An Autonomous Interference Detection and ... - Semantic Scholar
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
1
An Autonomous Interference Detection and Filtering Approach Applied to Wind Profilers V. K. Anandan and D. B. V. Jagannatham
Index Terms—Atmospheric radar, interference, moments estimation, MST radar, notch filter, wind profiler.
I. I NTRODUCTION
I
is generally observed that, for the Gadanki MST radar [13], the conventional method limits reliable Doppler profiling up to heights of typically 12–14 km. One approach to identify the atmospheric signal is using Gaussian fitting. By Gaussian fitting, all three lower order moments can be estimated. Most of the time, for a time-averaged signal, this holds good. It is also reported that atmospheric signals are not always having Gaussian distribution [2]. A very high frequency (VHF) radar is also capable of backscattering rain echoes, and the distribution of these echoes generally falls under Gamma distribution [17]. During this observation, if signals are contaminated with interference signals, moments of clear air and rain echoes through a simple fitting algorithm may not be accurate. To improve the performance of the peak detection algorithm, various approaches have been reported in literature. One might apply a consensus average or a median estimator, rather than a simple average, to the spectra before computing the spectral moments [4], [8], [15]. The main motivation for the consensus algorithm was to extend the reliable averages to low SNR. The problem with both median and consensus methods, however, is that they depend upon the number of samples of the desired data by around one-third of the total samples, and this could be lower if there are more samples for the average. Merritt [10] developed a more effective method that makes use of signal statistics to selectively average the data and that is not restricted by the number of contaminated samples. The method assumes only that the radar dwells on a particular volume of atmosphere, which is long enough for the atmosphere to be observed uncorrupted part of the time. Approaches to filtering the time-series data prior to spectral processing and modifications to spectral processing were considered in [7] and [9] to address clutter issues. All of these statistical averaging and filtering techniques are intended mainly to deal with spectral data contaminated with signals from nonatmospheric sources such as ground clutter, aircraft, birds, insects, etc. An adaptive moments estimation technique based on certain criteria, set up for the Doppler window, SNR, and wind shear parameters, which are used to adaptively track the signal in the range-Doppler spectral frame, has been developed by Anandan et al. [1] at the National Atmospheric Research Laboratory for the MST radar. The adaptive moments estimation algorithm has significant advantage in terms of better height coverage compared to the conventional single-peak detection method. For identifying signals from regions of low SNR, which would improve the reliability and height coverage of Doppler profiles, however, some kind of an adaptive method needs to
IE E Pr E oo f
Abstract—An autonomous interference detection and filtering algorithm has been developed for removing the interference bands generated in the Doppler spectra of mesosphere– stratosphere–troposphere (MST) radar signals. The technique, implemented with the MST radar at Gadanki (13.5◦ N, 79◦ E), is based on identifying interferencelike band signals using a statistical signal variance approach for fixing the amplitude threshold, through which detecting the interference frequency and designing an adaptable notch filter to filter the undesired frequency bands are done. Signals received from the MST radar are sometimes contaminated with interference signals received from other objects or generated within the system through arcing of high-power devices. Multiple interference bands with different characteristics are observed in the power spectra, which contaminate the wind information and other atmospheric signals. The autonomous interference detection and filtering algorithm is applied to various cases, and it is found that the interference signals could effectively be removed, leaving behind the original signals. By this approach, the effective number of signal samples obtained is increased, which helps one to improve the temporal resolution.
DENTIFYING the signal and computing the three loworder spectral moments are central to the problem of extracting information from the Doppler spectrum of the mesosphere–stratosphere–troposphere (MST) radar signal. These radars have mainly been used in lower atmospheric wind profiling, along with probing of mesosphere and ionosphere. The method of analyzing the MST radar spectral data is based on identifying the most prominent peak of the Doppler spectrum for each range gate and computing the three low-order spectral moments and signal-to-noise ratio (SNR) using the expressions given by Woodman [16]. SNR is defined as the ratio of total signal power to noise power in the coherent filter bandwidth. This simple method, however, has severe limitation in terms of height coverage, because the MST radar signals are characterized by rapidly falling SNR and often contaminated with interference signals of nonatmospheric origin. It
Manuscript received January 16, 2009; revised May 22, 2009 and August 3, 2009. V. K. Anandan is with the Radar Development Area, Indian Space Research Organisation Telemetry, Tracking and Command Network, Bangalore 560 058, India (e-mail: [email protected]). D. B. V. Jagannatham is with the Department of Electronics and Communication Engineering, C.R. Engineering College, Tirupati 507 501, India. Digital Object Identifier 10.1109/TGRS.2009.2034258
0196-2892/$26.00 © 2009 IEEE
2
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
II. AUTONOMOUS D ETECTION OF I NTERFERENCE F REQUENCY AND F ILTERING A LGORITHM An algorithm has been developed, which aims at filtering interference signals automatically in a range-Doppler spectral frame with background noise. Interference-frequency identification is carried out in the power spectral domain, and removal is done by applying a notch filter on time-series I & Q signals. The time-series data received from the radar is first subjected to power-spectrum computation using fast Fourier transform (FFT). Mean-noise-level estimation is carried out for determining the basic threshold value for the identification of interferencelike signals from each range gate. This procedure is repeated for all range gates and identifies the Doppler frequency above the threshold value set. In the second level, across all range gates, the selected frequency points were checked. Once the number of range gates (more than 15%) having the same Doppler frequency is found, that frequency will be selected for filtering purposes. Depending upon the value of interference frequency and known observational bandwidth, an adaptable second-order notch filter is designed. The original time-series data are passed through the notch filter. Once all interference frequencies are filtered, the data are subjected to power-spectrum computation, and from there, three lower order moments were estimated. The block diagram of the various steps involved in the algorithm is shown in Fig. 1, and the sequence of steps involved in the implementation of the algorithm will be detailed in the following.
IE E Pr E oo f
be used. An adaptive method based on constructing chains of profiles by maximizing an energy function and using a neural network approach for detecting the most likely profile has been developed by Clothiaux et al. [3]. The performance of the method has successfully been demonstrated with 404-MHz wind-profiler spectral data taken in low-altitude mode that showed extensive periods when either the SNR was poor or the atmospheric signal power was significantly less than that of the ground clutter. More recently, a wind confidence algorithm (NCAR improved moments algorithm—NIMA) and an automatic moments estimation technique (NCAR winds and confidence algorithm) was developed and implemented for wind profilers [5], [10]. The NIMA method implements combinational mathematical analysis, fuzzy logic synthesis, and global image processing algorithms. For the Gadanki MST radar having a peak power-aperture product of 3 × 1010 Wm2 , it has been found that the power spectra are sometimes contaminated with externally generated interference signals or arcing of high-power devices in the radar system. The system-generated interference will be rectified during the maintenance. However, for a given scientific observation, the received signal, along with interference, needs to be processed for retrieving the valuable information. Most of the algorithms that employ quality check depend upon the number of data points available. The moments estimation from spectra contaminated with interference signals is not possible with most of the aforementioned algorithms, and usually, those spectra that are contaminated with interference signals are not considered in the computation, which reduces the temporal resolution of the observation. Here, we present a new approach on interference detection and removal technique, which is developed and being used in signals received from the MST radar. The method is based on identifying the interference signal, which has defined characteristics in the coherent bandwidth using the statistical variance approach, and designing an appropriate notch filter. The algorithm checks the number of interference bands present in the observational bandwidth and filters them out automatically. This algorithm has specific advantage in analyzing the radar-backscattered echoes contaminated with interference signals of nonatmospheric origin. Generally, highresolution wind profiles are recorded at 30-s interval, and if the data are contaminated with multiple interference signals, analysis will become very difficult. The observational parameters of atmospheric radars are programmable as per the user requirement. Depending upon the interpulse period and pulsewidth, the observational bandwidth and range resolution will be changing. Thus, an adaptable filtering approach is most desirable for the removal of unwanted signals. The autonomous interference-signal detection and removal algorithm helps in analysis without manual intervention by using an adaptable notch filter. The algorithm is applied to various types of signals under varied conditions and has shown consistent results. A stepwise description of the algorithm applied to the data from the MST radar is given in Section II. The results are discussed in Section III, and the important conclusions are presented in Section IV.
A. Power-Spectrum Computation and Noise Removal (Step 1)
After preserving the original time-series data in memory, the presence of interference signals is being checked. The timeseries data received from the radar are subjected to powerspectrum estimation using an FFT algorithm. The raw Doppler power spectra computed are subjected to low-pass filtering (i.e., smoothing) to reduce the level of noise fluctuations that appear particularly prominent in low-SNR regions. Low-pass filtering is implemented with a three-point running average of the Doppler spectrum. Then, the mean noise level is estimated for each range gate using an objective method based on Gaussian statistics [6]. In a very general sense, it can be viewed as the point at which the ratio of mean square to variance of noise power spectral density crosses unity, which is considered to be the mean noise floor of the series. The brief description of the procedure is as follows. The spectrum is reordered in ascending order for identifying the mean noise level. The original series of power spectra is identified in terms of Pi (i = 1, . . . , N ). The reordered spectra in ascending order is represented as Ai (i = 1, . . . , N ). There are two variables to be determined (P and Q). From the lowest spectral points onward, these variables were computed and checked whether their ratio (RM ) crossed the threshold or not as per (3). Index i corresponds to the spectral point being used, and it varies from one to N , i.e., the series length. Variable M corresponds to the number of spectral points being used to check the threshold value (RM ). The noise-level threshold shall be estimated to the maximum level L [(4)] such that the set of
ANANDAN AND JAGANNATHAM: APPROACH APPLIED TO WIND PROFILERS
3
where M is the number of spectra that was averaged for obtaining the data. Step 3) Noise level (L) = Pk , M, where k = min|1,
such that RM >1 if no M meets the aforementioned criterion
. (4)
Because the mean noise level represents the statistical property of the medium under observation, it is used as the basic criterion for deciding the threshold value of signal detection in each range gate. The mean noise level for each range gate is subtracted from the corresponding power spectrum before the interference-signal detection process starts. B. Threshold Detection and Identification of Interference Frequencies
IE E Pr E oo f
Fig. 1.
It is observed that, in general, interference frequency has defined characteristics with nearly constant amplitude and appears at the same frequency in multiple range gates. In the case of system-generated interferences, multiple frequencies were observed, having spread in limited number of range gates. This makes the detection altogether difficult with the normal technique. To reduce the possibility of false alarm of identifying the interference frequency, the threshold value is chosen to be not less than 3 dB above the mean noise floor. The 3-dB threshold is chosen after careful evaluation of the noise power spectral density for a number of observations. It is observed that the standard deviation of noise power over its mean is around 3 dB. Thus, choosing the threshold value of 3 dB makes sure that no noise fluctuations were identified as interference frequency. The running average will reduce the noise spikes in the power spectrum, which will bring down the variance of the noise spectrum, thereby reducing the threshold value. Once the signal is identified above the threshold value, the same is carried out on the entire range gate on the same frequency bin in the frame. If, any time, more than 15% of the total range gates in the frame are continuously reported above the threshold value, then that frequency is considered as interference frequency. This procedure is repeated until all the frequency bins have been checked in the entire range gate in the Doppler power spectrum. The observation window set for the lower atmosphere is about 22.5 km, which corresponds to 150 range gates for a resolution of 150 m. The atmospheric signal seldom falls on the same Doppler bin from range bin to range bin, and the interference signal does as well. Therefore, the continuity checking on the range gates for the same frequency helps one to isolate the atmospheric signal that is wrongly selected as interference frequency.
Algorithm flowchart of an autonomous interference filter.
spectral points below the level nearly satisfies the criterion. The details of the steps involved in the algorithm for mean-noiselevel determination are given as follows: Step 1) Reorder the spectrum {Pi , i = 1, . . . , N } in ascending order. Let this sequence be written as {Ai , i = 1, . . . , N } and Ai < Aj for i < j Step 2) Compute the following: PM = QM =
N Ai (i) i=1 N A2 i
i=1
and if QM > 0, RM =
i
(1) C. Interference Filter Design −
2 PM
2 PM , (QM ∗ M )
(2) for M = 1, . . . , N (3)
A notch filter is a filter that contains one or more deep notches or, ideally, perfect nulls in its frequency-response characteristics. Notch filters are useful in many applications where specific frequency components must be eliminated. To create a null in the frequency response of a filter at a frequency ω0 by
4
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
Fig. 2. Notch-filter frequency–magnitude responses.
introducing a pair of complex conjugate zeros on the unit circle at an angle ω0 [12] Z1,2 = e±jω0 .
(5)
Thus, the system function for a finite-impulse response (FIR) notch filter is simply
IE E Pr E oo f
H(Z) = b0 (1 − ejω0 z −1 )(1 − e−jω0 z −1 ) = b0 1 − 2 cos ω0 z −1 + z −2 .
(6)
In general, in an FIR notch filter, the notch has a relatively large bandwidth, which means that the other frequency components around the desired null are severely attenuated. To reduce the bandwidth of the null, a more sophisticated and longer FIR filter could be designed. It is also possible to improve the frequency-response characteristics by introducing poles in the system function. The purpose of the poles is to introduce a resonance in the vicinity of the null and to therefore reduce the bandwidth of the notch. If we place a pair of complex poles at P1,2 = re±jω0
1 − 2 cos ω0 Z −1 + Z −2 . 1 − 2r cos ω0 Z −1 + r2 Z −2
(8)
Fig. 2 shows the frequency–magnitude responses of the notch filter with poles located at 0.95 and 0.80. For the demonstration of filter responses, a notch frequency of 2 Hz and a signal bandwidth of 8 Hz are chosen. A notch depth that is better than −50 dB and notch bandwidths of 0.29 Hz at −3 dB (around 3.6% of filter bandwidth) and 0.023 Hz at −20 dB (around 0.28% of filter bandwidth) are observed in the frequency–magnitude-response curve for r at 0.95. Currently, the value of r at 0.95 is chosen for the interference filter design. If further reduction in notch bandwidth is desired, that can be achieved by designing a higher order filter. The difference equation of the notch filter is given by Y (n) = X(n) − 2 cos ω0 X(n − 1) + X(n − 2) + 2r cos ω0 Y (n − 1) − r2 Y (n − 2).
for a number of frequency points identified for removal. Once the filtering process is over, power-spectrum computation is carried out for further processing of the signal. The first three lower order moments are estimated from the power spectrum, which represents the total power, mean Doppler, and Doppler width. III. R ESULTS AND D ISCUSSION
(7)
the system function for the resulting filter is H(Z) = b0
Fig. 3. Range-versus-Doppler-power-spectrum plot of atmospheric signals contaminated with interference signals.
(9)
The aforesaid equation is being used for filtering out the interference bands in the signal. The filtering process continues
The algorithm developed has been tested on a number observations related to atmospheric signals received from the Gadanki MST radar under various conditions. Two cases are presented here, i.e., one for the observation of the lower atmosphere used for wind profiling up to an altitude of 22 km, and the other one for the observation related to the ionosphere. Fig. 3 shows the Doppler power spectra of the signals for the range of 3.6–21.7 km. The plots are normalized to individual range gates. In general, the radar-backscattered power reduces at a rate of (range)−2 for the atmospheric targets. Therefore, as range increases, the returned power reduces. The power received from the radar is recorded in arbitrary value and not in absolute. To have a better presentation, plots are range normalized to show all the range gates with strongest signal as reference. To show the entire range gates (around 150) having signals, individual range-normalized plots are shown. There are three interference bands observed in the spectra at around 1, 5.1, and −3.8 Hz. It is observed that interferences also have varied characteristics. The interference at 1 Hz is present throughout the range gates. In the lower range gates, it is not visible because atmospheric signals are stronger than interference signals. The interference at 5.1 Hz starts at around 14 km and continues until the end of the range gate, whereas
5
IE E Pr E oo f
ANANDAN AND JAGANNATHAM: APPROACH APPLIED TO WIND PROFILERS
Fig. 5. Range-versus-Doppler-power-spectrum plot of the lower atmospheric wind profile signal contaminated with interference signals. The mean Doppler estimated in each range bin is plotted along with it.
Fig. 4. Range-versus-Doppler power-spectrum plot of the lower atmospheric wind profile signal after removing the interference signals.
the interference at 3.8 Hz starts around 8 km and ends at 14.5 km. It is observed that, in general, the externally generated interference present throughout the Doppler-range gate spectra and intermittent interference bands are usually generated from high-power system arcing or malfunctioning. Fig. 4 shows the spectra after applying the algorithm. The algorithm could detect all the interference bands and remove them effectively. The wind profile is clearly visible up to 21 km. Once the interference bands are removed, the mean noise level is estimated for each range gate using an objective method based on Gaussian statistics [6]. The mean noise level for each range gate is subtracted from the corresponding power spectrum. The three lower order moments are estimated using the expressions given by Woodman [16], which correspond to the total power, mean Doppler, and Doppler width. The strongest peak available in the spectrum is identified for the estimation of moments. The estimation of moments before and after filtering is shown in Figs. 5 and 6. The mean Doppler estimated is plotted against the Doppler power spectra for all range gates. From Fig. 5, it is observed that detection fails after 12 km because interference signals dominate over atmospheric backscattered signals. Whereas in the case of filtered signal spectrum, the signal could be detected correctly, and the wind profile is estimated up to 21 km, as shown in Fig. 6. In the second case, the observation related to the ionospheric signal is presented. The wind profiler works at VHF, which is also capable of detecting ionospheric signals when the beam is pointed transverse to the Earth’s magnetic field. This observation is taken when the radar beam is pointed 14◦ off zenith. Fig. 7 shows the 3-D plot of ionospheric observation for a height range of 99–129 km. There are two interference
Fig. 6. Range-versus-Doppler-power-spectrum plot of the lower atmospheric wind profile signal after the removal of interference signals. The mean Doppler estimated in each range bin is plotted along with it.
bands observed in the power spectra around 60 and −31 Hz. Ionospheric signals are generally sporadic in nature and very weak during this observation period. Ionospheric signals are present at two range gates at 111 and 112.2 km. The interference signal at −31 Hz passes through atmospheric signals. The interference removal algorithm is applied to this data frame. Fig. 8 shows the power-spectrum plot after the interference removal. Here, the interference bands are also effectively removed from the power spectrum and show very clearly signals
6
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
from the atmospheric radar. There are two cases presented, i.e., one for lower atmospheric signals corresponding to the wind profile up to an altitude of 21 km, and the other for ionospheric signals where the interference bands go through atmospheric signals. In both cases, multiple interference bands were removed effectively. It has also been demonstrated that moments estimation is improved by removing the interference bands and detecting the signal at higher heights having less SNR. The algorithm has been applied to a number of data frames that are contaminated with interference signals, and in all cases, the algorithm could detect and remove the interference bands. This interference detection and filtering algorithm helps in reducing the number of frames to be rejected during quality checking, thereby obtaining high temporal resolution in atmospheric wind profiling. R EFERENCES
IE E Pr E oo f
[1] V. K. Anandan, P. Balamuralidhar, P. B. Rao, A. R. Jain, and C. J. Pan, “An adaptive moments estimation technique applied to MST radar echoes,” J. Atmos. Ocean. Technol., vol. 22, no. 4, pp. 396–408, Apr. 2005. [2] V. K. Anandan, G. Ramachandra Reddy, and P. B. Rao, “Spectral analysis of atmospheric radar signal using higher order spectral estimation technique,” IEEE Trans. Geosci. Remote Sens., vol. GRS-39, no. 9, pp. 1890– 1895, Sep. 2001. [3] E. E. Clothiaux, R. S. Rene, D. W. Thomson, T. P. Ackerman, and S. R. Williams, “A first-guess feature-based algorithm for estimating wind speed in clear-air Doppler radar spectra,” J. Atmos. Ocean. Technol., vol. 11, no. 4, pp. 888–908, Aug. 1994. [4] M. A. Fischler and R. C. Boltes, “Random sample consensus: A paradigm for model fitting with application to image analysis and automated cartography,” Commun. ACM, vol. 24, no. 6, pp. 381–395, Jun. 1981. [5] R. K. Goodrich, C. S. Morse, L. B. Cornman, and S. A. Cohn, “A horizontal wind and wind confidence algorithm for Doppler wind profilers,” J. Atmos. Ocean. Technol., vol. 19, no. 3, pp. 257–273, Mar. 2002. [6] P. H. Hildebrand and R. S. Sekhon, “Objective determination of the noise level in Doppler spectra,” J. Appl. Meteorol., vol. 13, no. 7, pp. 808–811, Oct. 1974. [7] J. R. Jordan and R. J. Latatis, “A potential source of bias in horizontal winds estimated using a 915-MHz acoustically enhanced profiler,” J. Atmos. Ocean. Technol., vol. 14, no. 3, pp. 543–546, Jun. 1997. [8] P. T. May and R. G. Strauch, “An examination of wind profiler signal processing algorithms,” J. Atmos. Ocean. Technol., vol. 6, no. 4, pp. 731– 735, Aug. 1989. [9] P. T. May and R. G. Strauch, “Reducing the effect of ground clutter on wind profiler velocity measurements,” J. Atmos. Ocean. Technol., vol. 15, no. 2, pp. 579–586, Apr. 1998. [10] D. A. Merritt, “A statistical averaging method for wind profiler Doppler spectra,” J. Atmos. Ocean. Technol., vol. 12, no. 5, pp. 985–995, Oct. 1995. [11] C. S. Morse, R. K. Goodrich, and L. B. Cornman, “The NIMA method for improved moment estimation from Doppler spectra,” J. Atmos. Ocean. Technol., vol. 19, no. 3, pp. 274–295, Mar. 2002. [12] J. G. Proakis and D. G. Manolakis, Digital Signal Processing, 3rd ed. Englewood Cliffs, NJ: Prentice-Hall, 1996. [13] P. B. Rao, A. R. Jain, P. Kishore, P. Balamuralidhar, S. H. Damle, and G. Viswanathan, “Indian MST radar. I: System description and sample vector wind measurements in ST mode,” Radio Sci., vol. 30, no. 4, pp. 1125–1138, 1995. [14] R. G. Strauch, B. L. Weber, A. S. Frisch, C. G. Little, D. A. Merritt, K. P. Moran, and D. C. Welsh, “The precision and relative accuracy of profiler wind measurements,” J. Atmos. Ocean. Technol., vol. 4, no. 4, pp. 563–571, Dec. 1987. [15] T. L. Wilfong, S. A. Smith, and R. L. Creasey, “High temporal resolution velocity estimates from a wind profiler,” J. Spacecr. Rockets, vol. 30, no. 3, pp. 348–354, May 1993. [16] R. F. Woodman, “Spectral moments estimation in MST radars,” Radio Sci., vol. 20, pp. 1185–1195, 1985. [17] C. W. Ulbrich, “Natural variations in the analytical form of the raindrop size distribution,” J. Clim. Appl. Meteorol., vol. 22, no. 10, pp. 1764–1775, Oct. 1983.
Fig. 7. Range-versus-Doppler-power-spectrum plot of ionospheric signals contaminated with interference signals.
Fig. 8. Range-versus-Doppler-power-spectrum plot of ionospheric signals after removing the interference signals.
at 111 and 112.2 km. The algorithm is tested on a number of data frames corrupted with interference and shows consistency in performance by removing the interference signals, leaving behind the atmospheric signals. IV. C ONCLUSION An autonomous interference detection and filtering algorithm has been developed and successfully tested on signals received
ANANDAN AND JAGANNATHAM: APPROACH APPLIED TO WIND PROFILERS
D. B. V. Jagannatham received the B.Tech. degree in electronics and communication engineering from Andhara University, Waltair, India, in 1990 and the M.Tech. degree from Jawaharlal Nehru Technological University, Hyderabad, India, in 1999, where he is currently working toward the Ph.D. degree in digital signal processing, mainly on nonlinear signal processing applications. He is currently a Professor with the Department of Electronics and Communication Engineering, C.R. Engineering College, Tirupati, India. His areas of interest are radar signal processing, algorithm development, and digital system design.
IE E Pr E oo f
V. K. Anandan received the M.Sc. and M.Tech. degrees from Cochin University of Science and Technology, Kochi, India, and the Ph.D. degree in electronics and communication engineering from Sri Venkateswara University, Tirupati, India. He has done his postdoctoral research at National Central University, Chung-Li, Taiwan. He joined the Indian Space Research Organisation (ISRO), Department of Space, Government of India, in 1991. He was responsible for developing real-time data acquisition and signal processing system for mesosphere–stratosphere–troposphere radar and closely associated with other atmospheric radar developments. He has developed the signal and data processing software package for atmospheric radars and sodars. He is currently with the Radar Development Area, ISRO Telemetry, Tracking and Command Network, Bangalore. His research contribution is in new signal processing algorithm development related to atmospheric radars, precipitation studies, and atmospheric boundary layer. He is currently involved in the development of wind profiler radar and cloud radar systems. He has published more than 40 research papers. His area of interest includes radar system developments, signal processing algorithm developments, and studies related to remote sensing of the atmosphere using different observational systems. Dr. Anandan is a Fellow of The Institution of Electronics and Telecommunication Engineers and an associate fellow of the Andhra Pradesh Academy of Sciences.
7
1
An Autonomous Interference Detection and Filtering Approach Applied to Wind Profilers V. K. Anandan and D. B. V. Jagannatham
Index Terms—Atmospheric radar, interference, moments estimation, MST radar, notch filter, wind profiler.
I. I NTRODUCTION
I
is generally observed that, for the Gadanki MST radar [13], the conventional method limits reliable Doppler profiling up to heights of typically 12–14 km. One approach to identify the atmospheric signal is using Gaussian fitting. By Gaussian fitting, all three lower order moments can be estimated. Most of the time, for a time-averaged signal, this holds good. It is also reported that atmospheric signals are not always having Gaussian distribution [2]. A very high frequency (VHF) radar is also capable of backscattering rain echoes, and the distribution of these echoes generally falls under Gamma distribution [17]. During this observation, if signals are contaminated with interference signals, moments of clear air and rain echoes through a simple fitting algorithm may not be accurate. To improve the performance of the peak detection algorithm, various approaches have been reported in literature. One might apply a consensus average or a median estimator, rather than a simple average, to the spectra before computing the spectral moments [4], [8], [15]. The main motivation for the consensus algorithm was to extend the reliable averages to low SNR. The problem with both median and consensus methods, however, is that they depend upon the number of samples of the desired data by around one-third of the total samples, and this could be lower if there are more samples for the average. Merritt [10] developed a more effective method that makes use of signal statistics to selectively average the data and that is not restricted by the number of contaminated samples. The method assumes only that the radar dwells on a particular volume of atmosphere, which is long enough for the atmosphere to be observed uncorrupted part of the time. Approaches to filtering the time-series data prior to spectral processing and modifications to spectral processing were considered in [7] and [9] to address clutter issues. All of these statistical averaging and filtering techniques are intended mainly to deal with spectral data contaminated with signals from nonatmospheric sources such as ground clutter, aircraft, birds, insects, etc. An adaptive moments estimation technique based on certain criteria, set up for the Doppler window, SNR, and wind shear parameters, which are used to adaptively track the signal in the range-Doppler spectral frame, has been developed by Anandan et al. [1] at the National Atmospheric Research Laboratory for the MST radar. The adaptive moments estimation algorithm has significant advantage in terms of better height coverage compared to the conventional single-peak detection method. For identifying signals from regions of low SNR, which would improve the reliability and height coverage of Doppler profiles, however, some kind of an adaptive method needs to
IE E Pr E oo f
Abstract—An autonomous interference detection and filtering algorithm has been developed for removing the interference bands generated in the Doppler spectra of mesosphere– stratosphere–troposphere (MST) radar signals. The technique, implemented with the MST radar at Gadanki (13.5◦ N, 79◦ E), is based on identifying interferencelike band signals using a statistical signal variance approach for fixing the amplitude threshold, through which detecting the interference frequency and designing an adaptable notch filter to filter the undesired frequency bands are done. Signals received from the MST radar are sometimes contaminated with interference signals received from other objects or generated within the system through arcing of high-power devices. Multiple interference bands with different characteristics are observed in the power spectra, which contaminate the wind information and other atmospheric signals. The autonomous interference detection and filtering algorithm is applied to various cases, and it is found that the interference signals could effectively be removed, leaving behind the original signals. By this approach, the effective number of signal samples obtained is increased, which helps one to improve the temporal resolution.
DENTIFYING the signal and computing the three loworder spectral moments are central to the problem of extracting information from the Doppler spectrum of the mesosphere–stratosphere–troposphere (MST) radar signal. These radars have mainly been used in lower atmospheric wind profiling, along with probing of mesosphere and ionosphere. The method of analyzing the MST radar spectral data is based on identifying the most prominent peak of the Doppler spectrum for each range gate and computing the three low-order spectral moments and signal-to-noise ratio (SNR) using the expressions given by Woodman [16]. SNR is defined as the ratio of total signal power to noise power in the coherent filter bandwidth. This simple method, however, has severe limitation in terms of height coverage, because the MST radar signals are characterized by rapidly falling SNR and often contaminated with interference signals of nonatmospheric origin. It
Manuscript received January 16, 2009; revised May 22, 2009 and August 3, 2009. V. K. Anandan is with the Radar Development Area, Indian Space Research Organisation Telemetry, Tracking and Command Network, Bangalore 560 058, India (e-mail: [email protected]). D. B. V. Jagannatham is with the Department of Electronics and Communication Engineering, C.R. Engineering College, Tirupati 507 501, India. Digital Object Identifier 10.1109/TGRS.2009.2034258
0196-2892/$26.00 © 2009 IEEE
2
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
II. AUTONOMOUS D ETECTION OF I NTERFERENCE F REQUENCY AND F ILTERING A LGORITHM An algorithm has been developed, which aims at filtering interference signals automatically in a range-Doppler spectral frame with background noise. Interference-frequency identification is carried out in the power spectral domain, and removal is done by applying a notch filter on time-series I & Q signals. The time-series data received from the radar is first subjected to power-spectrum computation using fast Fourier transform (FFT). Mean-noise-level estimation is carried out for determining the basic threshold value for the identification of interferencelike signals from each range gate. This procedure is repeated for all range gates and identifies the Doppler frequency above the threshold value set. In the second level, across all range gates, the selected frequency points were checked. Once the number of range gates (more than 15%) having the same Doppler frequency is found, that frequency will be selected for filtering purposes. Depending upon the value of interference frequency and known observational bandwidth, an adaptable second-order notch filter is designed. The original time-series data are passed through the notch filter. Once all interference frequencies are filtered, the data are subjected to power-spectrum computation, and from there, three lower order moments were estimated. The block diagram of the various steps involved in the algorithm is shown in Fig. 1, and the sequence of steps involved in the implementation of the algorithm will be detailed in the following.
IE E Pr E oo f
be used. An adaptive method based on constructing chains of profiles by maximizing an energy function and using a neural network approach for detecting the most likely profile has been developed by Clothiaux et al. [3]. The performance of the method has successfully been demonstrated with 404-MHz wind-profiler spectral data taken in low-altitude mode that showed extensive periods when either the SNR was poor or the atmospheric signal power was significantly less than that of the ground clutter. More recently, a wind confidence algorithm (NCAR improved moments algorithm—NIMA) and an automatic moments estimation technique (NCAR winds and confidence algorithm) was developed and implemented for wind profilers [5], [10]. The NIMA method implements combinational mathematical analysis, fuzzy logic synthesis, and global image processing algorithms. For the Gadanki MST radar having a peak power-aperture product of 3 × 1010 Wm2 , it has been found that the power spectra are sometimes contaminated with externally generated interference signals or arcing of high-power devices in the radar system. The system-generated interference will be rectified during the maintenance. However, for a given scientific observation, the received signal, along with interference, needs to be processed for retrieving the valuable information. Most of the algorithms that employ quality check depend upon the number of data points available. The moments estimation from spectra contaminated with interference signals is not possible with most of the aforementioned algorithms, and usually, those spectra that are contaminated with interference signals are not considered in the computation, which reduces the temporal resolution of the observation. Here, we present a new approach on interference detection and removal technique, which is developed and being used in signals received from the MST radar. The method is based on identifying the interference signal, which has defined characteristics in the coherent bandwidth using the statistical variance approach, and designing an appropriate notch filter. The algorithm checks the number of interference bands present in the observational bandwidth and filters them out automatically. This algorithm has specific advantage in analyzing the radar-backscattered echoes contaminated with interference signals of nonatmospheric origin. Generally, highresolution wind profiles are recorded at 30-s interval, and if the data are contaminated with multiple interference signals, analysis will become very difficult. The observational parameters of atmospheric radars are programmable as per the user requirement. Depending upon the interpulse period and pulsewidth, the observational bandwidth and range resolution will be changing. Thus, an adaptable filtering approach is most desirable for the removal of unwanted signals. The autonomous interference-signal detection and removal algorithm helps in analysis without manual intervention by using an adaptable notch filter. The algorithm is applied to various types of signals under varied conditions and has shown consistent results. A stepwise description of the algorithm applied to the data from the MST radar is given in Section II. The results are discussed in Section III, and the important conclusions are presented in Section IV.
A. Power-Spectrum Computation and Noise Removal (Step 1)
After preserving the original time-series data in memory, the presence of interference signals is being checked. The timeseries data received from the radar are subjected to powerspectrum estimation using an FFT algorithm. The raw Doppler power spectra computed are subjected to low-pass filtering (i.e., smoothing) to reduce the level of noise fluctuations that appear particularly prominent in low-SNR regions. Low-pass filtering is implemented with a three-point running average of the Doppler spectrum. Then, the mean noise level is estimated for each range gate using an objective method based on Gaussian statistics [6]. In a very general sense, it can be viewed as the point at which the ratio of mean square to variance of noise power spectral density crosses unity, which is considered to be the mean noise floor of the series. The brief description of the procedure is as follows. The spectrum is reordered in ascending order for identifying the mean noise level. The original series of power spectra is identified in terms of Pi (i = 1, . . . , N ). The reordered spectra in ascending order is represented as Ai (i = 1, . . . , N ). There are two variables to be determined (P and Q). From the lowest spectral points onward, these variables were computed and checked whether their ratio (RM ) crossed the threshold or not as per (3). Index i corresponds to the spectral point being used, and it varies from one to N , i.e., the series length. Variable M corresponds to the number of spectral points being used to check the threshold value (RM ). The noise-level threshold shall be estimated to the maximum level L [(4)] such that the set of
ANANDAN AND JAGANNATHAM: APPROACH APPLIED TO WIND PROFILERS
3
where M is the number of spectra that was averaged for obtaining the data. Step 3) Noise level (L) = Pk , M, where k = min|1,
such that RM >1 if no M meets the aforementioned criterion
. (4)
Because the mean noise level represents the statistical property of the medium under observation, it is used as the basic criterion for deciding the threshold value of signal detection in each range gate. The mean noise level for each range gate is subtracted from the corresponding power spectrum before the interference-signal detection process starts. B. Threshold Detection and Identification of Interference Frequencies
IE E Pr E oo f
Fig. 1.
It is observed that, in general, interference frequency has defined characteristics with nearly constant amplitude and appears at the same frequency in multiple range gates. In the case of system-generated interferences, multiple frequencies were observed, having spread in limited number of range gates. This makes the detection altogether difficult with the normal technique. To reduce the possibility of false alarm of identifying the interference frequency, the threshold value is chosen to be not less than 3 dB above the mean noise floor. The 3-dB threshold is chosen after careful evaluation of the noise power spectral density for a number of observations. It is observed that the standard deviation of noise power over its mean is around 3 dB. Thus, choosing the threshold value of 3 dB makes sure that no noise fluctuations were identified as interference frequency. The running average will reduce the noise spikes in the power spectrum, which will bring down the variance of the noise spectrum, thereby reducing the threshold value. Once the signal is identified above the threshold value, the same is carried out on the entire range gate on the same frequency bin in the frame. If, any time, more than 15% of the total range gates in the frame are continuously reported above the threshold value, then that frequency is considered as interference frequency. This procedure is repeated until all the frequency bins have been checked in the entire range gate in the Doppler power spectrum. The observation window set for the lower atmosphere is about 22.5 km, which corresponds to 150 range gates for a resolution of 150 m. The atmospheric signal seldom falls on the same Doppler bin from range bin to range bin, and the interference signal does as well. Therefore, the continuity checking on the range gates for the same frequency helps one to isolate the atmospheric signal that is wrongly selected as interference frequency.
Algorithm flowchart of an autonomous interference filter.
spectral points below the level nearly satisfies the criterion. The details of the steps involved in the algorithm for mean-noiselevel determination are given as follows: Step 1) Reorder the spectrum {Pi , i = 1, . . . , N } in ascending order. Let this sequence be written as {Ai , i = 1, . . . , N } and Ai < Aj for i < j Step 2) Compute the following: PM = QM =
N Ai (i) i=1 N A2 i
i=1
and if QM > 0, RM =
i
(1) C. Interference Filter Design −
2 PM
2 PM , (QM ∗ M )
(2) for M = 1, . . . , N (3)
A notch filter is a filter that contains one or more deep notches or, ideally, perfect nulls in its frequency-response characteristics. Notch filters are useful in many applications where specific frequency components must be eliminated. To create a null in the frequency response of a filter at a frequency ω0 by
4
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
Fig. 2. Notch-filter frequency–magnitude responses.
introducing a pair of complex conjugate zeros on the unit circle at an angle ω0 [12] Z1,2 = e±jω0 .
(5)
Thus, the system function for a finite-impulse response (FIR) notch filter is simply
IE E Pr E oo f
H(Z) = b0 (1 − ejω0 z −1 )(1 − e−jω0 z −1 ) = b0 1 − 2 cos ω0 z −1 + z −2 .
(6)
In general, in an FIR notch filter, the notch has a relatively large bandwidth, which means that the other frequency components around the desired null are severely attenuated. To reduce the bandwidth of the null, a more sophisticated and longer FIR filter could be designed. It is also possible to improve the frequency-response characteristics by introducing poles in the system function. The purpose of the poles is to introduce a resonance in the vicinity of the null and to therefore reduce the bandwidth of the notch. If we place a pair of complex poles at P1,2 = re±jω0
1 − 2 cos ω0 Z −1 + Z −2 . 1 − 2r cos ω0 Z −1 + r2 Z −2
(8)
Fig. 2 shows the frequency–magnitude responses of the notch filter with poles located at 0.95 and 0.80. For the demonstration of filter responses, a notch frequency of 2 Hz and a signal bandwidth of 8 Hz are chosen. A notch depth that is better than −50 dB and notch bandwidths of 0.29 Hz at −3 dB (around 3.6% of filter bandwidth) and 0.023 Hz at −20 dB (around 0.28% of filter bandwidth) are observed in the frequency–magnitude-response curve for r at 0.95. Currently, the value of r at 0.95 is chosen for the interference filter design. If further reduction in notch bandwidth is desired, that can be achieved by designing a higher order filter. The difference equation of the notch filter is given by Y (n) = X(n) − 2 cos ω0 X(n − 1) + X(n − 2) + 2r cos ω0 Y (n − 1) − r2 Y (n − 2).
for a number of frequency points identified for removal. Once the filtering process is over, power-spectrum computation is carried out for further processing of the signal. The first three lower order moments are estimated from the power spectrum, which represents the total power, mean Doppler, and Doppler width. III. R ESULTS AND D ISCUSSION
(7)
the system function for the resulting filter is H(Z) = b0
Fig. 3. Range-versus-Doppler-power-spectrum plot of atmospheric signals contaminated with interference signals.
(9)
The aforesaid equation is being used for filtering out the interference bands in the signal. The filtering process continues
The algorithm developed has been tested on a number observations related to atmospheric signals received from the Gadanki MST radar under various conditions. Two cases are presented here, i.e., one for the observation of the lower atmosphere used for wind profiling up to an altitude of 22 km, and the other one for the observation related to the ionosphere. Fig. 3 shows the Doppler power spectra of the signals for the range of 3.6–21.7 km. The plots are normalized to individual range gates. In general, the radar-backscattered power reduces at a rate of (range)−2 for the atmospheric targets. Therefore, as range increases, the returned power reduces. The power received from the radar is recorded in arbitrary value and not in absolute. To have a better presentation, plots are range normalized to show all the range gates with strongest signal as reference. To show the entire range gates (around 150) having signals, individual range-normalized plots are shown. There are three interference bands observed in the spectra at around 1, 5.1, and −3.8 Hz. It is observed that interferences also have varied characteristics. The interference at 1 Hz is present throughout the range gates. In the lower range gates, it is not visible because atmospheric signals are stronger than interference signals. The interference at 5.1 Hz starts at around 14 km and continues until the end of the range gate, whereas
5
IE E Pr E oo f
ANANDAN AND JAGANNATHAM: APPROACH APPLIED TO WIND PROFILERS
Fig. 5. Range-versus-Doppler-power-spectrum plot of the lower atmospheric wind profile signal contaminated with interference signals. The mean Doppler estimated in each range bin is plotted along with it.
Fig. 4. Range-versus-Doppler power-spectrum plot of the lower atmospheric wind profile signal after removing the interference signals.
the interference at 3.8 Hz starts around 8 km and ends at 14.5 km. It is observed that, in general, the externally generated interference present throughout the Doppler-range gate spectra and intermittent interference bands are usually generated from high-power system arcing or malfunctioning. Fig. 4 shows the spectra after applying the algorithm. The algorithm could detect all the interference bands and remove them effectively. The wind profile is clearly visible up to 21 km. Once the interference bands are removed, the mean noise level is estimated for each range gate using an objective method based on Gaussian statistics [6]. The mean noise level for each range gate is subtracted from the corresponding power spectrum. The three lower order moments are estimated using the expressions given by Woodman [16], which correspond to the total power, mean Doppler, and Doppler width. The strongest peak available in the spectrum is identified for the estimation of moments. The estimation of moments before and after filtering is shown in Figs. 5 and 6. The mean Doppler estimated is plotted against the Doppler power spectra for all range gates. From Fig. 5, it is observed that detection fails after 12 km because interference signals dominate over atmospheric backscattered signals. Whereas in the case of filtered signal spectrum, the signal could be detected correctly, and the wind profile is estimated up to 21 km, as shown in Fig. 6. In the second case, the observation related to the ionospheric signal is presented. The wind profiler works at VHF, which is also capable of detecting ionospheric signals when the beam is pointed transverse to the Earth’s magnetic field. This observation is taken when the radar beam is pointed 14◦ off zenith. Fig. 7 shows the 3-D plot of ionospheric observation for a height range of 99–129 km. There are two interference
Fig. 6. Range-versus-Doppler-power-spectrum plot of the lower atmospheric wind profile signal after the removal of interference signals. The mean Doppler estimated in each range bin is plotted along with it.
bands observed in the power spectra around 60 and −31 Hz. Ionospheric signals are generally sporadic in nature and very weak during this observation period. Ionospheric signals are present at two range gates at 111 and 112.2 km. The interference signal at −31 Hz passes through atmospheric signals. The interference removal algorithm is applied to this data frame. Fig. 8 shows the power-spectrum plot after the interference removal. Here, the interference bands are also effectively removed from the power spectrum and show very clearly signals
6
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
from the atmospheric radar. There are two cases presented, i.e., one for lower atmospheric signals corresponding to the wind profile up to an altitude of 21 km, and the other for ionospheric signals where the interference bands go through atmospheric signals. In both cases, multiple interference bands were removed effectively. It has also been demonstrated that moments estimation is improved by removing the interference bands and detecting the signal at higher heights having less SNR. The algorithm has been applied to a number of data frames that are contaminated with interference signals, and in all cases, the algorithm could detect and remove the interference bands. This interference detection and filtering algorithm helps in reducing the number of frames to be rejected during quality checking, thereby obtaining high temporal resolution in atmospheric wind profiling. R EFERENCES
IE E Pr E oo f
[1] V. K. Anandan, P. Balamuralidhar, P. B. Rao, A. R. Jain, and C. J. Pan, “An adaptive moments estimation technique applied to MST radar echoes,” J. Atmos. Ocean. Technol., vol. 22, no. 4, pp. 396–408, Apr. 2005. [2] V. K. Anandan, G. Ramachandra Reddy, and P. B. Rao, “Spectral analysis of atmospheric radar signal using higher order spectral estimation technique,” IEEE Trans. Geosci. Remote Sens., vol. GRS-39, no. 9, pp. 1890– 1895, Sep. 2001. [3] E. E. Clothiaux, R. S. Rene, D. W. Thomson, T. P. Ackerman, and S. R. Williams, “A first-guess feature-based algorithm for estimating wind speed in clear-air Doppler radar spectra,” J. Atmos. Ocean. Technol., vol. 11, no. 4, pp. 888–908, Aug. 1994. [4] M. A. Fischler and R. C. Boltes, “Random sample consensus: A paradigm for model fitting with application to image analysis and automated cartography,” Commun. ACM, vol. 24, no. 6, pp. 381–395, Jun. 1981. [5] R. K. Goodrich, C. S. Morse, L. B. Cornman, and S. A. Cohn, “A horizontal wind and wind confidence algorithm for Doppler wind profilers,” J. Atmos. Ocean. Technol., vol. 19, no. 3, pp. 257–273, Mar. 2002. [6] P. H. Hildebrand and R. S. Sekhon, “Objective determination of the noise level in Doppler spectra,” J. Appl. Meteorol., vol. 13, no. 7, pp. 808–811, Oct. 1974. [7] J. R. Jordan and R. J. Latatis, “A potential source of bias in horizontal winds estimated using a 915-MHz acoustically enhanced profiler,” J. Atmos. Ocean. Technol., vol. 14, no. 3, pp. 543–546, Jun. 1997. [8] P. T. May and R. G. Strauch, “An examination of wind profiler signal processing algorithms,” J. Atmos. Ocean. Technol., vol. 6, no. 4, pp. 731– 735, Aug. 1989. [9] P. T. May and R. G. Strauch, “Reducing the effect of ground clutter on wind profiler velocity measurements,” J. Atmos. Ocean. Technol., vol. 15, no. 2, pp. 579–586, Apr. 1998. [10] D. A. Merritt, “A statistical averaging method for wind profiler Doppler spectra,” J. Atmos. Ocean. Technol., vol. 12, no. 5, pp. 985–995, Oct. 1995. [11] C. S. Morse, R. K. Goodrich, and L. B. Cornman, “The NIMA method for improved moment estimation from Doppler spectra,” J. Atmos. Ocean. Technol., vol. 19, no. 3, pp. 274–295, Mar. 2002. [12] J. G. Proakis and D. G. Manolakis, Digital Signal Processing, 3rd ed. Englewood Cliffs, NJ: Prentice-Hall, 1996. [13] P. B. Rao, A. R. Jain, P. Kishore, P. Balamuralidhar, S. H. Damle, and G. Viswanathan, “Indian MST radar. I: System description and sample vector wind measurements in ST mode,” Radio Sci., vol. 30, no. 4, pp. 1125–1138, 1995. [14] R. G. Strauch, B. L. Weber, A. S. Frisch, C. G. Little, D. A. Merritt, K. P. Moran, and D. C. Welsh, “The precision and relative accuracy of profiler wind measurements,” J. Atmos. Ocean. Technol., vol. 4, no. 4, pp. 563–571, Dec. 1987. [15] T. L. Wilfong, S. A. Smith, and R. L. Creasey, “High temporal resolution velocity estimates from a wind profiler,” J. Spacecr. Rockets, vol. 30, no. 3, pp. 348–354, May 1993. [16] R. F. Woodman, “Spectral moments estimation in MST radars,” Radio Sci., vol. 20, pp. 1185–1195, 1985. [17] C. W. Ulbrich, “Natural variations in the analytical form of the raindrop size distribution,” J. Clim. Appl. Meteorol., vol. 22, no. 10, pp. 1764–1775, Oct. 1983.
Fig. 7. Range-versus-Doppler-power-spectrum plot of ionospheric signals contaminated with interference signals.
Fig. 8. Range-versus-Doppler-power-spectrum plot of ionospheric signals after removing the interference signals.
at 111 and 112.2 km. The algorithm is tested on a number of data frames corrupted with interference and shows consistency in performance by removing the interference signals, leaving behind the atmospheric signals. IV. C ONCLUSION An autonomous interference detection and filtering algorithm has been developed and successfully tested on signals received
ANANDAN AND JAGANNATHAM: APPROACH APPLIED TO WIND PROFILERS
D. B. V. Jagannatham received the B.Tech. degree in electronics and communication engineering from Andhara University, Waltair, India, in 1990 and the M.Tech. degree from Jawaharlal Nehru Technological University, Hyderabad, India, in 1999, where he is currently working toward the Ph.D. degree in digital signal processing, mainly on nonlinear signal processing applications. He is currently a Professor with the Department of Electronics and Communication Engineering, C.R. Engineering College, Tirupati, India. His areas of interest are radar signal processing, algorithm development, and digital system design.
IE E Pr E oo f
V. K. Anandan received the M.Sc. and M.Tech. degrees from Cochin University of Science and Technology, Kochi, India, and the Ph.D. degree in electronics and communication engineering from Sri Venkateswara University, Tirupati, India. He has done his postdoctoral research at National Central University, Chung-Li, Taiwan. He joined the Indian Space Research Organisation (ISRO), Department of Space, Government of India, in 1991. He was responsible for developing real-time data acquisition and signal processing system for mesosphere–stratosphere–troposphere radar and closely associated with other atmospheric radar developments. He has developed the signal and data processing software package for atmospheric radars and sodars. He is currently with the Radar Development Area, ISRO Telemetry, Tracking and Command Network, Bangalore. His research contribution is in new signal processing algorithm development related to atmospheric radars, precipitation studies, and atmospheric boundary layer. He is currently involved in the development of wind profiler radar and cloud radar systems. He has published more than 40 research papers. His area of interest includes radar system developments, signal processing algorithm developments, and studies related to remote sensing of the atmosphere using different observational systems. Dr. Anandan is a Fellow of The Institution of Electronics and Telecommunication Engineers and an associate fellow of the Andhra Pradesh Academy of Sciences.
7
