed radio noise spectrum remains very difficult. As an alternative to a theoretical approach, we apply an artificial neural network (ANN) to the radio noise spectrum.

[5] The engineering parameters of the satellite, radio noise receiver (actually outputs of the AGC circuit of the swept-frequency sounder receiver), and topside sounder, which are pertinent to the interpretation of the data and to the design of an ANN, are summarized in section 2. Section 3 describes the ANN. section 4 presents the results and discusses possible error sources. Some applications of the data to global mapping of the ionosphere are demonstrated in Section 5.

2. ISS-b Satellite and Topside Sounder

[6] The Ionosphere Sounding Satellite, ISS-b, was designed in the late 1960s as the first Japanese space mission following the successful launch of the Canadian topside-sounder satellite Alouette 1. The main objective of the satellite was global mapping of the ionospheric parameters. For this purpose, ISS-b carried a digital-topside sounder and a data recorder capable of storing data for one revolution period of the satellite orbit. However, the capacity of the data recorders, which used a magnetic tape based on the level of technology at that time, limited the design of the sounder and other equipment. Thus, the resolution of the ISS-b topside ionograms was poor compared with that of the Alouette and ISIS satellites. Despite this disadvantage, topside soundings for four revolutions a day - each orbit was selected to be distributed evenly in a longitudinal direction - were quite successful in globally mapping the ionospheric critical frequency [Matuura et al., 1981]. In the same context, the in situ electron density that will be discussed here is still valuable for the study of global morphology of the ionosphere, including a contribution to the improvement of empirical ionospheric models, such as IRI [Bilitza et al., 1993].

2.1. Topside Sounder

[7] The frequency sweep of the sounder was made stepwise with an increment of 100 kHz from 0.5 to 14.8 MHz. The pulse-repetition frequency was 9 Hz and the automatic gain control (AGC) voltage was recorded just prior to each RF-pulse transmission, which made it possible to evaluate the natural radio noise level from the AGC data without any contamination of the plasma waves that are locally excited by the pulse at frequencies close to the resonance conditions [Benson, 1982]. In the Alouette and ISIS topside sounders, a receiving-only mode was prepared [Hartz, 1969]. It took 16 s to obtain one frame of ionogram and the observation rate was one frame each 64 s, totally approximately 100 ionograms, and hence radio noise spectra were recorded during one revolution period.

[8] As is understood from the frequency resolution of the ionogram, scaling the in situ plasma parameters from resonance spikes and local cutoff frequencies of the echo traces is not easy because some of them drop between the consecutive transmission frequencies. However, with the aid of a geomagnetic field model and the known relationship among the characteristic frequencies, such as the electron gyrofrequency, the upper hybrid frequency, and the X-, O-, and Z-mode cutoff frequencies, we can determine the local electron plasma frequency (O-mode cutoff) that should be consistent with the other characteristic frequencies on a trial-and-error basis. However, it is almost impossible to apply this approach to process 150,000 ionograms.

2.2. Cosmic Radio Noise Spectra

[9] Two pairs of dipole antennas with tip-to-tip lengths of 36.8 and 11.4 m were used for the topside sounding and the radio noise measurement. Automatic gain control was done at the second IF amplifier with a 60-dB dynamic range. This dynamic range was not sufficient for measuring the noise intensity at frequencies lower than the extraordinary wave cutoff, which is not a problem for the present purpose as we are using a change in spectra just above the cutoff frequency where the received radio noise intensity changes from a sufficiently high level down to the threshold level. An example of dynamic spectra for one revolution period is shown as a time-frequency-intensity diagram in Figure 1. This is not a typical example, but it contains both extremely high and low electron densities during the one revolution period. This example was, however, chosen to discuss a limit of the application of the method, which will be presented later. In the figure, the horizontal axis is expressed by the frame number instead of the unit of time for the convenience of referencing points. As mentioned above, each frame is separated by 64 s and frame numbers are easily converted into times. The universal times for the first and last frames are indicated below the axis and some important information on the orbit is indicated at the top of the figure. The white dots connected with a line are the ionospheric critical frequency (fOF2) scaled from the ionogram trace, and the sine-wave-like curve in the low frequency part is the electron gyrofrequency calculated from the IGRF model.

Figure 1.

Dynamic spectrum of radio noise observed as AGC voltage of the topside sounder on board the ISS-b satellite for about one revolution period. White dots connected by a line are the ionospheric critical frequencies scaled from the ionogram and the sine-wave-like line in the lower part is electron gyrofrequency calculated from the IGRF model.

[10] In the lower frequency part of the diagram, measurements of the noise intensity were not available as input signals dropped to a level lower than the AGC threshold, which is shown as a white area. We refer to the frequency at which the AGC voltage first drops to the threshold level, with decreasing frequency from the upper end of the sweep, as the AGC cutoff or f0. In other words, f0 is the high-frequency edge of the white area.

[11] The vertical wedge-shaped dark area at frame numbers 65 through 95 indicates interference by the ground transmissions at frequencies higher than fOF2. The horizontal continuous lines that appear throughout the whole period are contamination from the spurious signals generated by the electronics on board the spacecraft. Intense noise band is observed at the lowest frequency part just above the gyrofrequency, the intensity being correlated with the local plasma parameters.

3. Artificial Neural Network

3.1. Target

[12] The major purpose of our application of an ANN is to estimate the in situ electron density. Although selecting the plasma frequency (fN) as a target parameter is quite direct and natural, the extraordinary wave-cutoff frequency (fX) and upper hybrid frequency (fT) are also possible target parameters. To train the ANN, we manually scale the above parameters with the aid of a geomagnetic field model to obtain an fN, fX, fT, and a model-calculated electron gyrofrequency (fB) that are consistent with each other. Thus, selecting one of three parameters as a target is expected to be basically equivalent if fB is included in the input parameter set to the ANN. We will train and test the ANN for each target parameter and differences in efficiency will be discussed later.

3.2. Input

[13] As described in the previous section, the AGC cutoff, f0, should be the main input parameter. The radio noise spectrum starting from f0 up to a 2-MHz interval is also selected as an input parameter without using the whole sweep-range spectrum. This is to avoid an unfavorable effect from the noise band at frequencies lower than fT. The 2-MHz interval is a compromise between using a wide interval to accommodate a large amount of information on the spectrum change and avoiding interference from the ground transmissions above foF2. Although we have a conversion curve from the raw telemetry data (8-bits) to the input radio noise strength, we will use 8-bit raw data (Ai) as input information to the ANN, because the ANN is expected to learn the conversion curve also.

[14] As radio noises consist of ordinary and extraordinary wave components, cosmic radio noise spectrum near the cutoff frequency is also a function of the geomagnetic field. Thus, fB calculated by a geomagnetic field model is included in the input parameters even for the run in which fN is chosen as a target.

[15] Nonuniform sky distribution of the brightness in the short-wave band has been reported. Kotaki et al. [1981] observed the brightness in the direction of the galactic center to be twice as high as in the direction of the antigalactic center at 2.5 and 5 MHz. When the satellite is in the shadow of the Earth against the galactic center, f0 might be higher compared with when the satellite is on the galactic-center side. Thus, the galactic coordinate should be an input parameter to the ANN. Actually, we chose the angle (γ) between the zenith of the satellite and the direction of the galactic center.

3.3. Neural Network

[16] The basic structure of the ANN is a feed-forward back-propagation type consisting of three layers; an input layer, a hidden layer, and an output layer, as shown in Figure 2. The input layer has 23 input nodes, which are f0, Ai (i = 1 − 20) at a frequency of (f0 + 0.1i) MHz, fB, and γ. The output layer has one node of the target parameter selected from fN, fX, or fT. The number of nodes for the hidden layer is determined on a trial-and-error basis. The best efficiency was obtained for 8 nodes, although it was not very sensitive between 6 and 10. The sigmoid function is applied as an activation function, and hence, f0, fB, and the target parameters are normalized by dividing by 10, Ais are divided by 256, and γ is divided by π, so each parameter takes a value between 0 and 1.

Figure 2.

Structure of the artificial neural network for estimating local electron density parameters.

3.4. Training of the ANN

[17] The performance of the ANN strongly depends on the selection of the training data set. In making use of trained ANN, if the input data are outside the range of the data values used in the training, the reliability of the ANN output decreases. Thus, the training data set should cover as wide a range of values as possible. The in situ plasma frequencies were 1 to 2.5 MHz at most points in the orbit at 1000-km height during the ISS-b observation. A preliminary run of the ANN trained by a data set that consisted of one revolution period yielded reasonably precise results for many points, but it committed a large error near the magnetic equator where the electron density was high (Figure 1). The range of the electron density variations over one revolution around the Earth varies from orbit to orbit depending on the solar activity, equatorial crossing longitude, season, and other conditions as well as on the latitude and local time. We selected five orbits in which the electron density varies the most. The shaded portion in Figure 3 shows a histogram of the frequency distribution for the chosen orbits, and which indicates that only a few points are at frequency above 3.5 MHz. To compensate for the small number of points at high frequencies, five more orbits exhibiting high electron density near the magnetic equator are selected and data points around the density maximum are added to the training data set. Data points above 3.5 MHz increased by about four times; the frequency-distribution histogram for the final data set is shown by the solid curve in Figure 3. The local times of the equatorial crossing from south to north of the five orbits are 0859, 0547, 2200, 1535, and 1157 LT; the equatorial crossing time from north to south are shifted by about 12 hours. Their equator crossing longitudes are 16, 126, 28, 41, and 123 degrees East, respectively. The longitude versus the local solar time distribution of the data points is summarized in Figure 4.

Figure 3.

Histogram of the in situ electron plasma frequency of the training data set for the five full passes (shaded portion) and for the total set supplemented to the higher frequency part (solid line).

Figure 4.

Longitude versus local time distribution of the data points used for the training. For the full-pass data, high latitudes (>45°) and low latitudes (<45°) are distinguished by the circles and horizontal bars, and supplemental data are shown by pluses, which are mostly at low latitudes. Equator crossing points are indicated by the solid triangles.

[18] The first priority in selecting the training data set was to obtain as wide a range of electron density variation as possible. As a consequence, the orbits were selected during the period of high solar activity and in the season close to the equinox, i.e., January to April, in 1979. Confirming the validity of the trained ANN for different conditions is described in the next section.

4. Results and Discussion

[19] After training the ANN by using about 600 ionograms and CRN spectra, the electron plasma frequency fN for the pass presented in Figure 1 was evaluated; the data set in Figure 1 was not used for the training. First, we trained the ANN for the three target cases, fN, fX, and fT, and the results were compared in terms of plasma frequency, i.e., when fX or fT is chosen as a target parameter the results are converted to fN using model fB values. The best results were obtained when fX or fT was chosen as a target, while, when fN was chosen, the errors increased slightly at the points with small fN values (at frame numbers 88 through 92). Figure 5 shows the ANN-determined fN for the target fX (solid line), along with manually scaled fN (circles) and their differences (plus signs connected with a line). At frame numbers 64 through 68 and at number 74, scaled data are not available because of the poor quality of the ionograms. However, the ANN-determined values at those points seem reasonably precise when we compare them with the scaled values before and after that interval.

Figure 5.

Comparison between the electron plasma frequencies estimated by the ANN and those manually scaled from resonance spikes in the ionogram for the pass shown in Figure 1.

[20] At frames 17 to 19 and 88 to 92 in Figure 5, errors are noted to slightly increase. As mentioned already, this particular test data set contains large and small values of fN compared with an averaged condition, and the above mentioned increases in error seem to be due to the small number of data points in the training data set for large and small fN values as shown in Figure 3. Also, in the latter interval (at frames 88 to 92), where the scaled fNs are low, the antenna gain drops rapidly with decreasing frequency irrespective of the plasma condition because of the limited length of the antenna, and the AGC cutoff tends to converge at a constant value, even though the ANN uses information on the radio noise over a 2-MHz bandwidth above the cutoff. This constrains the application of the method to the study of ionospheric phenomena, and the electron densities that are less than 0.6 × 104 cm−3 or log ne < 3.8 are less reliable.

[21] The method described in this paper stands on the assumption that the radio noises received at the satellite have steady extraterrestrial origins and the ionosphere is stratified. However, on some occasions, these assumptions are not applicable. We will discuss them as possible error sources in the application of the method.

4.1. Plasma Bubbles

[22] A plasma bubble is a localized low-electron density region along the magnetic field line with transverse dimensions of hundreds kilometers in the ionosphere generated by a nonlinear evolution of Rayleigh-Taylor instability at the equatorial latitudes. The density depression reaches two or more orders and its region has a sharp boundary [McClure et al., 1977]. There are a variety of cases of field-aligned extension and radio noises observed at a satellite that may change depending on the case [Dyson and Benson, 1978]. When the bubble structure extends down to the bottomside penetrating the maximum electron density altitude, ground radio wave transmissions could be ducted by the bubble and reach the satellite. In this case, cosmic radio noise observations are affected by intense interferences, which results in the ANN misdetermining the in situ electron density to be lower than the actual electron density. There might be another case in which the electron density at the altitude of the F region peak inside the bubble is sufficiently high compared with the depleted topside electron density on the same magnetic field line. In other words, bubbles have relatively limited field-aligned dimensions and are abruptly terminated at both ends. In such a case, at the frequencies just above the local cutoff, the satellite is in a cavity that is shielded from ground radio wave transmissions and extraterrestrial radio noises. Observed cosmic radio noises should be those penetrated at the apex of the bubble where the ambient electron density is the lowest. Thus the ANN-determined electron density tends to be high compared with the real in situ electron density. In actual cases, the two effects compete and the results may be complicated. Such data should not be used for the training of the ANN.

[23] Figure 6 shows an example of the plasma bubble encounter. At frame numbers 34 through 37, the intense interference of ground radio transmissions are depicted as a dark area and the scaled fNs exhibit density depression. The corresponding ionograms exhibit severe backscatter echoes starting at zero range. The ANN-determined fNs also show density depression, but they are shallower than the scaled ones indicating the shielding effect. In this observation, the size of the bubble along the satellite pass appears to be fairly large because the satellite trajectory was almost in a magnetic meridional plane at the longitude of 50°W.

Figure 6.

Dynamic spectrum of radio noise and comparison between the scaled and estimated plasma frequencies for the orbit in which an intense plasma bubble was observed.

4.2. Solar Radio Bursts

[24] Another example of severe error sources is solar radio bursts. During a solar maximum period, the Sun often emits impulsive intense radio waves at short-wave and higher frequencies. The solar noise received at the satellite has an ionospheric cutoff depending on the solar zenith angle [Hartz and Roger, 1964]. When the solar zenith angle is large, solar radio noises do not affect the CRN observation near the local plasma cutoff. However, when the Sun is near the zenith, solar radio noises penetrate to the satellite even at frequencies close to the local cutoff. Such an example is shown in Figure 7. In the figure, the intense solar radio bursts observed are shown by the dark vertical lines. The white dots connected by a line indicate the solar zenith angle at the satellite.

Figure 7.

Same as Figure 6 but for the orbit in which intense solar radio bursts were observed. The white dots connected by a line are the solar zenith angle.

[25] The lower part of Figure 7 shows the ANN-determined and manually scaled fNs and their differences (this data set was not used for the training). We noted severe errors at frame numbers 57 to 59, and moderate errors at several previous frames. In these frames, solar noise intensity is high and its cutoff frequency is close to the normal CRN cutoff. While at frames 64 and 65, intense solar noises are observed but their cutoff frequencies are higher than the CRN cutoff because of the relatively high solar zenith angle (about 60°); nevertheless, fN was determined successfully.

[26] Thus, the intense solar radio noises adversely impact the application of the method. One possible way to overcome this effect is to train the ANN using a data set that includes solar radio burst events. This procedure resumed the accuracy somewhat for the observation affected by the solar noise, but overall efficiency was lowered. Instead of doing extended training to save such data, we will discard the data affected by intense solar radio noises near the local cutoff frequency during an application.

4.3. Solar Flux Changes

[27] Before applying the trained ANN to studying the ionosphere by using a large database, we further confirm the validity of the ANN for different geophysical conditions. The training of the ANN was conducted using a data set observed in a relatively short period in which solar radio flux at 10.7-cm wavelength or F10.7 solar activity index was around 200. However, the solar activity changed widely during the whole period. Changes in the solar activity result in changes in the atmospheric temperature, and therefore, changes in the scale height of the plasma distribution. Although Hartz and Roger [1964] calculated that the cone angle within which a radio signal reaches an antenna is not very sensitive to changes in the scale height, which is described in the introduction section, we examine if the ANN can predict a correct value of fN in the case of other solar activities.

[28] Figure 8 shows the ANN result for an orbit (August 12, 1978; about 6 months after the launch) when the monthly averaged F10.7 index was 114, which is the lowest level during the whole period. The difference between the scaled and the ANN-determined fNs is large at several frames, but no systematic errors are recognized. The relatively large errors at frame numbers 19, 24, and 82 are due to the interferences of ground transmissions and the intermodulation in the receiver circuit under the condition of a low ionospheric critical frequency when solar activity is low. Another example of ANN results is shown in Figure 9 for an orbit (November 10, 1980; close to the end of the satellite's mission life) when the monthly averaged F10.7 index was 218. The agreement of the scaled and ANN-determined fNs is quite good. Two examples indicate that the validity of the trained ANN does not change with solar activity at least during moderate to high activity levels, nor did any of the equipment appear to have deteriorated throughout the mission life.

Figure 8.

Same as Figure 5 for the lowest solar activity period (F10.7 = 114), which was 6 months after the launch.

Figure 9.

Same as Figure 5 for the highest solar activity period (F10.7 = 218), which was close to the end of the mission.

4.4. Geomagnetic Storms

[29] During big magnetic storms, the ionospheric distribution may be greatly disturbed, especially at high latitudes. Significant disturbances for applying the ANN are the scale height changes and the horizontal inhomogeneity. As the magnetic activity is not considered in the ANN, we will examine the accuracy of the ANN evaluation of the electron density for the case of a big storm that occurred on September 29, 1978, which was one of the most severe storms during the ISS-b observation period. The three hourly Kp index during the pass is 8 and the Kp summation on that day is 47-.

[30] The ANN result is shown in Figure 10. In the latter half of the pass, several points are missing for both the manual scalings and the ANN evaluations, where the ionogram data were not complete because of bit errors. Such dropouts sometimes occurred during severe magnetic storms. The peculiar feature of this pass is the irregular structure of the electron density at frame numbers 30 through 56, where the electron density largely varies from frame to frame. The corresponding ionograms exhibit severe backscattering signals near the zero range indicating large amplitude irregularities [Dyson and Winningham, 1974]. A detailed analysis of the irregularities is beyond this paper but we discuss the validity of the ANN application for extraordinary cases. The manually scaled and ANN-evaluated fNs show fairly good agreement except for a few points. Examining individual ionograms for those points, we note that the radio noise spectra are not smooth even below the critical frequency at frames 30 and 31. In these ionograms, extra traces of field-aligned ducting, which suggest localized low critical frequency nearby, and the radio noise spectra are probably affected by the ground transmissions. This results in the ANN-evaluated fN decreasing at frame number 31. On the other hand, if the spacecraft locates close to a wall of high electron density, the extraterrestrial radio noise half in the sky near the cutoff frequency might be blocked by the wall, which raises the ANN-evaluated plasma frequency. Although the substantiality of the density walls is not very clear from the topside soundings with low spatial resolution, presence of such walls is easily speculated from the severe backscattering signals and abrupt density changes along the satellite trajectory [Dyson and Winningham, 1974].

Figure 10.

Same as Figure 5 for during a severe magnetic storm.

[31] In conclusion, the trained ANN is confirmed to be valid even under severe magnetic storm conditions. However, the accuracy might be somewhat lowered when the spatial homogeneity is significantly broken at high latitudes.

5. Application to Global Mapping

[32] To demonstrate the potential of the ANN application to satellite data, we analyzed the global distribution of the ionosphere by using harmonic functional expansion. The orbit of the ISS-b satellite is nearly circular at 1000 km with an inclination angle of 70°, and four months are required to cover all local times at all latitudes due to the nature of the orbit. Three kinds of maps were drawn based on the data obtained during the period from November, 1979 to February, 1980, i.e., winter in the northern hemisphere. They are local time versus latitude, constant UT map, and constant LT map.

[33] Figure 11a shows the local time variation of the electron density (log ne; ne in cm−3) at various latitudes. Figure 11b is the distribution of the data points used for the functional expansion. As a latitude parameter, the modified dip latitude (λ*) [Jones et al., 1966] is used, where λ* is nearly equal to the dip angle or twice the magnetic latitude near the equator, and it becomes parallel to the geographic latitude at high latitudes. Thus, we note that the equatorial part is stretched in the latitudinal direction. As a longitude parameter in the functional expansion, the hour angle is used and, therefore, possible geographical variations are averaged over the whole longitude. Enhancement of the electron density centered at the magnetic equator can be clearly seen at two local time periods; one around 1400 LT and the other around 2000 LT. These enhancements are directly related to the local time variation of the zonal electric field which points eastward during daytime and westward during nighttime with a prereversal eastward enhancement in the evening; resulting E × B drift is upward (daytime and evening) and downward (nighttime). The density enhancement at 2000 LT corresponds to the strong prereversal enhancement of the upward E × B drift that is typically observed during a solar maximum period [Fejer et al., 1979]. The off-equatorial minimum around 0400 LT associated with a general hemispheric asymmetry is another interesting feature.

Figure 11.

Local time and latitudinal variation of the in situ electron density at about 1000 km drawn by the functional fitting for the data obtained during approximately four months which correspond to the northern hemisphere winter season (a) and distribution of the data points used for the mapping (b).

[34] We next expand the data to see a snapshot-like distribution of the ionosphere at a given time. Figure 12 shows the geographical distribution of the electron density for the constant universal time (UT map) centered at 0300 UT with a 3-hour width. The modified dip coordinate is used for the latitudinal expansion. The two density-enhanced regions are seen at 165° and 270° near the magnetic equator indicated by the dashed line. They are the same as the enhancement regions seen in Figure 11a, but quantitative differences are also noted reflecting possible longitudinal peculiarity.

Figure 12.

Geographical distribution of the in situ electron density at a fixed universal time (0300 UT ± 1.5 hr) during the same period as Figure 11.

[35] Although UT maps are intuitive, constant local time maps (LT maps) are suitable for examining longitudinal peculiarity. One of the distinguished features in Figures 11a and 12 is the density peak around 2000 LT at the magnetic equator. Figure 13 is the geographical variation of the density corresponding to the enhancement drawn for the constant local time condition centered at 2000 LT and for a 4-hour width. The enhancement clearly appears to be not homogeneous along the longitude but relatively high at the American longitudes. Such longitudinal peculiarity may reflect longitudinal variation of the Earth's magnetic field configuration; here we only demonstrate the potential of the ANN data processing, and a detailed analysis of the maps will be presented in the future.

Figure 13.

Geographical distribution of the in situ electron density at a fixed local time (2000 LT ± 2 hr) during the same period as Figure 11.

6. Summary

[36] We have developed a method to evaluate in situ electron density in the topside ionosphere from the cosmic radio noise cutoff by using an artificial neural network (ANN). The ANN successfully evaluates electron density for various conditions except for intense overhead solar radio burst events and plasma bubble encounters. Although the ANN is trained by using a data set under limited geophysical conditions, such as solar and geomagnetic activities, the ANN is validated to be applicable to extended geophysical conditions. This stands on the fact that the cosmic radio noise spectrum near the plasma cutoff frequency is determined by the extraterrestrial radio environment but does not depend greatly on the scale height of the ionospheric plasma distribution. Engineering parameters, including frequency characteristics of the receiver and an antenna peculiar to the spacecraft, as well as even the calibration curve of the raw telemetry data, are not explicitly required in the input data set; however, the ANN learns those parameters through the training. Therefore, the method can be applied to any other radio receiver systems if training data are available.

[37] The potential of applying the method to numerous ISS-b cosmic radio noise measurements was demonstrated by drawing global distribution maps of the topside electron density that reveal several interesting ionospheric structures. Thus, the method is believed to contribute to ionospheric studies. Detailed analyses of the global structure including a comparison with other experiments and existing empirical models will be presented in future.