A new incoherent scatter technique in the EISCAT Svalbard Radar



[1] A special incoherent scatter technique in the European Incoherent Scatter (EISCAT) Svalbard Radar (ESR) was tested in November 1999. After the 70-MHz IF stage the standard ESR receiver was replaced by a spectrum analyzer acting as a combination of a down-converter, an AD converter, and a quadrature detector and giving complex digital samples at a rate of 4 MHz. The samples were fed into a PCI-bus-based programmable digital I/O card, which performed a four-sample summing operation to give an effective sampling rate of 1 MHz, large enough to span all the frequency channels used in the experiment. Finally, the resulting samples were stored on hard disk. Hence the total multichannel signal was stored instead of separate lagged products for each frequency channel, which is the procedure in the standard hardware. This solution has some benefits; for example, the ground clutter can be eliminated with only a small loss in statistical accuracy, and the true phase of the transmitter waveform in phase-coded experiments can be measured better than in the present ESR system. In this paper the technical arrangements of this setup are presented. A new incoherent scatter experiment is also described, and the first results obtained using the hardware are shown.

1. Introduction

[2] A standard incoherent scatter measurement collects average values of lagged products of signal samples, which give estimates of the signal autocorrelation function at certain ranges and lags. In the European Incoherent Scatter (EISCAT) Svalbard Radar (ESR), which is located on Svalbard within the polar cap, the signal is down-converted to 7.5 ± 1.9 MHz passband in two stages and sampled at a rate of 10.0 MHz. Then the receiver branches out into six parallel digital channels. At each channel a chosen transmitting frequency is down-converted to zero frequency by means of numeric complex mixing and filtered by a low-pass digital filter. The samples then enter a sample buffer to be used by the digital signal processor, which calculates the lagged products. For details of the ESR radar, see Wannberg et al. [1997].

[3] The present paper describes an ESR experiment which stores the radar signal itself instead of its autocorrelation function. Special hardware is connected to the standard radar receiver, and hard disks are used which are large enough to store the high data flow. The rapidly increasing hard disk sizes and decreasing prices have now facilitated such measurements. Sampling and storing of the radar signal have been done before [e.g., Djuth et al., 1995]. The workstation-based Millstone Hill Incoherent Scatter Data Acquisition System (MIDAS-W) is capable of the same task, although it is designed to process the signal in real time [Holt et al., 2000].

[4] The purpose of this paper is to introduce the hardware, consisting of a spectrum analyzer, a fast programmable I/O card, a desktop computer, and large-capacity hard disks. The spectrum analyzer is used for down-conversion and sampling, and the computer stores the signal values on disks. The autocorrelation function estimates are computed afterward in the off-line signal processing. In addition, a new type of incoherent scatter experiment is presented, the advantages offered by storing the radar signal are explained, and finally, the applicability of this approach is demonstrated with real observations. No software for importing lag estimates into the standard incoherent scatter analysis package is yet available, and therefore the results are presented in terms of lag estimates rather than in terms of plasma parameters. Because of the occurrence of a sporadic-E layer within the observation period, the success of the method is convincingly demonstrated.

2. Experimental Setup

[5] In Figure 1 the standard ESR receiver is drawn on the left. The branch on the right portrays the new hardware. The spectra on the left and on the right (the widths of the spectrum peaks are exaggerated) describe the signal at different stages on the signal path in the conventional and in the present hardware, respectively. Here two transmitting frequencies around the nominal radar frequency of 500 MHz are assumed.

Figure 1.

(left) ESR receiver and (right) additional receiver connected into it, together with the incoherent scatter spectra at two frequencies in different stages of the signal paths.

[6] In the first mixer stage of the ESR receiver the nominal frequency is down-converted to 70 MHz. The oscillator frequency in the second mixer stage is 77.5 MHz, i.e., higher than 70 MHz, and therefore the spectra in the second IF stage lie around 7.5 MHz with inverted order of the spectrum peaks. Then the signal is sampled at a frequency of 10.0 MHz. This sampling rate produces down-conversion so that a group of spectral peaks will appear around zero frequency as shown by the fourth spectrum on the left-hand side of Figure 1. The group is repeated on the frequency axis at intervals of 10.0 MHz. The digital signal is then fed to parallel channel boards. There is one board for each transmitting frequency, and each board contains a numerical complex mixer, which converts a given transmitting frequency to zero. After subsequent filtering and decimation the spectrum consists only of the chosen peak, which is repeated at fixed intervals on the frequency axis. In Figure 1 this is 50 kHz, corresponding to a 20-μs sample interval after decimation. Finally, the samples are fed to a digital signal processor which calculates the lagged products.

[7] In the present experiment the ESR receiver is bypassed after the first IF amplifier. The 70-MHz signal is taken into a spectrum analyzer (Rohde & Schwarz, FSEM 30-B77), which is used here in an unconventional manner as a down-converter, AD converter, numerical quadrature detector, and digital filter. A cheaper and less sophisticated setup could quite well be used; the Rohde & Schwarz analyzer was chosen only because it was readily available.

[8] In the spectrum analyzer the signal is first mixed down to 6.4 MHz and then sampled by an AD converter at a frequency of 20.0 MHz. Thus the spectrum consists of groups of peaks at intervals of 20.0 MHz as indicated by the fourth spectrum on the right. Next the numerical complex mixer with a frequency of 6.4 MHz shifts the nominal radar frequency to zero. Then a digital filter and a decimator are used to choose the two peaks around zero frequency and to create their replicas at fixed intervals on the frequency axis. When the decimator produces complex samples at 4-MHz frequency, the spectrum of the output signal has groups of peaks at intervals of 4 MHz.

[9] The samples from the spectrum analyzer are fed to a PCI-bus-based programmable I/O card (GURSIP by Invers Oy, Finland). The card adds four samples together, which results in a sampling frequency of 1 MHz. A desktop computer (Power Mac G4) is finally used to store the data on hard disk.

[10] Unlike a conventional receiver which separates each transmission frequency into its own channel, this arrangement produces a single data stream which contains all frequency channels in a sequence of I/Q detected complex samples. Henceforth such data will be called multichannel complex data. The data flow is 4 Mb/s, which makes 14.4 GB/h. This means that the hard disks at the present time are big enough to allow measurements through a whole night.

3. Experiment

[11] In this paper, an incoherent scatter experiment designed for the new technique is presented. It consists of two different modulations at two frequencies (500.25 and 499.75 MHz) as shown in Figure 2. A 22-bit code with a sign sequence + − − − + + + − − − − + + + + + − + − + − + is first transmitted at 500.25 MHz. Each bit is further modulated by a 5-bit Barker code + + + − + with a bit length of 6 μs. After a 6-μs gap a second modulation with a 5-bit code + + + + − is transmitted at 499.75 MHz. Also this code is modulated by a 5-bit Barker code with a 6-μs bit length. Hence the bit length in the basic modulations is 30 μs, and the total lengths of the 22-bit and 5-bit codes are 660 μs and 150 μs, respectively. The length of the full transmission pattern is 816 μs.

Figure 2.

Full cycle of the applied experiment. Modulations at 499.75 MHz and 500.25 MHz are shown at the top and at the bottom, respectively.

[12] The experiment gives lag profiles at multiple values of the bit length of the 22-bit and 5-bit codes, i.e., at n × 30 μs, where n = 0, 1, …, 21. In addition to these full lags it is also possible to obtain fractional lags [Huuskonen et al., 1996]. In practice, fractional lags n × 30 ± 2 μs and n × 30 ± 4 μs are calculated. Inclusion of these lags can be used to improve the range resolution of the experiment.

[13] The transmissions are repeated at intervals of 4000 μs. This is such a short interval that no separate background measurement can be carried out. Calibration is arranged by having an 80-μs noise injection within every second transmission cycle. Then the total length of the experimental cycle is 8000 μs. The injection power is obtained by subtracting the power profile without injection from that with injection. When the scattering power and background power do not change, the subtraction gives the injection power.

[14] This experiment produces range ambiguities. In conventional experiments, range ambiguities are avoided because they are considered to be disastrous. This is not necessarily true, however, since statistical inversion can be used to remove the ambiguities. A second possibility is offered by a decoding filter as suggested by Sulzer [1989]. The 22-bit code in Figure 2 is constructed in such a manner that it behaves well in the inversion of as many lag profiles as possible. The choice of the code is based on a careful study of all 22-bit code sequences. In the selection the short lags are weighted most.

[15] The role of the 5-bit code at 499.75 MHz is to allow an easy inversion solution for lags 1–4. For these short lags, profiles from the 22-bit and the 5-bit codes can be combined through a set of linear equations, which can be solved by means of statistical inversion to give unambiguous lag profiles. In order to obtain unambiguous profiles for longer lags, regularization is applied in the inversion. This is made by putting the inversion solution to zero at the top and bottom of a profile and limiting its point-to-point variations statistically elsewhere in such a manner that it changes slowly except within sporadic-E layers. The method is similar to the regularization used by Nygrén et al. [1997] in satellite tomography. Although regularization is not necessary for lags 1–4, it is still used in order to improve the quality of the result.

4. Data Processing

[16] The measurements were carried out on 16 November 1999. Samples were taken at 1-μs intervals, including the pulse transmission times. Since the attenuated transmitted signal is available in the receiver, the data collected on hard disks contain both the transmission waveforms and the incoherent scatter signal with clutter and noise injection. All frequency channels are present in the same data flow. A sequence of samples taken between the start times of two successive transmission periods is called a sample profile. The data processing involves frequency channel separation, clutter removal, Barker decoding, calculation of lag profiles, removing range ambiguities, combining main and fractional lags, and finding plasma parameters by fitting the plasma theory to observed autocorrelation functions. All these steps, except the last one, are demonstrated in sections 4.14.5.

4.1. Channel Separation

[17] The bottom of Figure 3 displays the real part of two successive sample profiles of the multichannel complex data. These data points are taken at intervals of 1 μs, and they contain the two signals at ±0.25 MHz frequencies. Blowups of some parts of the experiment cycle are shown at the top of the figure.

Figure 3.

Real part of two successive sample profiles of multichannel complex data.

[18] Transmissions at the two frequencies are first seen as a strong signal at the bottom of Figure 3. The first blowup at the top shows the end of the transmission at 500.25 MHz (the 0.25-MHz signal) and the beginning of the transmission at 499.75 MHz (the −0.25-MHz signal), separated by the 6-μs gap.

[19] The second blowup at the top of Figure 3 portrays the end of the −0.25-MHz signal and the beginning of reception. The receiver is protected for a period of 18 μs after the transmission ends. When the protector is opened, initially a strong signal is received. Its strength is mainly due to ground clutter, which must be removed before low-altitude data can be analyzed.

[20] The third blowup at the top of Figure 3 displays the very end of the second sample profile. This consists of an incoherent scatter signal which is not contaminated by ground clutter. The contribution of noise injection increases the noise level during the time interval 7850–7930 μs.

[21] The channel detection is made by converting each frequency separately to zero and subsequent low-pass filtering. Filtering can be done in various ways, but the shortest length of the filter impulse response to separate the two channels is 2 μs. Therefore, in this paper, filtering is done simply by calculating two-sample averages. In practice, the program calculates a running average and a subsequent decimation so that the time resolution of the filtered samples is 2 μs. Results for the 0.25-MHz and −0.25-MHz signals are drawn in Figures 4 and 5 respectively. The two bottom panels contain the real and imaginary parts, and the top panels contain the phase angle calculated from these signals. Only the modulation part and the beginning of the received signal are shown in these figures.

Figure 4.

Phase and real and imaginary parts after the detection of the 500.25-MHz signal. Only the start of the profile showing the transmission and the beginning of the clutter signal is portrayed.

Figure 5.

Same as Figure 4, but for the 499.75-MHz signal.

[22] In Figure 4 the 0.25-MHz signal is converted to zero and the −0.25-MHz signal is filtered out. The 22-bit code with an embedded 5-bit Barker code is visible within the time interval 0–660 μs. The amplitude of the real part gradually decreases during the transmission, and correspondingly, the imaginary part increases. This is due to a progressive phase change in the transmitter, also visible in the top panel. This phase behavior does not remain the same, but changes from one experiment cycle to another. The 5-bit code (within the time interval 666–816 μs) is removed by filtering so that only peaks appear, indicating the steps in its real and imaginary parts. The scattering signal with ground clutter is seen after 834 μs at the end of the plots. Scattering due to the 5-bit code transmission is also suppressed by filtering; that is, channel separation takes place. The random character of the scattering signal is revealed by the stochastic behavior of its phase.

[23] In Figure 5 the −0.25-MHz signal is converted to zero frequency. Then the 0.25-MHz signal is shifted to 0.5 MHz and suppressed by filtering. The result is that the modulation pattern of the Barker-coded 5-bit code is visible within the time interval 666–816 μs and only the scattering signal due to this modulation is observed.

4.2. Clutter Removal

[24] The next step in data processing is to remove the clutter due to mountains and sea at long distances. Although this clutter is obtained from the sidelobes of the antenna, it dominates the scattering signal up to a range of 90 km. The clutter can be eliminated by making use of its long correlation time. The present practice in ESR experiments is to subtract two sample profiles at heights where disturbing clutter appears. The time separation of the profiles is chosen to be longer than the correlation time of the scattering signal but clearly shorter than that of the clutter signal. Then the clutter disappears in the subtraction. The differences of the profiles are used in calculating the lagged products in real time. The drawback of this procedure is that one half of the measurements are lost, so that a double integration time is needed in order to regain the statistical accuracy in the lower part of the lag profiles [see Turunen et al., 2000]. A method that does not suffer from this drawback is “DC subtraction,” which has been used in ESR measurements by Röttger [2000].

[25] A different method of clutter removal can be applied when, instead of the lagged products, the whole original signal is available from a longer time interval. Then one can subtract from each sample profile the mean of a number of sample profiles rather than a single one. This method was already discussed by Turunen et al. [2000], but it is not in routine use in the ESR radar. Assume that the ith sample profile is xi(j) = ci(j) + si(j) (here j = 1, 2, 3 … is the sample number), where ci(j) is the clutter profile and si(j) is the profile of the scattering signal plus noise. By subtracting a mean of n profiles from a single sample profile we obtain

equation image

[26] If n is not too great, the clutter terms in this equation cancel because of the long correlation time of the clutter signal. Also, the expectation value of the mean scattering signal is zero, and its variance decreases with increasing value of n. Hence, with a suitable value of n, yi(j) = si(j). The benefit of this method is that the elimination of the clutter has only a small effect on the signal variance.

[27] Tests showed that subtraction of a mean of 10 sample profiles suppresses the clutter equally well as does the method used in ESR. Yet the loss in signal statistics is negligible in comparison with the standard method. Increasing the number of sample profiles in the average would still slightly improve the statistics, but at some stage the clutter reduction would start getting worse. Tests showed that with an integration time of 2 s the subtraction of the mean does not work well anymore. This indicates that the clutter is not stationary in timescales of seconds.

[28] The top panel of Figure 6 portrays the real part of a single sample profile after detection of the 499.75-MHz channel (this is the same sample profile as that in the middle panel of Figure 5). Only the profile after the end of transmission is shown. The same profile after clutter suppression is plotted in the bottom panel (notice the order-of-magnitude difference in the scales of the vertical axes). Before suppression the clutter extends nearly to 1400 μs, i.e., about 600 μs after the end of transmission, corresponding to 90 km in range. The suppression removes the clutter very efficiently, so that it is negligible at about 100 μs after the end of the transmission. This corresponds to a range of 15 km, which suggests that the remaining clutter has a strong atmospheric component. It is also possible that the nonlinearity in the receiver front end has distorted the signal.

Figure 6.

Effect of clutter suppression for the 499.75-MHz signal.

4.3. Barker Decoding

[29] The last operation in the amplitude domain is Barker decoding. Conventionally, Barker decoding is done by convolution of the data with an inverted Barker code (i.e., by means of the matched filter). This generates sidelobes, which are not small in short Barker codes. In our case of a 5-bit code the total power in the sidelobes is 16% of the power in the central peak. Here we use, therefore, a different decoding method, which does not produce sidelobes. The method was originally suggested by Sulzer [1989]. There are other methods which are capable of suppressing the sidelobes [Hua and Oksman, 1990], but they do not remove them completely.

[30] The method can be understood in the following manner. We define a function

equation image

where the numbers bi are the coefficients of the nB-bit Barker code and Δ is the bit length. Next we assume that h(t) is the impulse response of a filter giving a δ function as an output when the input is b(t), i.e., (h * b)(t) = δ(t). Since the Fourier transform of a δ function is a constant, this gives H(ν)B(ν) = const, where H(ν) = equation image{h} and B(ν) = equation image{b} are the Fourier transforms of h(t) and b(t), respectively. By choosing the constant as a scaling factor of unity, one can then solve the impulse response of the filter as equation image.

[31] This impulse response consists of δ peaks with various heights and signs, and it has an infinite length. The peak heights rapidly approach zero in both directions. In decoding, however, the calculation of the impulse response is not needed. If y(t) is the signal before decoding, the decoded signal is yd(t) = equation image(t), and its Fourier transform is equation image. Hence the decoded signal is given by

equation image

[32] The application of this method does not require the sampling interval to be equal to the bit length of the Barker code. This is demonstrated in Figure 7, which corresponds to our case of a 5-bit Barker code with a 6-μs bit length and oversampling at 2-μs intervals.

Figure 7.

Method of decoding the 5-bit Barker code without sidelobes.

[33] The top left-hand panel in the top row of Figure 7 shows the oversampled 5-bit code, and the panel below shows its discrete Fourier transform. Correspondingly, the oversampled b(t) and the reciprocal of its discrete Fourier transform are plotted in the top two panels on the right. The latter is then multiplied by the Fourier transform of the sampled Barker code, and finally, the discrete inverse Fourier transform of the product is computed. The result is shown in the bottom panel, which indicates that the operation has compressed the Barker code into a single pulse with a length equal to the bit length. Thus no sidelobes are generated.

[34] In the case of the two modulations in Figure 2 the decoding applies separately to each 5-bit Barker code, and sequences of positive and negative peaks are obtained. The peak widths correspond to the bit length of the Barker code (i.e., 6 μs in the present case), and their separations correspond to the bit length of the basic phase code (i.e., 30 μs).

[35] All this means that when the method is applied to the received signal, the sampling interval can be any fraction of the bit length. The same sampling interval must be applied to b(t). Decoding is then carried out as indicated by equation (3) using discrete Fourier transforms of the samples of the signal and b(t).

4.4. Calculation of Lag Profiles

[36] After Barker decoding, the different lag profiles can be calculated. Each measured lag value is a weighted sum of the true lag values within the height interval covered by the range ambiguity function of this lag. The range ambiguity function has a peaked structure; in the case of the first full lag (30 μs) of the 5-bit code, for instance, it consists of four peaks. Each of them corresponds to a pair of successive bits in the 5-bit code. The peak separation is 30 × 150 m = 4.5 km.

[37] As an example, real parts of 30-μs lag profiles are plotted in Figure 8. The data are selected from a time period when an intense sporadic-E layer was present. The top panel shows the 30-μs lag profiles of the 5-bit and 22-bit codes. As expected, the layer is seen as four peaks (three positive and one negative) in the 5-bit lag profile. The separation of the peaks is 30 μs, which corresponds to the separation of the peaks in the range ambiguity function, i.e., 30 × 150 m = 4.5 km in range. Because of the thickness of the layer, the peaks are somewhat broader than the 6-μs bit length. If conventional Barker decoding were used, sidelobes would appear on both sides of the peaks. In principle, we would expect 21 peaks in the 22-bit lag profile. However, since the leading edge of the transmission has already passed the layer when the receiver protection is opened, only 14 peaks are observed.

Figure 8.

(top) Real parts of 30-μs lag profiles for the two modulations and (bottom) real part of the unambiguous lag profile obtained by means of inversion.

[38] The next step is the production of unambiguous lag profiles. This involves statistical inversion [see, e.g. Nygrén et al., 1997], and it will be only briefly outlined here. A detailed description will be given in a later work. In some previous works [Huuskonen et al., 1988; Pollari et al., 1989] the sidelobes of Barker codes were also removed by inversion. Here the problem is analogous except that the unwanted peaks are of the same size as the wanted peak.

[39] Since the sampling interval is 2 μs after decimation, the successive elements in a lag profile are separated by 300 m in range. Therefore we choose range gates with a 300-m gate separation, and the unknowns of the inversion problem are the true lag values at these gates. At altitudes above the highest gate the true lag value can be put to zero. The measured lag values are linear combinations of the unknowns, and the coefficients in these equations are determined by the range ambiguity function.

[40] The short lags are obtained from both the 22-bit and the 5-bit codes. Thus the number of measured lag values (i.e., the number of linear equations) is larger than the number of unknowns, and statistical inversion can be directly applied. In the case of longer lags the number of measurements and unknowns are the same, and therefore regularization is necessary in the inversion. The regularization method is similar to that used by Nygrén et al. [1997] in ionospheric tomography. The basic idea is to control statistically the point-to-point change of the lag value in terms of a suitable variance of this quantity. Although it is not necessary, the same regularization is applied to the short lags as well. The bottom panel of Figure 8 shows the real part of the 30-μs lag obtained by inverting the two lag profiles in the top panel. The result contains a single peak just below 105 km.

[41] In this manner, unambiguous lag profiles are obtained both for main lags and fractional lags. Each lag has its own range ambiguity function. Oversampling and fractional lags in the case of alternating codes was analyzed by Huuskonen et al. [1996], and they show the shapes and widths of range ambiguity functions thus formed. In the present experiment the width of the range ambiguity functions of the main lags is 900 m, as determined by the bit length (6 μs) of the 5-bit Barker code. The range ambiguity functions of fractional lags are more narrow; in the case of lags n × 30 ± 2 μs and n × 30 ± 4 μs the widths are 600 m and 300 m, respectively. Statistical inversion is used to combine the full and fractional lags to single lag values which have a range resolution of 300 m. Also this part of the analysis will be explained in greater detail in a separate work.

[42] The digital filtering and decimation lead to filtered samples at 2-μs intervals, which corresponds to a 300-m gate separation. If the data processing is started from the second sample in the sample profile instead of the first one, the range gates will be shifted upward by 150 m. When the two resulting lag profiles are merged, the gate separation will be 150 m.

4.5. Results

[43] An example of the resulting lag profiles is shown in Figure 9, which portrays the real part of the first lag at heights 85–135 km within the time interval 2030–2048 UT on 16 November 1999. These profiles are obtained from the 30-μs full lag and the fractional lags 26, 28, 32, and 34 μs. The temporal resolution is 10 s, and the gate separation is 150 m. Each lag value is indicated by a colored rectangle in an arbitrary color scale.

Figure 9.

Real part of the first lag obtained from the 30-μs full lag and the fractional lags 26, 28, 32, and 34 μs. The gate separation is 150 m, and the temporal resolution is 10 s. An arbitrary color scale is chosen to extend from dark blue to dark red.

[44] Since the autocorrelation function is scaled by the signal power, the observed variations reflect changes in electron density (although not only in electron density). The results reveal a slowly descending sporadic-E layer at about 110-km altitude. It consists of an intense layer at the bottom and a weaker, more structured layer at slightly greater heights. The thickness of both sublayers is about 1 km. In addition, steeply descending weak structures above the sporadic-E layer and an enhancement between 90 and 100 km altitudes are visible.

5. Discussion

[45] The signal samples taken in an incoherent scatter radar receiver could in principle be processed in real time or they could be stored for post processing. Since the sampling interval is of the order of 1–20 μs, a large memory capacity is needed for storing the flow of the original voltage samples. Therefore the most common approach is real-time processing which greatly compresses the data.

[46] A usual way of real-time processing is to calculate lagged products of the radar signal, averaged typically over 5–10 s periods. This implies that the various frequency channels are separated before calculating the lagged products. The frequency separation can take place either in the analog or in the digital part of the receiver. The former was the method in the old EISCAT UHF and VHF receivers [Folkestad et al., 1983], and the latter is used in the ESR [Wannberg et al., 1997] as well as in the EISCAT UHF and VHF radars after their refurbishing in 2000.

[47] Instead of lagged products, spectra can be calculated in real time from digital samples. Such a method, used in observations of plasma line enhancements due to ionospheric heating, was described by Djuth et al. [1986]. In this case the results were stored on magnetic tape after integration over 40–200 s. A similar method has also been used, for example, by Sulzer and Fejer [1991].

[48] An interesting means of circumventing the problem of storing large amounts of digital raw data has been applied by Djuth et al. [1990], who recorded the radar signal with an analogue tape recorder. The transmitted pulse shape of the 450-MHz Arecibo radar was also recorded. The data were then digitized with a separate data-taking system during playback with a reduced speed.

[49] More recently, storing quite large amounts of digital data has been possible. Djuth et al. [1995] report preserving about 45 min of unintegrated voltage samples on all data channels of their experiment. These authors emphasize the importance of this procedure; it allows changes to be made in the digital processing as desired and also investigating the radar returns on a pulse-to-pulse basis.

[50] With the development of workstations and desktop computers it has been realized that the data processing of the radar signals can be based solely on these tools and appropriate software instead of hardware designed for this specific purpose. The workstation-based Millstone Hill Incoherent Scatter Data Acquisition System (MIDAS-W) is described by Holt et al. [2000]. The 2.25-MHz IF signal is sampled by a commercial AD converter at 1-MHz rate, and the data samples are collected by a workstation. The data are then distributed via a fast network to other workstations carrying out different tasks in real time. These include digital filtering, decimation, I/Q separation, correlation, and integration, as well as real-time display and a possibility of storing raw data on disk.

[51] The data acquisition arrangements introduced in the present paper compare best with the MIDAS-W system. In both cases the IF signal is sampled directly, and the subsequent digital data processing is carried out by means of general-purpose computers rather than specialized hardware. Holt et al. [2000] use the Millstone Hill receiver, but we connect our own receiver to the ESR 70-MHz IF stage. In principle, the whole ESR receiver after the 500-MHz preamplifier could be replaced by our own, much cheaper equipment. The MIDAS-W system has real-time processing with tasks distributed to computers in a network, whereas we use a single computer which receives multichannel complex data from a programmable I/O card and stores them on disk. In our system, those tasks which MIDAS-W carries out in real time are left for off-line processing. Hence the hardware is more simple, and even the software is presumably less sophisticated, since there are no timing problems or communication tasks between the different workstations in the network.

[52] As pointed out already by Djuth et al. [1995], storing the radar signal has considerable benefits. It does not pose such limitations to data processing as real-time processing does. One is free to use any integration time, and data can be cleaned from disturbances like satellite echoes, for instance. The phase stability and the accuracy of the phase modulation can also be checked, which may be important in phase-coded experiments. The detection of phase allows us to observe pulse-to-pulse coherence of the received signal, which can be used for obtaining very precise velocity measurements of coherent objects. This makes the radar an ideal tool in the search for space debris. In the ESR, ground clutter creates a considerable difficulty. It can be removed from the stored raw data in a better way than it can in the standard ESR data acquisition system. A further benefit is that the statistical errors can be estimated more accurately than they can from the conventional data. It is also easier to design completely new types of experiments, as it is no longer necessary to design a real-time correlation algorithm.

[53] The radar experiment described in the present paper consists of two transmissions with different phase modulations. Unlike in standard experiments, these modulations produce range ambiguities. Each bit in the phase modulation is further modulated by a Barker code. Reception is oversampled in order to produce fractional lags, which improve the experiment.

[54] In the analysis, Barker decoding is done using a decoding filter which completely removes the sidelobes. This method was originally suggested by Sulzer [1989]; here it is applied to oversampled data using Fourier transforms as indicated by equation (3). Sulzer [1989] has also pointed out that the range ambiguities produced by a single phase code can be removed by means of an appropriate decoding filter. Instead, we use statistical inversion, which allows the inclusion of regularization. Statistical inversion is also used for combining measurements of full lags and fractional lags to produce single lag values. The bit patterns of the basic modulations were chosen in such a manner that they behave well in the inversion. A computer search of good codes is feasible, since the number of possible patterns in a 22-bit code is not too great.

[55] The purpose of presenting the experiment is to demonstrate that phase modulations producing range ambiguities do work in practice. The benefit of such codes is that transmissions over long cycles are not needed as in the case of alternating codes, for instance. This makes it easier to suppress the ground clutter. The experiment also facilitates extremely short integration periods, which may be an important property in rapidly changing ionospheric plasma.


[56] EISCAT is an international association supported by Finland (SA), France (CNRS), the Federal Republic of Germany (MPG), Japan (NIPR), Norway (NFR), Sweden (NFR), and the United Kingdom (PPARC). Financial support from the Academy of Finland is gratefully acknowledged. The work of B.D. is financed by the Finnish Graduate School in Astronomy and Space Physics. We are also grateful to T. Ulich for suggesting improvements in the manuscript.