Vous êtes sur la page 1sur 14

2678

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 37, NO. 6, NOVEMBER 1999

Adaptive Parametric Algorithms for Processing Coherent Doppler-Lidar Signal


Jean-Luc Zarader, Alain Dabas, Pierre H. Flamant, Bruno Gas, and Olivier Adam
Abstract In this paper, we study the autoregressive and moving average (ARMA) lter for lidar signal processing. After a short presentation of the atmospheric laser Doppler instrument project (ALADIN), we introduce the objective of this paper, which is to extract the Doppler frequency and to retrieve the spectral width of a noised lidar signal. A general presentation of ARMA lters and parametric adaptive algorithms (PAAs) is provided. Then we present results about the choice of the model, the Doppler frequency estimate, and the spectral width estimate. Finally, we study the possible estimate of SNR, which is biased by the rst estimates (Doppler frequency and spectral width). Index Terms Adaptive signal processing, laser radar.

I. INTRODUCTION OHERENT Doppler lidars have been developed for more than 15 years now [25] and have proved very powerful and useful for remotely measuring winds in the atmosphere. The measurement consists of transmitting a laser pulse into the atmosphere. Along its propagation path, the laser radiation is scattered by the aerosol particles drifting with the wind. Part of the energy is scattered back toward the instrument, where it is captured by a telescope and analyzed. For a Doppler system, the analysis aims at a range-resolved estimate of the signal frequency. According to Dopplers effect, it is equal to the frequency of the transmitted laser pulse, which is monitored in real time, plus a shift proportional to the wind velocity along the line-of-sight. At the present time, various systems, either ground-based or airborne [8], [9], [23], [36], [41], have been developed and used operationally in meteorology to document particular features of atmospheric dynamics [1][6], [11], [13], [21], [36]. The global (i.e., at the scale of the Earth) and continuous survey of atmospheric wind eld, though very promising for weather prediction, still is not available. It requires spaceborne instruments that still are not in existence because they are at the limit of what the current technology can provide. Nevertheless, development programs such as the Space Readiness Coherent Lidar Experiment (SPARCLE) in the United States (see http://wwwghcc.msfc.nasa.gov/sparcle/)
Manuscript received August 7, 1997; revised November, 30, 1998. This work was supported by a grant from the European Space Agency (ESA). J.-L. Zarader and B. Gas are with the Laboratoire des Instruments et Syst` emes, Universit e Paris VI, Paris, France. A. Dabas is with the Centre National de Recherche M et eorologique, M et eoFrance, Toulouse, France. P. Flamant is with the Laboratoire de M et eorologie Dynamique, Palaiseau, France. O. Adam is with the Laboratoire dEtude et de Recherche en Informatique, Signaux et Syst` emes, Universit e Paris XII, Paris, France. Publisher Item Identier S 0196-2892(99)06267-1.

and the atmospheric laser Doppler instrument (ALADIN) in Europe [42]. The processing of signals in the digital domain is a difcult point in the design of an instrument. The main reason is that the signals contain a great deal of detection noise (the SNR is often below 0 dB) due to the limited powers delivered by existing lasers, the low backscatter coefcient of the atmosphere, and, for spaceborne applications in particular, the long distance from the targets. Additionally, the useful part of the signal, the atmospheric echo, is subject to random phase and amplitude uctuations known as speckles. These occasionally lead to local fadeouts. Hoping that better processing would alleviate the demanding specications for high-energy lasers and large-area telescopes, many studies have been devoted to lidar signal processing in the past. First, the precise nature of the signals has been claried [37], resulting in the provision of accurate and realistic signal simulation models [40]. Second, the best frequency estimators for Doppler lidar have been considered: Pulse-Pair [31], Poly-Pulse-Pair [29], parametric or nonparametric adaptive lters [27], [33], [45], and the so-called maximum likelihood estimator [16]. Their performances for the range-resolved retrieval of the signal frequency have been determined, both on simulated [16], [18], [19] and real [15], [18] signals. Though most of the studies cited above have focused on the retrieval of the mean frequency (that is, the rst order moment of the signal-power spectrum), the signal power and the width of its spectrum (zeroth and second order moments of the power spectrum, respectively) are also valuable parameters. The former measures the light intensity backscattered by the atmosphere, and thus provides information on the structure of the atmosphere (for instance, the presence and altitude of subvisible clouds). Also, it gives the SNR, which is a good piece of information for the assessment of the quality of Doppler measurements. Regarding the spectrum width, it contains information on the level of turbulence and windshear, since both are responsible for a broadening of the returned spectrum. The present paper investigates the possible application of adaptive autoregressive and moving average (ARMA) lters to coherent lidar for the joint estimate of the signal power, mean frequency, and spectrum width. To our knowledge, such an investigation has never been conducted before. Indeed, parametric spectral analysis already has been conducted on Doppler radars and lidars, which both deliver similar signals, but Keeler and Lee [26] and Mahapatra and Zrnic [29] dwell on the maximum entropy estimator, which is the adoption of

01962892/99$10.00 1999 IEEE

ZARADER et al.: ADAPTIVE PARAMETRIC ALGORITHMS

2679

an autoregressive (AR) lter [10]. Filters with MA parts also have been used, but the objective is then the determination of the signal spectrum width. At last, the application of notch lters [33] to lidar signals is investigated by Zarader et al. [45]. However, the purpose is then the estimation of the signal mean frequency only. This paper will be divided into six sections. In the second section, the signal model retained for the whole study is presented with a discussion of its ability to account for actual Doppler-lidar systems. Also, the principles of adaptive ARMA spectral analysis are reviewed with a particular emphasis on the numerical methods used for adapting the lter to the signal. The following sections (III, IV, and V) present the performances reached by adaptive ARMA for the estimate of the Doppler frequency, spectrum width, and signal power. The last section concludes the study. II. BACKGROUND A. Doppler-Lidar Signals Basically, two types of coherent Doppler lidars exist, depending on the laser technology. Though both deliver pulses with different characteristics (see below), regarding in particular the wavelength (around 2 m for solid-state and 10 m for CO lasers), the nature of the signals generated by the instrument is similar in both cases. After demodulation, it can be considered as a complex Gaussian process polluted by a statistically independent noise. The noise is generally assumed to be white, which in reality is assured by a proper matching of the anti-aliasing analog band-pass lter to the sampling frequency of the analog to digital (A/D) converter. The atmospheric return is never perfectly stationary, since it contains at least a range dependant power decrease. However, at long distances, and provided the optical and dynamic properties of the atmosphere can be considered spatially homogeneous along the line-of-sight (which is often the case for clear-air measurements), the stationarity is nearly met, at least to the second order. Therefore, stationarity generally is considered a basic assumption of lidar signal processing studies. Then, the statistical properties of the signal samples are entirely . characterized by the autocorrelation function At a given time, the lidar signal results from the addition of a great number of laser pulses reected by the backscattering where is the distance targets with a time delay (equal to instrument target) and a frequency Doppler shift [40]. Since the distance from the target varies over many half-wavelengths, it follows that the reected pulses add randomly at detection level. They sometimes add constructively, sometimes destructively. It results in random phase and amplitude uctuations known as speckles [14]. At a given time, the probability den, the amplitude follows sity of the phase is uniform over a Rayleigh distribution, and the power follows a negative exponential distribution [12], [22]. Then, speckle uctuations evolve in time due to the renewal of the targets inside the probed volume (caused by the propagation of the laser pulse) and the relative movements of the targets remaining inside. is basically The result is that the autocorrelation function

equal to the autocorrelation of the laser pulse plus a decorrelating impact from all the wind velocity uctuations inside the illuminated volume [37]. Correspondingly, the signal spectral power density (the Fourier transform of the autocorrelation) is basically equal to the power spectrum of the laser pulse (that is, the square magnitude of its Fourier transform) plus some broadening by intrapulse velocity uctuations. In many studies, the autocorrelation function is given a Gaussian shape (1) where is the power of the atmospheric echo, and correlation time. Since the spectral power density Fourier transform of sets the is the

(2) is the signal spectral width. In reality, the parameter Gaussian autocorrelations are met with solid-state lidars, because delivered pulses have a Gaussian power prole and no signicant chirp [15], [17]. The autocorrelation function above is a good model in this case. With CO lidars however, the pulse power prole combines a short spike and long a tail. Furthermore, it contains a signicant frequency chirp [44]. Then, the autocorrelation cannot be Gaussian. However, should the pulse energy within the spike be limited as well as the frequency chirp, the autocorrelation should not be too far from Gaussian. Therefore, (1) and (2) will be systematically considered in the following. As a consequence, the synthetic signals we shall use in the next section to test the performances of the proposed estimators will be generated with Zrnics simulator [46]. is related priIn real applications, the spectrum width marily to the laser pulse [17]. Solid-state lasers deliver pulses varying from s to with no chirp and a duration s [19]. The spectrum width thus ranges from kHz to MHz. For 10 m systems, the duration is longer (1.5 s to 3 s), but the large frequency chirps nevertheless broaden the spectra from 250 to 500 kHz. For both types of system, the duration of the processing windows may vary s to s or more, depending on the required from m and m, respectively). So the range resolution ( , which is proportional to nondimensional product the number of independent realizations of speckle uctuations to within one processing window, might vary from or more (that is, about two decades). B. Parametric Estimation Parametric models generated in signal processing are used for process identication, prediction of signals, or even (for the present subject matter) spectral analysis [28]. The interesting aspect of these models rests in their high-frequency resolution. We can distinguish two types of lters. First of all, AR (or all pole lters) adapted to the analysis of signals with one or several peaks in their spectrum. Therefore, they are interesting for the processing of lidar signals. Yet, the performances of such lters depend on the number of samples and on

2680

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 37, NO. 6, NOVEMBER 1999

noise level. To reduce the effect of disturbances, one tends to increase the lter order (that is, the number of coefcients or parameters). This technique amounts to using an ARMA model. can be dened by the following An ARMA lter relationship: (1) represents white noise and represents the lidar where signal. Spectral density can be determined from the transfer function of the lter. It can be expressed as follows by a transform [34]: (2)

C. Algorithms The cost function of these algorithms corresponds to the average power of the a posteriori prediction error. At time , it can be expressed as (5) with

(6) is the a posteriori prediction, and are the where input (or observation) vector and the vector of the estimated parameters, respectively (7)

, we obtain the spectral density of Assuming that the signal obtained by ltering input with variance . We can now write (3) Contrary to notch lters (ANFs), ARMA lters with free coefcients enable one to estimate Doppler frequency and spectral width, and to obtain some information concerning the SNR. The rst problem met upon carrying an ARMA modeling, is the choice of the lter order. There exist different tests, called performance-complexity criteria, that allow one to obtain a model with a reduced number of parameters, satisfying signal processing. For instance, we can mention the Aka ke criterion (Final Prediction Error) [28], dened by (4) and are, respectively, the number of samples where is the a and the number of parameters of the model. posteriori prediction error. Hereafter, we use this criterion in order to determine the coefcients of the lter. One uses parametric adaptive algorithms (PAAs) [24], [30], [43]. The coefcients are calculated recursively. Several methods have been developed so that parameters can converge toward their optimum values. Nonetheless, the principle remains the same: minimizing a cost function (or criterion) that depends on the error made between the lidar signal value and the value predicted by the lter. This error is called the prediction error. In part II-C, we shall detail algorithms based on the exact least square criterion, such as: 1) recursive least squares (RLS); 2) recursive maximum likelihood (RML); 3) output error (OE). We shall then compare the results obtained in the estimate of the Doppler frequency and the spectral width of lidar signals in Sections III and IV. Finally, in Section V, we shall present the results obtained for the evaluation of the SNR.

and (8) The criterion can be written as follows for a block of data: (9) This global criterion, calculated from initial time, ensures a uniform progression of coefcients. The following methods give an estimate of the coefcients of the model, minimizing this criterion. 1) Algorithm of Recursive Least Squares (RLS): The RLS algorithm is used to update the lter parameters (10) is the gain matrix. It varies with the power of the input signal. Its initialization is important and results in a more or less fast convergence of coefcients toward their optimum value. The problem with this algorithm is that minimization of the criterion is made on all past times (from initial time 1 to time ). This therefore assumes that the set of coefcients does not change with time, or from a physical point of view, that the signal is stationary (constant Doppler frequency during the shot). To take into account nonstationary aspects of the lidar signals, several modications can be introduced. They concern either the calculation of the errorcost function, or the content of the input vector. The criterion dened previously takes all data into account with equal weight. To study the nonstationary process, it is necessary to introduce a weight favoring recent evolutions of the signal to compare it to past behavior. The new errorcost function shows a forgetting factor (11) is within zero and one. The forgetting factor also where can be found in the evolution of the gain matrix .

ZARADER et al.: ADAPTIVE PARAMETRIC ALGORITHMS

2681

Several proposals were made concerning the choice of the forgetting factor. : We obtain the recursive least squares algorithm a) again. with within zero and one: In practice, b) it is chosen within the interval [0.8; 1]. Previous data are forgotten at an exponentially growing speed. The weight is at its maximum for the last error calculated. tunes the time As a matter of fact, the parameter for instance, the estimate resolution. When is performed on the whole signal from the beginning , the to step . On the contrary, when is less than weight given to component , and is thus negligible. Through , the time window corresponding to an estimate is xed. When the weighting function is a negative exponential, there exists no straightforward denition for the time resolution. For instance, if we set the time resolution of , when the weight an estimate as the time delay becomes lower than and , the time resolution is equal to ten time . When and , it is equal samples to 20 and 50 time samples, respectively. with and within zero and c) one, chosen close to one: The forgetting factor varies. This forgetting is important during the rst iterations (with convergence), and then it decreases through processing. The forgetting factor asymptotically tends to one and thus prevents the adaptation gain from decreasing too quickly. term that enables one One also can introduce a . In fact, to maintain a constant trace on gain matrix if the trace of this matrix decreases too quickly, updating does not occur. One avoids decreasing the gain so as to be able to follow possible variations of parameters with time. For the study of nonstationary processes through their parameters, this technique provides for permanent adjustment of the coefcients of the models. It is possible to combine these possibilities using a priori, known signal characteristics. In our study, a forgetting factor enables minimization of the convergence time by weighting the rst estimates, which were not realistic. However, since real signals are not very stationary in terms of frequency, it is necessary to use either a variable forgetting factor, or a factor providing for constant trace of the gain matrix. We started analyzing the lidar signals using a variation forgetting factor so as to minimize convergence time, and then maintained a constant gain matrix trace to provide for follow-up of possible evolutions of the characteristics of the signal. For all applications, the forgetting factor remains smaller than 0.98, which corresponds [(for an exponential (see previous section)], to a range resolution function at of 180 m for a sampling frequency MHz. 2) Algorithm of Recursive Maximum Likelihood (RML): In estimation theory, a likelihood variable allows us to know whether the associated magnitude is close to the result expected. This algorithm repeats this idea in ltering the obser-

vation vector with an estimate of the model associated to lter inputs. This model is a representation of the most likely biases introduced through the various inputs. Thus, the ltering stage is intended to remove the component generated by noises that are present in inputs. Therefore, vector is ltered by model , where is the estimator of the coefcients of part MA at time . is Consequently, the a priori prediction of (12) with

(13) and are the output and the a posteriori prediction errors, . respectively, ltered by The parameters are as follows:

(14) This ltering tends to accelerate decorrelation between the vector of observations and the prediction error. Results obtained are likely to be more interesting than those drawn from the RLS algorithm. Yet, in practice, they are contingent upon the . precision of the estimate of Furthermore, upon initializing, as a correct estimate of the lter coefcients is not available, processing has to start with the RLS algorithm to then return to this algorithm. Lidar signals are slightly or highly noised, depending on atmospheric conditions. The use of this method can be useful for signals with a low SNR. Extraction of the signal conveying the signicant information will be facilitated by the preltering phase. 3) Algorithm of Output Error (OE): This algorithm differs from the RLS algorithm by the choice of the vector of observations. Signal samples are replaced by their a posteriori estimates

(15) This amounts to considering that signal samples are biased and that, consequently, the information contained in prediction is more accurate. Thus, estimate of the signal depends indirectly on the disturbance through the adaptation algorithm. In the case of a good modelization, the a posteriori error tends asymptotically toward a white noise, guaranteeing an unbiased estimate of parameters. In this method, the criteria take into account the information contained in previous prediction errors. Performances obtained using the exact least squares method can be improved for noisy signals with a large spectral width. There again, it is appropriate to begin the processing using the RLS algorithm and then to continue with the OE algorithm. To conclude, it can be noticed that a better matching between the lter model inferring the estimator and the signal leads to a better accuracy. In particular, the CramerRao lower

2682

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 37, NO. 6, NOVEMBER 1999

bound (CRLB) for frequency estimation is asymptotically reached when the signal ts the lter model [28]. So it would be necessary to search for the more appropriate ARMA model to improve the performance and get closer to the CRLB (see III-A). It turns out that for a limited range gate duration, as considered in the present study, the CRLB cannot be reached but only approached within a certain factor (about three to ten) depending on the model to be used. III. MEAN FREQUENCY ESTIMATE These algorithms were developed and applied to lidar signals. Our aim is to estimate the wind velocity on a single-shoot basis (without accumulation on successive realizations) in order to decrease the speckle noise effect [7]. To compare the results obtained by different algorithms, we used the simulation model proposed by Zrnic [46]. These signals, for a constant velocity, allow one to evaluate the bias and the standard deviation of the estimator. Then, in order to replace this study in an experimental case, we have used a signal model with a time-varying frequency, which corresponds to varying velocity. As previously indicated (see II-C.1), for these two kinds of signals, the forgetting factor remains smaller than 0.98. Below, we are presenting the bias and the standard deviation of the estimator calculated on 200 lidar signals of 2,000 real-valued samples. As often is the case in adaptive signal processing (and for reasons of complexity), we have processed real signals. One of the advantages is that it is not necessary to create lines and as for complex signals. points) can be written as For the th shot, the bias (on follows: (16) and being real and estimated Doppler frequencies with for the th shot. shots, the bias average is For (17) Variance on the estimate (for the th shot) is (18) For shots, the variance average is (19)

density of bad estimates, or outliers. Three independent parameters are therefore necessary to characterize the whole distribution. However, a comprehensive characterization of the probability density functions of the measurements is far beyond the scope of the present article. It would require determining the type of distribution for the return power and spectral width, and to our knowledge, these have never been studied yet. This would call for a long work that would have a very limited scope, since the distributions are likely to depend on the actual laser pulse used for the measurements. A. Choice of the Model Order As previously indicated (see end of II-C.3), performances of the ARMA models depend on the order of parts MA and AR. The number of inputs retained corresponds to a representation that is in conformity with lidar signals. Thus, we tend to increase the order when signals are very noisy. Table I shows results obtained (using the RLS algorithm and after convergence) for the bias and the standard deviation of the Doppler frequency estimate in relation to the characteristics of lidar signals (noise, spectral width). The rst line shows the bias and the second line gives the standard deviation for the model considered. To convert these results to velocity, it is necessary to know the frequency/velocity relation. For example, for a 10 m laser, 1 m/s is equivalent MHz, to 200 kHz. Then, for a sampling frequency corresponds to 400 kHz ( a spectral width of m/s). Table I can be read: If MHz, the spectral , the is , and the width dB, the standard deviation is , which corresponds to kHz, inferior to 1 m/s (Note: stands throughout the text). for The best results obtained using the FPE criterion are shown in heavy print. One notes that for a large spectral width model gives the best performance. For (0.05), the smaller widths, results can be improved slightly by increasing or a model. the order. For instance, one can use an One can explain this result by the fact that, for a large spectral width, the spectrum displays less extreme values. One therefore can be satised with a model with a smaller number of poles and zeros. To conclude, the model we have chosen is the lter, which offers the advantage of a small complexity (reduced calculation time). Furthermore, one will note that it is possible to extract the Doppler frequency from the calculation of the poles of the denominator (order 2) of the transfer function. Thus, one avoids the long spectrum calculation. B. Convergence Time

Results obtained for the estimate of the Doppler frequency depend on the SNR and on the spectral width. The quality of the various measurements (wind velocity, return power, and spectral width) will be characterized by biases and standard deviations. Those characteristics are not enough to fully describe their whole probability distributions. For wind velocity for instance, such distribution is the addition of a Gaussian density of good estimates plus a uniform

Convergence time corresponds to the time necessary to reach an estimate error in the Doppler frequency less than 5%. Thus, it is possible to evaluate the number of iterations necessary to obtain a rst reliable estimate of the Doppler frequency. Fig. 1 shows the inuence of initialization of the gain matrix used in parametric algorithms. Each gure corresponds to the overlay of ten estimates of the Doppler frequency on

ZARADER et al.: ADAPTIVE PARAMETRIC ALGORITHMS

2683

TABLE I BIAS AND STANDARD DEVIATION OF DOPPLER FREQUENCY ESTIMATE FOR DIFFERENT ARMA MODELS. THE DOPPLER FREQUENCY IS fD = 0:2Fs . THE FIRST COLUMN REPRESENTS THE ARMA MODEL, LINE b CORRESPONDS TO BIAS, AND LINE sd CORRESPONDS TO STANDARD DEVIATION. BIAS AND STANDARD DEVIATION ARE CALCULATED FOR EVERY SNR (5 dB, 5 dB, AND 15 dB) AND EVERY SPECTRAL WIDTH (0:01Fs , 0:03Fs , AND 0:05Fs ). NOTE: e x STANDS FOR 100x , AS IN THE BODY OF THE TEXT

the rst 500 points of ten different signals. These signals , are simulated using Zrnics model (spectral width dB, and real Doppler frequency (solid line). SNR A large initial value of the gain matrix enables one to increase the power of the rst errors, and the consequence is to reduce convergence time (to the detriment of the estimate variance). In fact, the adaptation gain also intervenes in the updating of coefcients. When this gain is large, variations also will be large, from one estimate to the next. Therefore, one therefore must nd a convergence-variance middle term. Since this convergence time corresponds to a blind zone, within 100 and 1000). we chose a large initial gain ( It must be noted that one surely can improve these results in different ways. For instance, we can initialize lter coefcients at a value close to the Doppler frequency, with the latter being in its turn estimated by another algorithm. Also, we can make two estimates: the rst one being made in the direction of increasing time and the second in the direction of decreasing time. C. Comparison of Algorithms In Fig. 2(a), we show the bias in the frequency estimate obtained after convergence of the RLS algorithm. On Fig. 2(b), we show the standard deviation. Calculations were made model for three spectral widths: , with an , and . The normalized Doppler frequency is . For each SNR, bias and standard deviation are calculated on 200 signals of 2000 samples. One notes that, regardless of the spectral width, bias and standard deviation plots are practically identical. The choice of the model prevails over the choice of the algorithm. In fact,

contrary to ANF, the model has a larger number of degrees of freedom and therefore, it is able to make a better approximation of the lidar signal spectrum. On these plots, we note that the bias is below for signals for which the SNR is above dB. For slightly dB, noised signals, estimates are biased slightly. Below to at dB. the bias increases from Within the same range, the relative error of the Doppler frequency increases from 5% to 16%. Standard deviations are of the same magnitude for all three spectral widths. They are small for signals with an SNR above dB. They exceed in the opposite case, reaching at dB. Fig. 3 shows the results obtained with algorithm OE. The conditions of experimentation are the same with the RLS algorithm. There again, we note the similarity of the bias with the standard deviation. Bias remains relatively constant and below up to dB. For an SNR exceeding dB, results obtained by the RLS or OE algorithms are comparable. for an SNR of dB. Conversely, the bias reaches It is twice as big as the one obtained with the RLS algorithm. On the other hand, the standard deviation of the OE esticompared to the RLS estimate mate is reduced for SNRs below dB. Above , dB standard deviations are practically identical. Fig. 4 shows the bias and the standard deviation obtained with the RML algorithm. The conditions of experimentation are the same with the RLS and OE algorithms. dB. In this case, plots differ slightly for SNRs below This is due to the fact that for the RML algorithm, data are

2684

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 37, NO. 6, NOVEMBER 1999

(a)

(a)

(b)

(b) Fig. 2. Bias and standard deviation of the Doppler frequency estimate as a function of the SNR for fD = 0:2Fs . The signals are processed with an RLS algorithm and a model ARMA(2; 2). The three curves correspond to three different normalized spectral widths:  = 0:01, r = 0:03, and 3 = 0:05. (a) Normalized bias. (b) Normalized standard deviation. (c) Fig. 1. Doppler frequency estimate as a function of time (number of samples) using an ARMA(2; 2) model for different initial gain value. The Doppler frequency is fD = 0:2Fs . (a) Initial gain value = 0:25. (b) Initial gain value = 100 (c) Initial gain value = 1000.

In light of its simplicity, we retained the RLS algorithm to estimate the Doppler frequency.

D. Tracking of the Doppler Frequency Contrary to estimators of the FFT or PPP type, adaptive algorithms function on sliding windows of variable size. Therefore, it is possible to track, with a fair amount of accuracy, the evolution of the Doppler frequency with time. To assess this tracking capacity, we simulated a noisy signal with a variable frequency [Fig. 5(a)]. Frequency variations (from to ) are shown in Fig. 5(b). These variations are drawn from a real-wind distribution with more or less signicant slopes. This study enabled us to x values, allowing for good tracking such as the forgetting factor (0.9 to 0.98) and the threshold of the gain matrix trace (0.01). In Fig. 6 (plot ), we show the estimate of the Doppler model. frequency with the RLS algorithm for an dB. The SNR equals

ltered with an estimate of the MA model, and this estimate is all the more biased due to the fact that the noise is important. For the bias, the RML algorithm enables one to obtain results that are comparable to those given by the RLS algorithm, with the exception of signals in which spectral width is equal to 0.05 and performances decay as soon as the SNR is lower dB. than The standard deviation is of the same magnitude with all for SNRs three algorithms. It is almost constant dB. It reaches at dB. exceeding Bias is minimum with the RLS algorithm, and standard deviations are of the same magnitude with all three algorithms (except for highly noised signals, for which the OE algorithm enables us to obtain a smaller standard deviation).

ZARADER et al.: ADAPTIVE PARAMETRIC ALGORITHMS

2685

(a)

(a)

(b) Fig. 3. Bias and standard deviation of the Doppler frequency estimate as a function of the SNR for fD = 0:2Fs . The signals are processed with an OE algorithm and a model ARMA(2; 2). The three curves correspond to three different normalized spectral widths:  = 0:01, r = 0:03, and 3 = 0:05. (a) Normalized bias. (b) Normalized standard deviation.

(b) Fig. 4. Bias and standard deviation of the Doppler frequency estimate as a function of the SNR for fD = 0:2Fs . The signals are processed with an RML algorithm and a model ARMA(2; 2). The three curves correspond to three different normalized spectral widths:  = 0:01, r = 0:03, and 3 = 0:05. (a) Normalized bias. (b) Normalized standard deviation.

When one compares the estimate for the Doppler frequency (plot ), one notes that for slow variations of the frequency, the tracking is good. It fails only when the frequency gets close to the Shannon frequency (samples 2100 to 2500), probably because of spectral aliasing. This problem can be solved easily by limiting the Doppler frequency to a value between and . IV. SPECTRUM WIDTH We studied the estimate of the spectral width for two reasons. 1) It is a characteristic of weather disturbances. 2) Even if it is a rough value, a value of the spectral width can be used by another estimator of the spectral width (for instance by Levin [27]). A. Estimate To make an estimate of the spectral width, one proceeds in three steps. First, one calculates the spectrum from the

coefcients of the lter. Next, one ts a Gaussian shape and computes its standard deviation. Finally, the standard deviation is compared to the real spectrum width. Fig. 7 sums up the various steps. Fig. 7(a) shows two plots, representing the spectrum calculated from the coefcients of the ARMA model (solid line) and the tted Gaussian shape (dashed line). In Fig. 7(b), we draw the tted Gaussian curve and the initial spectrum (solid line). These gures were obtained in processing a real-valued lidar signal with a spectral and an SNR of ratio dB. width As in the Doppler frequency study, we dened a bias and a variance. For the th shot, the deviation between the estimated width and the real width can be written as follows: (20) and the average for shots is (21)

2686

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 37, NO. 6, NOVEMBER 1999

TABLE II BIAS AND STANDARD DEVIATION OF SPECTRAL WIDTH ESTIMATE (I.E., NORMALIZED-TO-SAMPLING FREQUENCY) FOR DIFFERENT ARMA MODELS. THE DOPPLER FREQUENCY IS fD = 0:2Fs . THE FIRST COLUMN REPRESENTS THE ARMA MODEL, LINE b CORRESPONDS TO BIAS, AND LINE sd CORRESPONDS TO STANDARD DEVIATION. BIAS AND STANDARD DEVIATION ARE CALCULATED FOR EVERY SNR (5 dB, 5 dB, AND 15 dB) AND EVERY SPECTRAL WIDTH (0:01Fs , 0:03Fs , AND 0:05Fs ). NOTE: e x STANDS FOR 100x , AS IN THE BODY OF THE TEXT

The estimate variance for the th shot is (22) For -shots, the variance average is (23) B. Results As we have done for the estimate of the Doppler frequency, in Table II, we show the bias and standard deviation obtained on the estimate of the spectral width. We varied the order of the model. There again, we note that the results obtained are rather close when the AR orders range from two to eight. However, we note that we obtain slightly better performances when using , , or . For the models with a higher order estimate of the spectral width, we retained the model. Fig. 8 shows the bias and the standard deviation obtained with the RLS algorithm for the spectral width estimate and for signals with spectral widths of 0.01, 0.03, and 0.05. The . For each SNR, the bias and Doppler frequency is of the standard deviation are calculated on 200 Lidar signals of 2000 samples. The bias for the estimate of the spectral width is below 0.063. As in the case of the Doppler frequency, bias is noticeably smaller for signals for which the SNR exceeds dB. Also, it is below 0.02 for signals with three spectral widths. For a highly noised signal, the spectrum, inferred from

the parameters of the model, is attened, which causes an overestimation of the width (0.07). At dB, the relative error is of 40% (130% and 500%, respectively) for a spectral width 0.05 (0.03 and 0.01, respectively). Standard deviation is below 0.005 for signals with a small dB. For spectral width and for which the SNR exceeds the noise levels considered, the standard deviation remains below 0.02 regardless of the width. For the other algorithms, the results are nearly the same. For example, the bias and the standard deviation with the dB are and RML algorithm and an , respectively. The estimate of the spectral width is dB. It is strongly biased at fair for SNRs greater than lower SNRs. V. SIGNAL POWER In this section, we will present the performances of parais the metric estimators for the evaluation of the SNR. If noised discrete lidar signal (24) and are the useful signal and the noise, where respectively, at time . Evaluating the SNR enables us to do the following. 1) Use it as a validation criterion for the estimates of spectral widths. The smaller the SNR, the lesser condence we can have in the estimate of the Doppler frequency. 2) Optimize the use of the Doppler frequency estimator (Levin for example).

ZARADER et al.: ADAPTIVE PARAMETRIC ALGORITHMS

2687

(a)

(a)

(b) Fig. 5. Frequency unstationary signal as a function of time (number of samples). (a) Noisy signal. (b) True Doppler frequency.

(b) Fig. 7. Estimate of signal spectral width (for fD = 0:2Fs ). (a) Solid line: Estimated Spectrum. Dashed Line: Gaussian approximation. (b) Solid line: True Gaussian Spectrum. Dashed Line: Gaussian approximation.

Fig. 6. Tracking of the signal Doppler frequency as a function of time (number of samples). (a) True Doppler frequency. (b) Estimated Doppler frequency.

A. Estimate The SNR can be obtained from the estimated spectrum of lidar signals. Fig. 9 shows the spectral estimate obtained with an model on real-valued Zrnic signals with SNRs dB. The corresponding spectral width ranging from 5 dB to , and the normalized Doppler frequency is . is In Fig. 9(a) and (b), we note that the average estimated dB) to 0.5 ( noise increases from 0.2 (

dB) when SNR decreases. We also note that the width of the Gaussian curve is overestimated when SNR is small (see Section IV-B). To estimate the SNR (Fig. 10), we considered normalized powers associated with 1) the noised spectrum (solid line) calculated from the ; parameters of the model, referred to as 2) the Gaussian curve (dashed line) calculated from the (this Gaussian previous spectrum, referred to as curve can be inferred from the fact that the estimate of the spectral width depends on the precision obtained with ARMA models); 3) the average level (constant value) of the noise, referred (this level is calculated from the noise level to as differentiated from the Gaussian curve and is all the more signicant when the SNR is small).

2688

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 37, NO. 6, NOVEMBER 1999

TABLE III Spt POWERS ESTIMATES WITH AN ARMA(2; 2) MODEL AS A FUNCTION OF SNR. THE SIGNALS ARE PROCESSED WITH THE RLS ALGORITHM. Am IS THE NOISE POWER LEVEL, Ag ^ IS THE POWER UNDER THE GAUSSIAN CURVE, AND ASpt IS THE POWER OF THE ESTIMATED SPECTRUM FOR THREE-SIGNAL WIDTH 0:01Fs , 0:03Fs , AND 0:05Fs
Am

Ag

^,

AND

TABLE IV P OWERS ESTIMATES WITH AN ARMA(8; 4) MODEL AS A FUNCTION OF SNR. THE SIGNALS Spt ARE PROCESSED WITH THE RLS ALGORITHM. Am IS THE NOISE POWER LEVEL, Ag ^ IS THE POWER UNDER THE GAUSSIAN CURVE, AND ASpt IS THE POWER OF THE ESTIMATED SPECTRUM FOR THREE SIGNAL WIDTH 0:01Fs , 0:03Fs , AND 0:05Fs
Am

Ag

^,

AND

At rst, the spectrum of the noisy signal is calculated. Then, we extract the gaussian shape as indicated in IV-A. is estimated. Then, the constant noise level B. Results We have carried out this study for the two models previously and . retained, i.e., and Table III shows the results obtained for magnitudes with an model, with signals of 0.01, 0.03, and 0.05 spectral widths, and with an SNR exceeding dB. The parametric algorithm used is the RLS algorithm. gives an information on the normalized SNR. In theory, this ratio should decrease when SNR decreases. gives identical information, but no estimate of the spectral width is required. Estimate is biased when SNR is small. Reading this Table III, we can make three comments. increases when noise increases, regardless of the 1) real spectral width.

2) Ratio decreases when the SNR varies from 5 dB. Then, contrary to the expected result, dB to this ratio increases. This behavior can be explained by dB, the bias on the estimate the fact that below is therefore of the width becomes important. Power . overestimated, and the same applies to ratio decreases when SNR decreases. 3) In the end, dB, values of this ratio are very However, below close to one. This shows that the signal is drowned by the noise. Therefore, it seems difcult to infer the corresponding SNR from it. Table IV shows that the results obtained with the model are better than those obtained with model. This can be explained by a more the appropriate modelization of the signal. increases when noise increases. Here again, we note that ratio depends on the Conversely, evolution of the decreases to dB ( dB and spectral width.

ZARADER et al.: ADAPTIVE PARAMETRIC ALGORITHMS

2689

(a)

(a)

(b) Fig. 8. Bias and standard deviation of spectral width estimate (fD = 0:2Fs ) with an RLS algorithm and a model ARMA(8; 4) as a function of SNR. The three curves correspond to three different spectral widths: = 0:01, = 0:03, and = 0:05. (a) Normalized Bias. (b) Normalized Standard Deviation.

(b)

dB, respectively) for a 0.01 width (0.03 and 0.05, respectively). This behavior is linked to the estimate of the spectral width. In fact, we showed previously (Fig. 8) that the relative bias, over the spectral width, is less important than the size of the ratio remains real width. Finally, the behavior of the globally unchanged compared to the behavior of the ratio noted model. with the To conclude, an accurate estimate of the SNR using an ARMA model remains difcult and poorly reliable at low dB). This is due primarily to the fact that SNRs (below this evaluation is done after the estimate of the spectral width.

Fig. 9. Spectral estimation with a model ARMA(8; 4) for two different SNRs, using the RLS algorithm. The Doppler frequency is fD = 0:2Fs . (a) SNR = 5 dB. (b) SNR = 5 dB.

VI. CONCLUSION Parametric methods, and especially the ARMA model, are well suited for the processing of lidar signals. Coefcients of the AR part enable one to locate the typical peak in the spectrum, and the MA part helps improve the parameter estimates.

These algorithms, developed for adaptive signal processing, provide for the updating of the parameters of lters upon every presentation of a new sample of the lidar signal. Moreover, primarily in the case of frequency nonstationary signals, it is possible to optimize convergence using various methods (forgetting factor and constant trace). The use of parametric ARMA models allows us to 1) know the spectrum estimated upon each sample of the lidar echo; 2) estimate the Doppler frequency (small bias and standard deviation even for noised signals); 3) estimate the spectral width (evaluation of atmospheric disturbances, interesting information to optimize other estimators, LEVIN for example); 4) estimate the SNR (needs the spectral width estimate).

2690

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 37, NO. 6, NOVEMBER 1999

Fig. 10. Power estimate. Solid line: estimated spectrum; dashed line: Gaussian approximation; constant line: estimated noise-power level.

However, as we have shown, the last estimate, SNR, is strongly biased. At last, the prospectives of the present work are both a validation of performances on real signals and a comparison of these performances with other possible predictive estimators based on neural networks.

REFERENCES
[1] R. M. Banta, L. D. Olivier, E. T. Holloway, R. A. Kropi, B. W. Bartram, R. E. Cupp, and M. J. Post, Smoke-column observations from two forest-res using Doppler lidar and Doppler radar, J. Appl. Meteorol., vol. 31, pp. 13281349, Nov. 1992. [2] R. M. Banta, L. D. Olivier, and D. H. Levinson, Evolution of the Monterey sea-breeze layer as observed by pulsed Doppler lidar, J. Atmos. Ocean. Technol., vol. 50, pp. 39593982, Dec. 1993. [3] R. M. Banta, L. D. Olivier, W. D. Neff, D. H. Levinson, and D. Rufeux, Inuence of canyon-induced ows on ow dispersion over adjacent plains, Theor. Appl. Climatol., vol. 52, no. 1, pp. 2742, 1995. [4] R. M. Banta, Sea breezes shallow and deep on the California coast, Month. Weather Rev., vol. 123, pp. 36143622, 1995. [5] R. M. Banta, L. D. Olivier, P. H. Gudiksen, and R. Lange, Implications of small-scale ow features to modeling dispersion over complex terrain, J. Appl. Meteorol., vol. 35, pp. 330342, Dec. 1996. [6] R. M. Banta, P. B. Shepson, J. W. Bottenheim, K. G. Anlauf, H. A. Wiebe, A. Gallant, T. Biesenthal, L. D. Olivier, C.-J. Zhu, I. G. McKendry, and D. G. Steyn, Nocturnal cleansing ows in a tributary valley, Atmos. Environ., vol. 31, pp. 21472162, May 1997. [7] M. F. Barth, R. B. Chadwick, and D. W. Van De Kamp, Data processing algorithms used by NOAAs wind proler demonstration network, Ann. Geophys., vol. 12, pp. 518528, July 1994. [8] J. W. Bilbro and W. W. Vaughan, Wind eld measurement in the non precipitous regions surrounding severe storms by an airborne pulsed Doppler lidar system, Bull. Amer. Meteorol. Soc., vol. 59, pp. 10951100, Sept. 1980. [9] J. Bilbro, G. Ficht, D. Fitzjarrald, M. Krause, and R. Lee, Airborne Doppler lidar wind eld measurements, Bull. Amer. Meteorol. Soc., vol. 65, pp. 348359, Apr. 1984. [10] J. P. Burg, Maximum entropy spectral analysis, 37th Annu. Meeting, Soc. Explor. Geophys., Oklahoma City, OK, 1967. [11] T. L. Clark, W. D. Hall, and R. M. Banta, Two- and three-dimensional simulations of the 9 January 1989 severe Boulder windstorm: Comparison with observations, J. Atmos. Sci., vol. 51, pp. 23172343, Aug. 1994. [12] J. H. Churnside and H. T. Yura, Speckle statistics of the atmospherically backscatter laser light, Appl. Opt., vol. 22, pp. 25592526, Sept. 1983.

[13] W. L. Eberhard, R. E. Cupp, and K. R. Healy, Doppler lidar measurements of proles of turbulence and momentum ux, J. Atmos. Ocean. Technol., vol. 6, pp. 809819, Oct. 1989. [14] P. H. Flamant, R. T. Menzies, and M. J. Kavaya, Evidence for speckle effects on pulsed CO2 lidar signal returns from remote targets, Appl. Opt., vol. 23, pp. 14121417, May 1984. [15] R. G. Frehlich, S. M. Hannon, and S. W Henderson, Performance of a 2-m coherent Doppler lidar for wind measurements, J. Atmos. Ocean. Technol., vol. 11, pp. 15171528, Dec. 1994. [16] R. G. Frehlich and M. J. Yadlowsky, Performance of mean frequency estimators for Doppler radar and lidar, J. Atmos. Ocean. Technol., vol. 11, pp. 12171230, Oct. 1994. [17] R. Frehlich, Comparison of 2- and 10-m coherent Doppler lidar performance, J. Atmos. Ocean. Technol., vol. 12, pp. 415420, Apr. 1995. [18] , Simulation of coherent Doppler lidar performances in the weak signal regime, J. Atmos. Ocean. Technol., vol. 13, pp. 646658, June 1996. [19] , Effects of wind turbulence on coherent Doppler lidar performance, J. Atmos. Ocean. Technol., vol. 14, pp. 5475, Feb. 1997. [20] R. G. Frehlich, S. M. Hannon, and S. W Henderson, Coherent Doppler lidar measurements of winds in the weak signal regime, Appl. Opt., vol. 36, pp. 34913499, May 1997. [21] T. Gal-Chen, M. Xu, and W. L. Eberhard, Estimations of atmospheric boundary layer uxes and other turbulence parameters from Doppler lidar data, J. Geophys. Res., vol. 97, no. 18, pp. 409418, 423, 1992. [22] J. W. Goodman, Statistical Optics. New York: Wiley, 1985. [23] F. F. Hall, R. M. Huffaker, R. M. Hardesty, M. E. Jackson, T. R. Lawrence, M. J. Post, R. A. Riichter, and B. F. Weber, Wind measurements accuracy of the NOAA pulsed infrared Doppler lidar, Appl. Opt., vol. 23, pp. 25032506, Nov. 1984. [24] S. Haykin, Adaptative Filter Theory. Englewood Cliffs, NJ: PrenticeHall, 1991. [25] R. M. Huffaker and R. M. Hardesty, Remote sensing of atmospheric wind velocities using solid-states and CO2 coherent lidar systems, Proc. IEEE, vol. 84, pp. 181204, Aug. 1996. [26] R. J. Keeler and R. W. Lee, Complex covariance/maximum entropy Doppler estimates for pulsed CO2 lidar, Int. Conf. Acoustics, Speech, and Signal Processing, Instr. Electron. Eng., Tulsa, OK, 1978. [27] M. J. Levin, Power spectrum parameter estimation, IEEE Trans. Inform. Theory, vol. IT-11, pp. 100107, Jan. 1965. [28] L. Ljung and T. Soderstrom, Theory and Practice of Recursive Identication. Cambridge, MA: MIT Press, 1983. [29] P. R. Mahapatra and D. S. Zrnic, Practical algorithms for mean velocity estimation in pulse Doppler weather radars using a small number of samples, IEEE Trans. Geosci. Remote Sensing, vol. GE-21, pp. 491501, Oct. 1983. [30] F. Michaut, M ethodes adaptatives pour le signal Collection trait e des nouvelles technologies, s erie traitement du signal. Paris, France: Hermes, 1992. [31] K. S. Miller and M. M. Rochwarger, A covariance approach to spectral moment estimation, IEEE Trans. Inform. Theory, vol. IT-18, pp. 588596, Sept. 1972. [32] A. Nehora , A minimal parameter adaptive notch lter with constrained poles and zeros, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-33, pp. 983996, Aug. 1985. [33] A. Nehora and D. Starer, Adaptive pole estimation, IEEE Trans. Acoust., Speech, Signal Processing, vol. 38, pp. 825838, May 1990. [34] L. R. Rabiner, J. H. McClellan, and T. W. Parks, FIR digital lter design techniques using weighted Chebychev approximation, in Proc. IEEE ICASSP, 1975, pp. 595610. [35] J. Rothermel, C. Kessinger, and D. L. Davis, Dual-Doppler lidar measurements of winds in the JAWS experiment, J. Atmos. Ocean. Technol., vol. 2, pp. 138147, June 1985. [36] J. Rothermel, D. R. Cutten, R. M. Hardesty, R. T. Menzies, J. N. Howell, S. C. Johnson, D. M. Tratt, L. D. Olivier, and R. M. Bante, The multicenter airborne coherent atmospheric wind sensor, MACAWS, Bull. Amer. Meteorol. Soc., Apr. 1998. [37] B. J. Rye, Spectral correlation of atmospheric lidar returns with range dependent backscatter, J. Opt. Soc. Amer., vol. A7, pp. 21992207, Dec. 1990. [38] B. J. Rye and R. M. Hardesty, Discrete spectral peak estimation in incoherent backscatter heterodyne lidar. I. Spectral accumulation and the Cramer-Rao lower bound, IEEE Trans. Geosci. Remote Sensing, vol. 31, pp. 1627, Jan. 1993. , Discrete spectral peak estimation in incoherent backscatter [39] heterodyne lidar. II. Correlogram accumulation, IEEE Trans. Geosci. Remote Sensing, vol. 31, pp. 2835, Jan. 1993.

ZARADER et al.: ADAPTIVE PARAMETRIC ALGORITHMS

2691

[40] P. Salamitou, A. Dabas, and P. H. Flamant, Simulation in the time domain for heterodyne coherent laser radar, Appl. Opt., vol. 34, pp. 499506, Jan. 1995. [41] R. L. Schweisow and M. P. Spowart, The NCAR airborne infrared lidar system: Status and applications, J. Atmos. Ocean. Technol., vol. 13, pp. 415, Feb. 1996. [42] A. Stoffelen, P. Flamant, D. Carson, and W. Wergen, Report for assessment, the nine candidates earth explorer missions, the atmospheric dynamics mission, ESA, vol. SP-1196, no. 4, 1996. [43] B. Widrow and S. D. Stearns, Adaptative Digital Processing. Englewood Cliffs, NJ: Prentice-Hall, 1985. [44] D. V. Willets and M. R. Harris, An investigation into the origin of frequency sweeping in a hybrid TEA CO2 laser, J. Phys., vol. D15, no. 2, pp. 5167, 1982. [45] J. L. Zarader, G. Ancellet, A. Dabas, N. K. MSirdi, and P. H. Flamant, Performance of an adaptive notch lter for spectral analysis of coherent lidar signals, J. Atmos. Ocean. Technol., vol. 13, pp. 1628, Feb. 1996. [46] D. S. Zrnic, Simulation of weatherlike Doppler spectra and signals, J. Atmos. Ocean. Technol., vol. 14, pp. 619620, June 1975.

Pierre H. Flamant received the Doctorat-es-science degree in physics from the Universit e de Paris VI, France, in 1979. He is currently Directeur de Recherche at the Centre National de la Recherche Scientique (CNRS). He also is leading a team of ten scientists at the Laboratoire de M et eorologie Dynamique, Palaiseau, France. His main research interests are in meteorology and climate processes, atmospheric ow, planetary boundary layer dynamics, radiative budget linked to semitransparent clouds and aerosols, and lidar physics and relevant disciplines like signal processing and data inversion techniques. Dr. Flamant is the chair of the International Coordination Group on Laser Atmospheric Studies (ICLAS) and a member of the International Radiation Commission. He also is a member of several working groups on space-based lidars at the European Space Agency, Paris, France, and Co-Investigator of the PICASSO-CENA program, recently approved by NASA and the French Space Agency CNES, to develop a space-based backscatter lidar to be launched in 2003. He has more than 60 published papers in peer-reviewed journals.

Jean-Luc Zarader received the Ph.D. degree in applied physics in 1989 from the Universit e Pierre et Marie Curie, Paris VI, France. He is currently Ma tre de Conf erences at the Laboratoire des Instruments et Syst emes on the Perception et Commande team, Universit e Paris VI. He is currently participating in the European airborne Doppler lidar Wind Infrared Doppler lidar (WIND) project and the European Space Agency project, ALADIN. He also is working on noisy speech coding from dynamics and predictive neural networks. His research interests are in adaptive signal processing and neural networks processing. The applications are in lidar signal processing and in speech recognition and coding. Dr. Zarader is a member of FRANcophone de lIng enieurie de la Langue (FRANCIL).

Bruno Gas received the Ph.D. degree in electronics from the Universit e dOrsay, Paris XI, France, in 1994. He is currently Maitre de Conferences at the Universite de Paris VI, France, and is working on the noisy speech coding problem and lidar spectrum analysis from neural networks. His main research interest is the study of neural predictive models for temporal data coding. Dr. Gas is member of the international group FRANcophone de lIng enieurie de la Langue (FRANCIL).

Alain Dabas received the engineering degree from Ecole Polytechnique, Palaiseau, France, in 1988, the engineering degree from the Ecole Nationale de M et eorologie Toulouse, France, in 1990, and the Ph.D. degree in atmospheric physics from the Universit e Pierre et Marie Curie, Paris VI, France, in 1993. He is currently working at the Centre National de Recherches M et eorologique, Toulouse, France, the research center for the French weather service M et eo-France. He also is participating in the ground-based portable Doppler lidar project (LVT), and the European airborne Doppler lidar project (WIND). His current research interests are in lidars (light detection and ranging) for the remote sensing of atmospheric wind. He has specialized in lidar signal processing (spectral analysis of detected RF signals) and data analysis (retrieval of atmospheric wind eld from lidar measurements).

Olivier Adam received the engineering degree from the Ecole Sup erieure dIng enieur en Electrotechnique et Electronique de Paris, France, in 1991. He received the Ph.D. in signal processing from the University Pierre et Marie Curie, Paris VI, France, in 1995. He is currently a Research Teacher with the University of Cr eteil, France, in the Laboratoire dEtudes et de Recherches en Instrumentation Signaux et Syst` emes (LERISS), Creteil, France. His research interests include biomedical domain and expert systems. He also is working with the detection of human auditory pathology and the classication with neural network approach.

Vous aimerez peut-être aussi