• Angevine, W. M., Avery S. K. , and Kok G. L. , 1993: Virtual heat flux measurements from a boundary-layer profiler-RASS compared to aircraft measurements. J. Appl. Meteor., 32 , 19011907.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Carter, D. A., Gage K. S. , Ecklund W. L. , Angevine W. M. , Johnston P. E. , Riddle A. C. , Wilson J. , and Williams C. R. , 1995: Developments in UHF lower tropospheric wind profiling at NOAA's Aeronomy Laboratory. Radio Sci., 30 , 977.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Cohn, S. A., Goodrich R. K. , Morse C. S. , Karplus E. , Mueller S. , Cornman L. B. , and Weekley R. A. , 2001: Radial velocity and wind measurement with NIMA: Comparisons with human estimation and aircraft measurements. J. Appl. Meteor., 40 , 704719.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Cornman, L. B., Goodrich R. K. , Morse C. S. , and Ecklund W. L. , 1998: A fuzzy logic method for improved moment estimation from Doppler spectra. J. Atmos. Oceanic Technol., 15 , 12871305.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Freund, J. E., 1992: Mathematical Statistics. 5th ed. Prentice Hall, 658 pp.

  • Goodrich, R. K., Morse C. S. , Cornman L. B. , and Cohn S. A. , 2002: A horizontal wind and wind confidence algorithm for Doppler wind profilers. J. Atmos. Oceanic Technol., 19 , 257273.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hilderbrand, P. H., and Sekhon R. S. , 1974: Objective determination of the noise level in Doppler spectra. J. Appl. Meteor., 13 , 808811.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Jordan, J. R., Lataitis R. J. , and Carter D. A. , 1997: Removing ground and intermittent clutter contamination from wind profiler signals using wavelet transforms. J. Atmos. Oceanic Technol., 14 , 12801297.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • May, P. T., and Strauch R. G. , 1998: Reducing the effect of ground clutter on wind profiler velocity measurements. J. Atmos. Oceanic Technol., 15 , 579586.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • May, P. T., Sato T. , Yamamoto M. , Kato S. , Tsuda T. , and Fukao S. , . 1989: Errors in the determination of wind speed by Doppler radar. J. Atmos. Oceanic Technol., 6 , 235242.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Morel, C., and Passi R. , 1986: Use of adaptive polynomials for real-time data editing. J. Atmos. Oceanic Technol., 3 , 494498.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Press, W. H., Flannery B. P. , Tenkolsky S. A. , and Vetterling W. T. , 1988: Numerical Recipes in C. Cambridge University Press, 735 pp.

  • Schumann, R. S., Taylor G. E. , Merceret F. J. , and Wilfong T. L. , 1999: Performance characteristics of the Kennedy Space Center 50-MHz Doppler radar wind profiler using the median filter/first guess data reduction algorithm. J. Atmos. Oceanic Technol., 16 , 532549.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • van de Kamp, D. W., 1988: Profiler Training Manual #1: Principles of Wind Profiler Operation. National Weather Service, 49 pp.

  • Weber, B. L., Wuertz D. B. , Welsh D. C. , and McPeek R. , 1993: Quality controls for profiler measurements of winds and RASS temperatures. J. Atmos. Oceanic Technol., 10 , 452464.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Yager, R. R., Ovchinnikov S. , Tong R. M. , and Nguyen H. T. , 1987: Fuzzy Sets and Applications, Selected Papers by L. A. Zadeh. John Wiley and Sons, 685 pp.

    • Search Google Scholar
    • Export Citation
  • Zimmerman, H. J., 1996: Fuzzy Set Theory and its Applications. 3d ed. Kluwer Academic Publishers, 435 pp.

  • Zrnić, D. S., 1975: Simulation of weather like Doppler spectra and signals. J. Appl. Meteor., 14 , 619620.

  • Zrnić, D. S., . 1980: Spectral statistics for complex colored discrete-time sequence. IEEE Trans. Acoust. Speech Signal Process., 28 , 596599.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • View in gallery

    Contour plots of example Doppler spectra from the Juneau Lemon Creek 915-MHz profiler including atmospheric, clutter, point target, and RFI signals. (a) Unfiltered dB scaled spectra overlaid with POP moments, (b) 1D median-filtered dB scaled spectra overlaid with NIMA moments with moments confidence shown in the right sidebar, first moment confidence in red, second moment confidence in green. Note that confidence is lower in low SNR regions

  • View in gallery

    This chart shows the primary external input and output data and important intermediate data. Each circle, aj, indicates processing that is described in section 4

  • View in gallery

    Illustrations of the results of NIMA processing on spectra from Fig. 1. (a) 2D median-filtered dB scaled spectral data. (b) Range-normalized dB scaled spectral data. (c) Clutter features detected by NIMA. (d) Atmospheric features detected by NIMA and their classifications. Feature numbers correspond to columns in Table 1

  • View in gallery

    Example of velocity-folded spectra from the Sha Lo Wan profiler. (a) The velocity folded 2D median-filtered spectra with a Nyquist velocity of 8.55 m s−1. (b) The unfolded 2D median-filtered spectra with an effective Nyquist velocity of 17.1 m s−1. Range gates below 464 m were not detected as being folded, so the spectral intensities at the additional Doppler velocities are set to the noise level

  • View in gallery

    Example of Doppler peak membership threshold adjustment processing for spectra from the Lemon Creek profiler under precipitation conditions. (a) The spectra and NIMA moments. Note the strong signal and velocity folding at the melting layer around the 522-m range gate. (b) The NIMA atmospheric features at the initial Doppler peak membership threshold of 0.15. Each feature is shown in a distinct color. (c) The NIMA atmospheric features at the raised threshold of 0.36

  • View in gallery

    Examples of membership functions for RFI feature detection. Membership characteristic values that lie outside the ranges shown receive the membership value associated with the closest value within the range shown. Note that the membership value is positive for an RFI feature; zero or negative for non-RFI features

  • View in gallery

    Portion of example spectrum from Fig. 1 at 440 m. Dashed line is the 2D median-filtered spectrum, dotted line is the noise floor, shaded areas are the spectral cutoff regions, solid line shows the Gaussian fit. The Gaussian-estimated midpoint x0 and the resulting first and second moments estimated from the Gaussian are indicated

  • View in gallery

    The idealized (without random noise) spectral profile used in the simulation study is based on the profile shown in Fig. 1

  • View in gallery

    Simulation results for atmospheric signal without contaminants. Average errors in the first and second moments (second moments as spectral half-width) over 1000 trials are shown for NIMA and SPA algorithms. Also shown are the theoretical signal-to-noise ratio and the average SNR as calculated by SPA

  • View in gallery

    The lowest four gates of the idealized simulated spectra including atmospheric signal and ground clutter. The asterisks indicate the true first moment (center point) and the true spectral width (outer points)

  • View in gallery

    Simulation results for atmospheric signal contaminated with clutter. Average errors in the first moments over 1000 trials are shown for NIMA and SPA algorithms. Also shown are the theoretical SNR for the atmospheric signal and the average SNR calculated by SPA based on its picked peak

  • View in gallery

    Simulated example of the effect of NIMA's RFI processing in the case of weak atmospheric signal. (a) With RFI processing enabled, NIMA selects the weaker atmospheric signal in preference to the stronger RFI signal. (b) With RFI processing disabled, NIMA selects the RFI signal

  • View in gallery

    Simulation results for atmospheric signal contaminated with clutter and RFI. Average errors in the first moments over 1000 trials are shown for NIMA and SPA algorithms. Also shown are the NIMA average first moment confidence values

  • View in gallery

    Simulation results for atmospheric signal contaminated with clutter, RFI, and point targets. Average errors in the first moments over 1000 trials are shown for NIMA and SPA algorithms. Also shown are the NIMA average first moment confidence values and the SPA average SNR calculated at each height

  • View in gallery

    Plots of first moment errors as a function of the fraction of 6000 sample spectra removed based on NIMA first moment confidence. The top x axis shows the confidence thresholds used to discard the fraction of the data indicated on the lower x axis. (a) A plot of average absolute error. (b) A plot of rms error

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 145 142 13
PDF Downloads 94 93 5

The NIMA Method for Improved Moment Estimation from Doppler Spectra

View More View Less
  • 1 National Center for Atmospheric Research, Boulder, Colorado
  • | 2 National Center for Atmospheric Research, and Department of Mathematics, University of Colorado, Boulder, Colorado
  • | 3 National Center for Atmospheric Research, Boulder, Colorado
© Get Permissions
Full access

Abstract

The NCAR Improved Moments Algorithm (NIMA) for estimating moments from wind measurement devices that measure Doppler spectra as a function of range is described in some detail. Although NIMA's main application has been for real-time processing of wind profiler data, it has also been successfully applied to Doppler lidar and weather radar data. Profiler spectra are often contaminated by a variety of sources including aircraft, birds, velocities exceeding the Nyquist velocity, radio frequency interference, and ground clutter. The NIMA method uses mathematical analysis, fuzzy logic synthesis, and global image processing algorithms to mimic human experts' ability to identify atmospheric signals in the presence of such contaminants. NIMA is configurable and its processing can be tuned to optimize performance for a given profiler site. Once configured, NIMA is a fully automated algorithm that runs in real time to produce Doppler moments and a confidence assessment of those moments. These confidence values are useful in the generation and assessment of wind and turbulence estimates and are important when these quantities are used in critical situations such as airport operations. A simulation study is used to compare NIMA performance with that of a simple peak picking algorithm in the presence of ground clutter, RFI, and point targets. Some performance results for the NIMA confidence algorithm are also given.

Corresponding author address: Corinne S. Morse, National Center for Atmospheric Research, P.O. Box 3000, Boulder, CO 80307-3000. Email: morse@ucar.edu

Abstract

The NCAR Improved Moments Algorithm (NIMA) for estimating moments from wind measurement devices that measure Doppler spectra as a function of range is described in some detail. Although NIMA's main application has been for real-time processing of wind profiler data, it has also been successfully applied to Doppler lidar and weather radar data. Profiler spectra are often contaminated by a variety of sources including aircraft, birds, velocities exceeding the Nyquist velocity, radio frequency interference, and ground clutter. The NIMA method uses mathematical analysis, fuzzy logic synthesis, and global image processing algorithms to mimic human experts' ability to identify atmospheric signals in the presence of such contaminants. NIMA is configurable and its processing can be tuned to optimize performance for a given profiler site. Once configured, NIMA is a fully automated algorithm that runs in real time to produce Doppler moments and a confidence assessment of those moments. These confidence values are useful in the generation and assessment of wind and turbulence estimates and are important when these quantities are used in critical situations such as airport operations. A simulation study is used to compare NIMA performance with that of a simple peak picking algorithm in the presence of ground clutter, RFI, and point targets. Some performance results for the NIMA confidence algorithm are also given.

Corresponding author address: Corinne S. Morse, National Center for Atmospheric Research, P.O. Box 3000, Boulder, CO 80307-3000. Email: morse@ucar.edu

1. Introduction

To fully utilize Doppler measurement devices for the real-time measurement of wind, wind shear, and turbulence, data from these devices must be quality controlled. When rapidly updated measurements are required, such as for an airport wind hazard warning system, the traditional method of long-term moment averaging or “consensus” methods (Weber et al. 1993) cannot be used. Analysis of the Doppler spectra by human experts provides reliable quality control (Schumann et al. 1999) but becomes impractical when continuous real-time processing is required. The National Center for Atmospheric Research (NCAR) Improved Moments Algorithm (NIMA) spectral processing algorithm produces a set of quality-controlled moments from each beam of spectral data at a skill level approaching that of a human expert.

The NIMA algorithm, composed of an integrated set of fuzzy logic modules, was first described by Cornman et al. (1998). This paper gave a generalized overview of the fuzzy logic and image-processing techniques used, particularly as applied to the problem of distinguishing ground clutter and atmospheric signal in the spectra from the Sha Lo Wan, China, profiler. NIMA was shown to have several advantages over traditional peak picking methods. Although NIMA was developed for the processing of profiler data and has been applied to boundary layer, tropospheric, and 50-MHz devices, it should be noted that the method is also applicable to any device that measures spectra as a function of range,1 and it has been successfully applied to both Doppler lidar and weather radar spectral data.

Since the original publication, the NIMA algorithm has been further developed and a relatively mature version has been employed operationally for over three years as part of a terrain-induced turbulence research study at the Juneau International Airport (Alaska). This paper describes the overall NIMA processing with an emphasis on the new components of the algorithm. These include a new fuzzy logic membership combination method, an algorithmic component to handle time-varying radio frequency interference (RFI) contamination, and perhaps most significantly, estimation of a quality control index (confidence) in the resulting moments. These confidence estimates can be used in subsequent processing, that is, wind and turbulence calculations based on the moments, to produce more accurate and reliable wind and turbulence estimates. For an operational warning system, confidence estimates can reduce false alarms, that is, a hazard warning may be suppressed in the case of calculated moments that indicate a wind-related hazard but have a low confidence. The details of the wind calculations, the confidence estimates for these values, and the resulting performance are the subject of related papers (Goodrich et al. 2002; Cohn et al. 2001).

An example of the spectral moments calculated by NIMA compared with those calculated by the Profiler On-line Program (POP; Carter et al. 1995) are shown in Fig. 1 for a beam of data from the Lemon Creek 915-MHz profiler in Juneau, Alaska. The plots show the contoured spectral intensity, or spectral surface, as a function of Doppler velocity (horizontal axis) and range (vertical axis). The first moment at each range gate is denoted by an asterisk and the spectral width (twice the square root of the second moment) is indicated by the length of the line segment centered at the first moment. Figure 1a shows the 30-s average spectral data and the moments calculated by POP. This data has several contaminants in addition to the atmospheric signal. The strong narrow signal near zero velocity and below 600 m is characteristic of ground clutter. The weak signal near 8 m s−1 Doppler velocity is RFI with characteristic constant Doppler shift and near constant intensity over all range gates. The region of localized spectral intensity around 440-m range and −9 m s−1 Doppler velocity is characteristic of a transient contaminant such as a bird. POP is based on locating the peak intensity of the signal at each range gate and produces good moments for clean spectra. However, when contaminants have spectral peak intensities greater than that of the atmospheric signal, erroneous moments can be generated as seen in Fig. 1a. In traditional applications, such erroneous moments are not problematic as they are eliminated through consensus averaging over time periods of 30 min or more. However, to support airport operations with the requirement for rapidly updated and highly accurate winds, quality control must be applied on each beam of data. This was the motivation for the development of NIMA. Figure 1b shows the NIMA moments overlaid on the spectra (median-filtered along the Doppler velocity axis) with the associated moment confidences in the right sidebar. Note that the NIMA moments are smooth as a function of range and that the first moment confidence is lowered for ranges with a low signal-to-noise ratio (SNR). This example dataset will be used to illustrate some of the components of NIMA processing described in section 4 and will be used as the basis of the simulation study presented in section 5.

The following sections will discuss the fuzzy logic combination methodology used in NIMA and how the fuzzy logic parameters can be tuned, followed by a description of the components of NIMA processing including confidence estimation. In the results section, NIMA performance is compared with that of a simple peak picking algorithm using a simulation study and the performance of the confidence estimates analyzed using human-truthed data.

2. Enhanced membership combination

Fuzzy logic (Yager et al. 1987) is used in NIMA because of its strength in addressing the natural ambiguities in measurement data, as well as in classification and pattern recognition. The basis of the fuzzy logic approach is to avoid strict thresholding until all information has been combined. That is, in contrast to a Boolean logic algorithm where strict thresholds for each test are used, fuzzy logic algorithms delay this thresholding until the final step. This leads to a much more robust and accurate result. To achieve this end, each of the disparate measurable quantities or characteristics to be used in the decision is translated into a common quantity, a membership value, using a membership function, which is specific to that characteristic. The original paper described the fuzzy logic combination of membership values at each point x in the radial velocity-range coordinate space of the spectral surface illustrated in Fig. 1. In that implementation, individual membership values MAi are combined using the center of gravity (Zimmerman 1996) method to create a total membership value MAT; that is, an algebraic weighted mean is calculated
i1520-0426-19-3-274-e1
If a particular membership characteristic, i = p, shows more skill than others, a larger weight αAp can be assigned in Eq. (1). In this way, the relative skill of the membership characteristics can be taken into account. In some cases the skill of the membership characteristic is dependent upon its associated membership value, for example, a membership characteristic that has high skill when its membership value is negative, but low skill when the membership value is positive. It would be desirable to combine membership values of this type in such a way that the total membership would be low if any such individual membership value were low. One possible combination rule with this property would be to take the minimum of the membership values; that is, MMT(x) = minj[MMj(x)] (Zimmerman 1996). This method does not, however, allow for the use of weights. An alternative method that does allow for weighting is to create a total membership value MGT by combining membership values MGj using a geometric weighted mean,
i1520-0426-19-3-274-e2
Note that the membership values MGj must all be positive. This combination method is particularly useful for membership characteristics that have high skill primarily when the membership value is zero or near zero. By assigning a small αG weight to this characteristic, its membership value will have a small effect on the total membership value if it is close to unity (and has low skill), but will have a large impact if it is near zero (and has high skill).
Each membership characteristic is assigned a combination method, that is, either algebraic or geometric. The membership combination paradigm used by NIMA for the total membership function MT includes both algebraic and geometric combination as
i1520-0426-19-3-274-e3
where MAT and MGT are the total membership values of the algebraic and geometric characteristics, respectively, and αAT and αGT are the sums of the weights of the algebraic and geometric characteristics, respectively. Typically each of the NIMA fuzzy modules use some combination of algebraic and geometric characteristics.

After the fuzzy combination step is the final defuzzification step, where a decision is made by comparing the total membership value MT(x) against the membership threshold value. If MT(x) exceeds the threshold then the data at x is a member otherwise, it is not a member. Section 4g gives a detailed example of the application of membership functions and combinations in NIMA.

3. NIMA configuration and tuning

Spectral characteristics can vary from site to site depending on local conditions including topography and the radar-processing parameters used. Thus it is desirable to be able to modify NIMA processing to optimize performance at a given site. For this reason, the processing in each of the NIMA fuzzy logic modules is controlled by a configuration file. This configuration file defines the membership function and the αA or αG weight associated with each membership characteristic as well as the final membership threshold. The offline adjustment of the threshold, weights, and membership functions to improve performance is known as tuning and generally proceeds as follows. The initial generation of a membership function for a membership characteristic is performed based on intuition and a sample of empirical evidence after examining “typical” spectra. The membership function is then adjusted based on viewing how well the calculated membership values compare with desired results for a larger and more varied set of spectra. The weights are adjusted based on how well the membership values compare with the desired result relative to each other. Once an initial set of membership functions and weights is chosen, a large number of spectra are processed and examined for cases where the resulting wind values look suspect, indicating the possibility of incorrectly calculated moments. These cases are then examined in detail to determine which processing module(s), if any, produced errors. The membership functions and weights are then modified until the error is mitigated. The large spectral sample is then reevaluated to ensure that no new errors were introduced by the modifications.

4. NIMA processing overview

The original paper described the NIMA processing philosophy and discussed the application of fuzzy logic at each point of the spectral surface to determine which are likely to be associated with clutter or atmospheric signal. Contiguous points on the spectral surface that exceed the membership threshold are grouped together to form a feature. The fuzzy logic module for classifying clutter results in a set of clutter features. The fuzzy logic module for classifying atmospheric signal results in a set of candidate atmospheric features. Another fuzzy logic module scores each of these candidate atmospheric features. The “best” atmospheric feature is used as a guide to which spectral points should be used in calculating the moments. Figure 2 shows the overall flow of NIMA processing as currently implemented and as described in the following sections.

The fuzzy logic approach used in NIMA is not limited to applications to points on the spectral surface. It can be applied to spectra at individual range gates, for example, to determine whether velocity folding is occurring. It can also be applied to a set of generated features, for example, to identify atmospheric or RFI features based on feature characteristics. A detailed example of the fuzzy logic processing as applied to RFI feature identification is given in section 4g.

a. Noise calculation

The method of Hilderbrand and Sekhon (1974) is used to calculate the noise level and its variance at each range gate of the Doppler spectra. The method also generates two statistical parameters that are saved and used in computing the moment confidence estimate. If the parameters indicate that there was a problem in determining the noise level from the spectra, confidence in the resulting moments will be decreased. For example, if the Hilderbrand and Sekhon chi-square test fails, implying that the noise distribution is not as expected, then the confidence value is lowered.

b. Median filtering

At several steps in the analysis, the spectra are median filtered. Prior to performing the 2D surface analysis described in section 4d, outliers are removed from the averaged spectra by using a 2D (range and Doppler velocity) median filter. Filtering results in a more consistent and useful local least squares fit; that is, it highlights the larger-scale features of the spectral intensity surface rather than the higher-frequency noise on the surface. The velocity and range halfwidths of the median filter are configurable in order to accommodate profilers with differing Doppler velocity and range resolution. Figure 3a shows the spectra from Fig. 1a after the application of the 2D median filter using halfwidths of two velocity bins and three range gates. Note that small point target contaminants are effectively removed by the filter.

Prior to performing the Gaussian fits to spectra described in section 4h, 1D median filtering of the averaged spectra along the velocity axis is performed to improve the performance of the fit. Although these 1D median-filtered spectra, as shown in Fig. 1b for the example spectra, have given reasonable results, some in the profiler community have expressed a desire for this choice to be configurable. NIMA can be configured to use 1D median-filtered, 2D median-filtered, or unfiltered spectra for the final moments calculation. The unfiltered spectra can be used if there is particular interest in detecting small-scale shears along range. Conversely, the 2D median-filtered spectra, which are smoothed in range as well as in velocity, can be used if there is a large penalty for false reports of shear or if the desire is to find shear over a larger scale.

c. Velocity folding

Ideally, the Nyquist velocity is chosen large enough during data collection to ensure that no radial velocities will exceed it. However, even if the Nyquist velocity is exceeded and the observed velocity is aliased or folded at some ranges, NIMA can determine the correct moments. NIMA detects the presence of velocity folding at a given range gate using a fuzzy logic module that considers membership characteristics such as the spectral intensities near the positive and negative Nyquist velocities relative to each other and relative to the noise level. Intensities that are above the noise level and close in value are indicative of velocity folding. When velocity folding is detected, the spectrum is unfolded by effectively doubling the Nyquist velocity and the number of velocity bins and copying the spectral data to satisfy the boundary conditions. Figure 4a shows a profile of 2D median-filtered spectra with velocity folding above 464 m. Because the spectral unfolding introduces some ambiguity itself, for example, the two strong signals that appear in Fig. 4b, unfolding is performed only at range gates where there is evidence to suggest that the spectral data are folded. The decision to unfold the spectra is made on a per range basis. Range gates that are not unfolded use the noise level to fill the spectral image region outside their Nyquist velocity. In Fig. 4b, the range gates below 464 m were not unfolded, resulting in less ambiguity as only one of the resulting signal images covers the entire range.

d. 2D surface analysis

After unfolding the spectra, where necessary, the resulting spectral surface is subjected to the 2D analysis described in Cornman et al. (1998). That is, local 2D quadratic least squares fits to the spectra are employed to determine the curvature and gradient. These values can then be used as membership characteristics for determining the location of clutter and atmospheric signal. The analysis is performed on several scalings of the spectral intensity: linear, dB, range-normalized linear, and range-normalized dB. In the normalized scalings, the linear or dB scaled data are normalized at each range to a value between zero and one (one being the maximum value for that range). This minimizes the effects of reduced intensity at the higher range gates as illustrated by the contrast between the dB-scaled spectral surface in Fig. 3a and the range-normalized dB-scaled surface shown in Fig. 3b.

e. Detecting ground clutter features

The detection of ground clutter was discussed in some detail in the original paper. The clutter detection algorithm consists of a fuzzy logic module to determine spectral points with high ground clutter membership, followed by global image processing to combine contiguous points with high membership values into distinct clutter features. Figure 3c shows the clutter features found in the example spectra. The locations of the clutter features are saved for use in later processing.

In the original paper, the curvature and gradient characteristics of the ground clutter spectral signal along with the expectation that clutter should be symmetric around zero velocity were used to distinguish it from the atmospheric signal. The current implementation of NIMA has added some additional characteristics based on the expectation that the clutter signal has a narrow Gaussian-like shape and/or corresponds to a user-configurable intensity template. (The Gaussian fit analysis is discussed further in section 4h.) When clutter and atmospheric signal are expected to appear in the same Doppler velocity bins, based on the calculated moments from previous beams in the same direction, total clutter membership can be downgraded for those bins by generating a negative clutter membership value for velocity bins near those first moment velocities.

The specific characteristics of clutter signals appear to vary widely between sites and typically a significant portion of a tuning effort is focused on this fuzzy module. For example, in applying NIMA to data from a profiler observing sea clutter in one of its beams, the membership function for the location characteristic was modified to detect the sea clutter signal that was consistently located slightly off zero at 0.5 m s−1 Doppler velocity.

f. Detecting atmospheric features

The detection of atmospheric signals proceeds similarly to that of the detection of ground clutter signal. A fuzzy logic module to determine spectral points with high “Doppler peak” membership is followed by global image processing to combine contiguous points with high membership values into distinct features. As described previously, the characteristics from the 2D surface analysis are combined with characteristics based on the spectral intensity, for example, the ratio of spectral intensity to the noise level or the range-normalized spectral intensity, to determine locations of atmospheric signal. Temporal history, that is, moments calculated from the previous beams in the same and opposite directions, is used to resolve potential ambiguities resulting from velocity unfolding. This is done by increasing the membership of points in the vicinity of the Doppler velocity of those moments. Clutter features detected in the spectra are used as a membership characteristic to explicitly downgrade atmospheric signal membership in regions where ground clutter was detected.

The membership threshold used to extract the atmospheric features is typically set low enough to ensure that a mostly continuous atmospheric feature is obtained, even under low SNR conditions. In the presence of an unusually strong signal, as illustrated by the precipitation spectra shown in Fig. 5a, this threshold may be too low to adequately separate the atmospheric feature, as shown in Fig. 5b (around range 500 m). An additional fuzzy logic module is used to determine whether the membership threshold should be raised. This module considers membership characteristics of the largest feature extracted. These characteristics include the average width of the feature and the existence of regions within the feature that did not exceed the threshold. If this test is positive, the set of features is discarded and a new set is created using a higher threshold. This process is repeated until the feature characteristics indicate that the threshold used was reasonable. Figure 5c illustrates the separated features generated when the higher threshold is applied.

After a set of well-separated candidate features are generated (illustrated in Fig. 3d for the example spectra of Fig. 1), they are then analyzed to identify which are most likely to correspond to atmospheric signal and which may correspond to RFI. Characteristics considered for atmospheric features are the sum of the atmospheric membership of the points included in the feature, the variance of the peak spectral intensity over range, the feature average width, and the feature average SNR. Also considered are the proximity of the feature to the Nyquist velocity, predicted RFI, and the signal from the previous beam in the same direction. The feature with the highest atmospheric feature membership is selected as the “best” candidate for the atmospheric signal.

In the event that this best feature does not cover all the range gates, two strategies are employed to attempt to remedy the situation. First, a fuzzy logic module is used to analyze the other features, which exceed the atmospheric feature membership threshold to determine which, if any, could be connected with the existing best feature to extend the feature over additional range gates. Membership characteristics for these disjoint (not connected) features include the difference in Doppler velocity between it and the best feature as well as the continuity in range of a combined feature. In Fig. 3d, feature #2 is a disjoint feature that can be connected to feature #1. Second, a strategy of lowering the atmospheric membership threshold in the vicinity of the feature top or bottom edges is used to attempt to extend the feature into the missing ranges.

g. Radio frequency interference (RFI)

RFI contamination was observed during the analysis of the Point Loma dataset described in the original paper. In this dataset, the RFI appeared at a constant Doppler velocity over range and time, and the RFI signals were easily characterized and identified. Application of NIMA to the data from the Juneau profilers showed that this characterization was overly restrictive and the original RFI algorithm was evolved into a more sophisticated fuzzy logic module. This fuzzy logic module analyzes various characteristics of the atmospheric features to determine whether they should be classified as RFI. RFI features in a single beam tend to be narrow and constant over range in width, intensity, and Doppler velocity. RFI may shift in Doppler velocity from beam to beam over time. If the RFI is observed consistently from beam to beam, this movement can be tracked and subsequently predicted. If an RFI feature does not cover all the ranges, a fuzzy logic module is applied to identify disjoint features, which may be associated with the RFI signal. This processing is the same as that used for detecting disjoint atmospheric features, except that the membership functions and relative weights are tuned to the more specific characteristics of RFI features. Like the ground clutter features, the location of an identified RFI feature is saved for use in the subsequent calculation of spectral moments.

The following detailed explanation of the RFI feature identification fuzzy logic module is given to illustrate the use of membership functions and the application of Eq. (3) used in each of NIMA's fuzzy logic modules. The results of the RFI fuzzy logic module processing that identify feature 4 in Fig. 3d as RFI are given in Table 1. The membership functions that convert the membership characteristics (C) to the membership values (M) in columns C and M, respectively, in Table 1, are shown in Fig. 6. The RFI identification module includes both algebraic and geometric membership characteristics. The algebraic characteristics are associated with measurable properties of an RFI feature and have membership values in the range [−1, 1]. The geometric characteristics are used to suppress identification of a given feature as RFI under circumstances where other characteristics may be misleading. For example, a feature covering very few range gates will nearly always have the very low feature width and midpoint variances that lead to high RFI membership values. Geometric membership values are in the range [0, 1]. For simplicity, this example does not include any of the history-dependent membership characteristics.

A constant and relatively narrow signal width in the velocity dimension, as a function of range, is characteristic of an RFI feature. The average feature width is calculated over the range gates included in the feature. The variance of the width over the range gates and that variance normalized by the number of range gates in the feature are also calculated. The membership functions for these characteristics are shown in Figs. 6a–c. The true RFI feature 4 (Fig. 3d) shows high RFI membership values, M, for each of these characteristics (Table 1). Membership based on width variance drops considerably for feature 3 when that variance is normalized over the number of range gates. The atmospheric feature 1 has negative RFI membership based on feature width.

Another RFI characteristic is a relatively constant Doppler velocity value for the given beam. This characteristic is measured by the variance of the feature velocity midpoint over the range gates included in the feature, that variance normalized by the number of range gates in the feature, and the slope of the linear fit of the feature midpoint versus the range gate number. The membership functions for these characteristics are shown in Figs. 6e–g. Note that membership is high for low variance and falls off for higher variances; membership is high for zero slope, but becomes negative if the feature has a large slope. Again the true RFI feature 4 shows high membership values (M) for each of these characteristics. Feature 3 shows high membership based on the midpoint variance, but this membership drops when the variance is normalized over the range gates and membership is negative based on the slope of the midpoint.

Unlike atmospheric signals, RFI signals tend to have a relatively constant spectral intensity as a function of range. The variance of the peak signal intensity over the range gates within the feature is calculated and the RFI membership function is shown in Fig. 6d. The atmospheric feature 1 has a very negative membership value because of its highly varying intensity.

In some instances, RFI may be observed consistently at a specific Doppler velocity, or it may change position over time. The proximity of an atmospheric feature to the expected position of RFI is used as an algebraic membership characteristic in the identification of RFI features. In order to determine this expected position, the average Doppler velocity of an identified RFI signal is tracked over time and successive beams using the adaptive polynomial technique of Morel and Passi (1986). In this computationally efficient technique, a quadratic model is used for the RFI position after accounting for velocity folding. The adaptive nature of the algorithm allows it to track both stationary and nonstationary RFI signals. The model is used to predict the expected RFI Doppler velocity position in the current beam. After the initial detection of RFI, the uncertainty in this prediction is large, but tends to decrease as more data are included in the model coefficient calculations. After a specified number of beams, for example, configured by default as 5, in which the RFI signal is not observed, tracking and position prediction are discontinued and removed from consideration in detecting RFI features in subsequent beams.

Two geometric membership characteristics are included with the goal of suppressing the classification of a feature as RFI under specific conditions. The first such characteristic is the number of range gates included in the feature. If there are too few, the width, midpoint, and intensity variances may be artificially low and could result in an erroneous classification. The membership function for this characteristic is shown in Fig. 6h. Both features 2 and 3 get a zero membership value for this characteristic keeping them from being classified as RFI. The other geometric characteristic is the average (over feature range gates) of the difference between the feature midpoint Doppler velocity and the expected atmospheric first moment. The expected first moment is predicted based on persistence from the previous beam in the same direction. RFI membership is low near the predicted first moment and is high away from it. This membership characteristic can protect the atmospheric signal from being classified as RFI if there is significant overlap of the two signals. Features whose total RFI membership value exceeds the membership threshold are classified as RFI. In the example in Table 1, feature 4 is classified as RFI based on its total membership value of 0.984, which exceeds the 0.90 threshold value.

h. Calculating Doppler moments

After the best candidate atmospheric feature is selected, the portion of the Doppler spectra associated with this feature is used to determine the data from which the zeroth, first, and second moments are calculated. Either the 2D median-filtered, the 1D median-filtered, or the unfiltered spectral data can be used for the moments calculation, as specified in the configuration file. The atmospheric feature can be considered as providing a “first guess” as to the range of Doppler velocities within which lies the desired atmospheric signal. The velocities that mark the limits of this signal are referred to as the spectral cutoffs. For clean spectra, the spectral cutoffs are simply those points where the signal power drops to the noise level on either side of the signal peak for a given range gate, that is, the “standard” method. This simple method may produce erroneous moments if there are overlapping signals, for example, overlapping atmospheric signal and ground clutter. For overlapping signals that have distinct peaks, the desired spectral cutoff is at the local minimum between the two signal peaks. The fuzzy logic module for spectral cutoff determination thus utilizes membership characteristics of the spectral surface, which can identify that local minimum, for example, curvature and slopes from the local 2D fits. This process results in the identification of points on the spectral surface whose “spectral cutoff” total membership value exceeds the spectral cutoff membership threshold. As shown in Fig. 7, these points tend to be grouped together in spectral cutoff regions. Because the entire spectral surface is analyzed, many such spectral cutoff regions may be generated, but only those that are adjacent to the atmospheric signal are actually used in the moment determination.

The spectral cutoff regions are used in conjunction with the atmospheric feature and the location of clutter and RFI features to select the spectral points that are used to calculate the moments at each range gate. In the absence of contaminants, the spectral data from the peak to the spectral cutoffs can be used directly for this calculation. However, when the atmospheric signal overlaps clutter, RFI, or other contaminants, the resulting moments are likely to be biased if contaminated points are included and also if they are simply excluded in the calculation. To minimize this bias, NIMA replaces the contaminated data by an estimate of the uncontaminated atmospheric spectral intensity. These estimates are made using the expectation that the atmospheric signal above the noise should be Gaussian in shape,
i1520-0426-19-3-274-e4
where s is the spectral intensity less the noise level intensity, x is the Doppler velocity, x and σ2 are the first and second Doppler moments, respectively.

At each range gate, the atmospheric feature is used to identify the spectral points to which the Gaussian model is applied. A series of fits to the model are tried, starting at the peak of the feature signal and incrementally expanding the number of points included around that peak. The confidence in each of these fits is calculated as described below and the fit with the highest confidence is selected as the Gaussian model. If points away from the peak are contaminated by noise or clutter, the confidence in the fit will tend to be lowered, thus a fit containing these contaminants will tend not to be selected. To minimize noise contamination, points that are less than one noise sigma above the noise floor are excluded from the fits. Spectral points in the cutoff regions as well as those within the clutter features or RFI features are also excluded from the fits. The spectral intensities of these excluded points are replaced with those predicted by the Gaussian model before calculating the moments by the standard method (van de Kamp 1988) using the linearly scaled spectral data. Figure 7 shows a portion of the example spectral curve at 440 m overlaid with the position of the spectral cutoff regions, the Gaussian curve fit, and the resulting moments.

The Gaussian model is applied by first applying a dB scaling, 10 log10(s), to the spectral data and expanding around the spectral midpoint velocity x0. This reduces the exponential form of Eq. (4) to the quadratic relationship,
10sABxx0Cxx02
where
i1520-0426-19-3-274-e5
Although fitting the Gaussian model in the linear spectral field would provide for more accuracy (May et al. 1989), the logarithmic scaling is used for computational simplicity. A linear least squares quadratic fit to the dB-scaled spectra data is performed to solve for A, B, and C. The singular value decomposition (SVD) (Press et al. 1988) method is used here because it is robust and provides a measure of how well-behaved the system of equations is, that is, the condition number. Gaussian parameter estimates for the first and second moments, x and σ2, respectively, are then calculated using Eq. (6):
i1520-0426-19-3-274-e6

A confidence in the Gaussian model is determined by a fuzzy logic module using characteristics including the chi-square probability that the data used fit the Gaussian model (Freund 1992). This is the probability that chi-square, if the model were correct, would be larger than the value calculated for the data used. Other characteristics used in determining confidence in the Gaussian model are the number of points used in the fit, the condition number of the SVD, the root-mean-square (rms) difference between the data and the model, the difference between the estimated midpoint x0 and the first moment estimated from the Gaussian model x and the coefficient of the quadratic term C, which must be negative for the Gaussian width σ to be real. The confidence in the Gaussian model fit to the spectral data is used later in assessing the confidence in the calculated moments.

i. Continuity of moments

After the first and second moments have been calculated at each range gate, they are evaluated with respect to the assumption that the atmosphere is continuous and there should be no sharp discontinuities in the Doppler moments as a function of range. Their continuity is tested using characteristics including the chi-square probabilities from local linear and quadratic fits of the moments as a function of range. By “local” it is meant over ±n range gates, centered at the given range. The value of n is typically 2, but is configurable. Fits are calculated centered at each range gate where there are sufficient surrounding range gates to satisfy the ±n requirement. A moment is judged based on all the fits in which it is included; that is, 2n + 1 fits except for the top and bottom n range gates. For each moment a “badness of fit” total membership value is calculated and the moment is classified as “bad” if that membership value exceeds the membership threshold. Missing moments are always classified as bad.

When bad first moment values are found at one or more range gates, an attempt is made to estimate those first moments based on continuity considerations. Because no one method is clearly the best, several methods are used to make an estimate, including interpolation based on quadratic and linear fits to moments in surrounding range gates. In some cases, the various estimation methods produce widely varying results, so the median rather than an average of the results of the various methods is used as the first moment estimate. This estimate is now used instead of the atmospheric feature to select a new set of spectral points as described in section 4h. The first and second moments are recalculated using this new set of spectral points. The Gaussian model fit to these points and its associated moments estimate and confidence are also calculated.

Using the “good” first moments and the recalculated moments, continuity is retested and a new set of bad moments may be identified. If there are any bad moments remaining, the first moment is reestimated using the continuity considerations and this estimated value is reported as the first moment. Using this final set of first moments, the continuity is tested again. The resulting badness membership at each range gate is saved for use in the final confidence estimation. A similar method is applied to the second moments. The final NIMA moments for the example spectra are shown in Fig. 1b. The first and second moments for the 1540–1925-m range gates were estimated values.

j. Moment confidence

There are cases where even a human expert has trouble locating all or part of the atmospheric signal region. Such cases can include data with a low SNR, or when ground clutter or RFI features are located near the atmospheric signal. A human expert may have lower confidence in the moment estimates when the atmospheric signal has discontinuities in range or when multiple features are present. A confidence index has been developed to mimic the methods that a human expert might use in estimating the quality of the data. A confidence estimate for the calculated moments is useful for downstream processing, that is, wind, wind shear, or turbulence estimation. Confidence values are in the range of zero to one, zero indicating no confidence and one indicating complete confidence in the moment value.

The confidence in the moments is estimated based on a variety of characteristics generated during the NIMA processing. The characteristics used for calculated moments include the confidence in the Gaussian model of the estimated atmospheric signal region, the difference between the calculated moment (van de Kamp 1988), and the moment estimated from that Gaussian model, the signal-to-noise ratio, the continuity badness score, and the noise statistical measures. Confidence is high when the Gaussian model is good, the moments estimated from the Gaussian model and the calculated moment agree, the signal-to-noise ratio is high, the moments are continuous, and the statistical parameters from the noise level determination are as expected.

Different characteristics are used for moments estimated from continuity considerations rather than calculated from the spectral data. Recall that this estimate is the median of estimates made using several models to interpolate or extrapolate based on moments in nearby range gates. The characteristics used for estimated moments include the variance between the different estimates calculated using the different methods, the distance in range gates from the closest calculated moment, and the continuity badness score of the estimate. For both calculated and estimated moments, the confidence is lowered if the moment is near the predicted position for RFI signal or near ground clutter, or if the feature upon which the moments were based was overly wide in the velocity dimension.

The moments confidence for the example spectra are shown in Fig. 1b. Note that the confidence is lowered in the region of weak signal, 1100–1265 m, and at the range gates, 1595–1925 m, where the moments were estimated rather than calculated.

5. Results

Two methods are used to demonstrate NIMA's performance in spectral moment and moment confidence estimation. The first is to perform a simulation where exact truth is known. The computed moments can be compared with the true values. The second is to compare computed moments with those determined by human experts. This requires the assumption that the human experts' moments may be used as a true value. The advantage of this second approach is that the ambiguity of real data may be studied. It will be shown using both methods that NIMA produces first moments with smaller errors than those of a peak-picking algorithm. In order to use the moment confidence estimates to identify poor quality moments, it is necessary to relate the confidence value to the measurement error. It will be shown using both methods that most of the large first moment errors are assigned a low confidence by NIMA. To be useful, a confidence value should decrease with increasing difference between the calculated moment and the true moment (as determined by a human expert or simulation). A practical way to determine whether this relationship holds is to consider the average difference for a set of calculated moments each with a confidence value that exceeds some confidence threshold. That average difference should decrease as the confidence threshold is increased. This property of the NIMA first moment confidence will be demonstrated using the human-truthed data.

a. Simulated data

1) Methodology

Zrnić (1975) suggests that a power spectrum may be simulated by using an exponential distribution with mean Sk + N where Sk is the idealized power spectrum to be simulated in spectral bin k, and N is the noise floor. The use of the exponential distribution is justified by Zrnić (1980) in the case of an atmospheric signal. The assumptions are that I and Q data measured by a profiler are independent zero mean Gaussian processes with identical variances. Here I and Q are assumed independent. The times series for I (or Q) is allowed to be correlated with itself. This allows for windowing effects in the times series processing. Under these assumptions, the Doppler spectrum will have an exponential distribution. May and Strauch (1998) suggest modeling ground clutter in a similar fashion. This is based on the observation that ground clutter has similar statistics to the atmospheric signal. To include RFI and point targets, a similar methodology is assumed. That is, each of these signals are assumed to be independent zero mean Gaussian processes, and the I and Q signals are the sum of the resulting I and Q values for each of the corresponding contaminants plus the noise plus the atmospheric signal. This is based on the fact that the Doppler spectrum of RFI has an appearance similar to a narrow atmospheric signal, and this similarity was the reason a special module had to be created in NIMA to distinguish RFI from the atmospheric signal. Other models are possible such as a deterministic model for RFI, but treating the RFI in a way similar to atmospheric signal gives a simple way to generate realizations that have similar statistics to atmospheric signals. Under these assumptions, it follows by the analysis in Zrnić (1980) that an appropriate model will be to make Sk the sum of several Gaussian models, one each for atmosphere, ground clutter, RFI, and point targets.

The model for the simulation is the data in Fig. 1. The stacked spectra from this example were examined by eye to determine the approximate mean μ, half-width σ, and peak intensity S0, of the component signals at each range gate. Models of the form
i1520-0426-19-3-274-e7
were produced for each component signal. The final Sk was a linear combination of these component models. The noise N was set based on noise estimates from the example spectra. The theoretical SNR was calculated as [(2π)1/2S0σ]/(2NVN), where S0 is the amplitude of the atmospheric signal amplitude and VN is the Nyquist velocity. Figure 8 shows the idealized simulated spectral surface and the true moments of the atmospheric signal. At each range gate and each bin an exponential random number generator was simulated by calculating (Sk + N) ln(R[0, 1]) where R[0, 1] is a uniformly distributed random variable on the interval [0, 1]. This produced one set of unaveraged spectra. Fifty such sets (the number of spectral averages used in the Juneau system) were averaged to generate one simulated profile. NIMA was applied to the profile, and the first moments and first moment confidences were calculated and the NIMA first moments were compared to the true first moments given by the atmospheric model. A thousand realizations were generated and at each range gate, the average error and its variance were calculated as were the average absolute error, its variance, and the average confidence from NIMA.

If a comparison of NIMA and POP is desired, an algorithm similar to POP must be implemented. The simple peak algorithm (SPA) finds the peak signal at each range gate and integrates to the noise floor to find the moments. This is similar to what was done in a simulation by May and Strauch (1998). This algorithm is different from POP as no I and Q processing is done in the simulation and no attempt has been made to implement the POP ground clutter rejection algorithm. The Radian system performs some I and Q averaging, a window function is used in converting from I and Q to Doppler spectra, and a notch filter is applied to reduce the effects of ground clutter at certain range gates. This I and Q processing gives spectra similar to the realizations produced here, thus SPA should have similar performance characteristics to that of POP. SPA is also run on each simulated profile, and comparisons are made to the true first moment values. Although errors were calculated for both first and second moments, the discussion emphasizes the first moment performance as the second moment analysis is more problematic. Second moment error comparisons are meaningful only when the first moments are correctly identified and in the real data there is greater ambiguity and human expert disagreement regarding the true second moment values because many of the spectra are not Gaussian.

2) Atmospheric signal

The first simulation includes only the atmospheric signal, that is, without ground clutter, RFI, or point targets. Figure 9 presents the NIMA and SPA first and second moment errors as a function of range and the associated theoretical SNR values from the idealized spectra compared with average SNR values calculated by SPA and NIMA. NIMA and SPA both perform well on signals with a high SNR when no other contaminants are present. However, notice that at ranges between 1500 and 2000 m, where the SNR is low, SPA has much larger first moment errors than NIMA. The variances of the SPA estimates ranged from 20 m2 s−2 to 110 m2 s−2 in these ranges. The NIMA confidences were low in these ranges (all were less than 0.4, where 0.67 is considered good), and the variances were all smaller than 0.5 m2 s−2, indicating NIMA's superior results in low SNR conditions. If all data with true SNR values smaller than −10 dB are removed, then both SPA and NIMA have average absolute errors smaller than 0.1 m s−1. For low SNR signals, NIMA tends to track the signal better than SPA, and furthermore low first moment confidence is given to these moments. Figure 9 also shows that the second moment errors (measured as errors in the spectral half-width) are comparable for NIMA and SPA for ranges where the first moment errors are small. Both algorithms tend to underestimate the second moments.

3) Ground clutter

The second simulation involves the addition of ground clutter to the atmospheric signal. Figure 10 shows the stacked spectra from the lowest four ranges where the atmospheric signal partially overlaps with the ground clutter. Comparing the performance results in Fig. 11 with those in Fig. 9 shows that for these ranges both SPA and NIMA show increased errors. The average bias for SPA in these ranges is 1.44 m s−1, while the average bias for NIMA is 0.13 m s−1. This shows that NIMA reduces the bias due to the clutter when the atmospheric signal is well separated from or only partially overlaps the clutter. This is because of the way NIMA first moments are computed (see section 4h). While NIMA uses a fit on the part of the atmospheric signal that does not overlap with the clutter, SPA simply integrates to the noise floor thus including the clutter in the calculation of the first moment. SPA also selects the ground clutter over the atmosphere where the clutter has a higher peak than the atmospheric signal, that is, 1155 m. Note that in the presence of clutter, the SPA-calculated SNR is no longer well correlated with the magnitude of the SPA first moment error because the clutter signal selected may have a good SNR. Various methods for ground clutter mitigation have been applied in the I and Q domain (Jordan et al. 1997; May and Strauch 1998). NIMA reduces the biases due to clutter in the spectral domain. Although the I and Q and clutter rejection techniques used by the Radian system were not included in this simulation, examples in the dataset used in the human verification effort indicated that POP does select ground clutter over atmospheric signals when the atmospheric signal is weak. POP allows for clutter mitigation at user-specified range gates, but the presence of clutter varies at different sites. For example, in Juneau clutter signals can be present up to the 1000-m range. It should be noted that the bias in the NIMA first moment estimate could be larger depending on the amount of overlap between the atmospheric signal and the clutter, which in turn depends on the locations and second moments of the two signals. Such cases do cause problems for NIMA and this is a topic of current investigation. This problem has a particular impact in vertical beam estimates of the vertical wind.

4) Radio frequency interference

In the next simulation, RFI was added to the ground clutter and the atmospheric signals. The RFI problem is acute in the Juneau system since two of the profilers are in direct line of sight of each other. The RFI can have a similar appearance to a narrow atmospheric signal. This was the motivation for adding an RFI detection module to NIMA. Before this module was added to NIMA, RFI was sometimes selected as the atmospheric signal especially in cases where the atmospheric signal was weak. To illustrate this point, the simulated atmospheric signal was attenuated by a factor of 0.15 to produce a weak signal. Figure 12a shows a realization in which NIMA selects the weak atmospheric signal in preference to the stronger RFI signal, while Fig. 12b shows NIMA with the RFI module disabled selecting the RFI as the atmospheric signal. SPA also selects RFI whenever the RFI signal has a higher peak than the atmospheric signal. To analyze this, the atmospheric signal intensity was returned to its original value and Fig. 13 shows a comparison of NIMA and SPA average errors in the presence of ground clutter and RFI. Notice that there are additional large errors when SPA selects the RFI instead of the atmospheric signal and that these errors are consistent with the errors in the POP moments shown in Fig. 1a. Notice that NIMA does have some average errors around 0.5 m s−1, but the moments associated with these errors have a very small average confidence. These errors occur at heights where the SNR is low. All data with an average confidence above 0.4 have an average absolute error smaller than 0.3 m s−1.

5) Point targets

The final simulation adds a point target to the previous signals. SPA selects this point target in the range gates where the point target has a higher peak than the other signals (around 500 m). NIMA does not select the point target. The point target is often removed by the median filter (see Fig. 1a vs Fig. 3a). In cases with larger point targets, a feature might be formed, but such a feature would receive a low score as it covers few ranges. Such isolated targets usually do not pose much of a problem for NIMA. Figure 14 compares NIMA and SPA average errors in the case when the simulation contains the atmospheric signal, ground clutter, RFI, and a point target. Applying a confidence threshold of 0.4 to these data results in an average absolute error of less than 0.3 m s−1. Notice the large number of errors for SPA as compared to the case when only an atmospheric signal was present. If contaminants are present, a peak finding algorithm will select a contaminant if the contaminant has a higher peak value than the atmospheric signal. If a contaminant is selected by the peak finding algorithm, then the computed SNR will tend to be larger than the true SNR. So although SNR can be useful in assessing data quality (Angevine et al. 1993), it can be misleading if used alone as a confidence index for a peak finding algorithm.

6) Summary of simulation results

In these realistic, though limited, simulations, NIMA tracks weak signals with less error than SPA. NIMA mitigates ground clutter biases (in cases where well-defined separate atmospheric and clutter peaks are seen in the stacked spectra) and tends not to select RFI nor point targets. The first moment confidence may be used to identify poor quality data. The next section will show that these conclusions also hold for the human-truthed data.

b. Human-truthed data

The version of NIMA described above was installed in December 1998 for use on three 915-MHz wind profilers located in Juneau, Alaska. In 1999, the 10-min confidence-weighted average winds (Goodrich et al. 2002) were more reliable and consistent than those reported during the previous year when an earlier version of NIMA without the enhanced RFI processing had been used. This empirical evidence is corroborated by a more formal verification effort, described in Cohn et al. (2001). In this verification study, the first moments from NIMA were compared to first moments estimated by human experts. The dataset consisted of 172 profiles, that is, average spectra from single beams, each containing 36 range gates. The profiles were selected to test NIMA under a variety of atmospheric and contaminant conditions, with the greatest number selected to be near times when a research aircraft was flying near the profiler. In addition, some of the spectra were chosen at times when NIMA winds showed large temporal discontinuities. This was a challenging dataset containing spectra with weak signals, ground clutter, RFI, and point targets. To evaluate these profiles, a graphical interface was developed (Cohn et al. 2001) that allowed the human expert to insert the spectral cutoffs. Estimates of the first and second moments were then computed using a quadratic fit to the dB-scaled spectral data, Eqs. (5)–(6). The human expert had the additional option of selecting the location of the first moment if the quadratic fit did not give an adequate estimate. The human expert also entered a confidence value for each moment.

This dataset of over 6000 spectra was processed by NIMA and POP. The first moments output from each algorithm were then compared to the first moments given by the human experts. A summary of the performance of the two algorithms is given in Table 2, where the average error, average absolute error, and rms errors are reported. There were a fair number of outliers in the NIMA first moments as signified by the rms error of 1.32 m s−1. The largest outliers were caused by weak signal cases as well as cases where NIMA erroneously selected a part of the ground clutter as the atmospheric signal. The comparison of the POP first moments with the human-generated ones resulted in errors roughly three times larger than those from NIMA. The largest POP outliers were caused by weak signal cases, ground clutter, and cases where RFI was selected instead of the atmospheric signal. RFI is an unavoidable problem in Juneau since two of the profilers are in a direct line of sight of each other.

The first moment confidence values were also evaluated. It should be expected that the average absolute error and the rms error should become smaller as data with low NIMA confidence are removed. In Fig. 15a, the horizontal axis represents the fraction of the data removed via confidence thresholding. The curve represents the average absolute error as a function of the fraction of data removed. Thus the average absolute error of the data is reduced to about 0.3 m s−1 (roughly the velocity resolution of the Juneau profilers) if the lowest 20% of the data as ranked by NIMA confidence is removed. This corresponds to using only data above the NIMA first moment confidence threshold of 0.676. If SNR is used as a surrogate confidence for POP, nearly 60% of the data with lowest SNR must be removed to obtain the same level of performance. The average absolute error for NIMA first moments levels off at about 0.2 m s−1 when about half of the data are removed. This is less than the velocity resolution and so increasing the first moment confidence threshold does not significantly further reduce the absolute error. Thus NIMA first moment confidence has the property that increasing the confidence threshold decreases the average absolute error. A good quality control index should have this property.

Figure 15b is an analysis of the rms error of NIMA first moments as a function of the percent of data removed as ranked by NIMA first moment confidence. The entire dataset has an rms error of around 1.3 m s−1. If 20% of the data are removed, this rms error is reduced to around 0.75 m s−1. This indicates some significant outliers have been removed by applying a NIMA first moment confidence threshold of 0.676. Similarly, if the data with the lowest 20% SNR is removed, the rms error of the POP moments is reduced from 4.71 to 1.4 m s−1. There are still significant outliers because POP often selected the RFI or ground clutter whose resulting SNR was not low. If 30% of the data are removed based on NIMA confidence, the rms error for NIMA first moments drops below 0.5 m s−1. The NIMA first moment confidence values show skill in removing outliers, as is shown in Fig. 15b. This is further confirmed by Cohn et al. (2001) where the distribution of errors as a function of confidence is shown.

NIMA in general produces more accurate moments than traditional algorithms based on picking the peak spectral intensity, particularly in spectra containing contaminants, but in some cases the algorithm is not able to give an accurate first moment estimate. In these cases, a low confidence is usually given to the moment value. In many of these cases the human expert also has difficulty in estimating the moment. First moment estimates with high confidence values have high agreement with human-expert-produced first moment estimates. The confidence estimates can be used in the generation of derived products such as winds, wind shear, and turbulence and can contribute to confidence estimates for those derived products (Goodrich et al. 2002). These confidence estimates can be useful in improving operational performance, that is, reducing false alarm rates, by limiting the use of moments or derived products for which confidence is low.

6. Conclusions and future work

NIMA is a mature algorithm. It has been employed operationally at the Juneau International Airport since 1998. The first moment values have been used to produce winds and the first moment confidence values used as factors in the wind confidence estimates (Goodrich et al. 2002; Cohn et al. 2001). The long-term goal in Juneau is to produce a warning system. This will require reliable and rapidly updated wind and wind shear estimates as well as turbulence estimates. Each of these products will require a confidence value to help reduce false alarms. When the confidence value is low for a product, no warning will be produced based on that product. Where these products are produced from profiler data, a factor in each of these confidence values will be the first moment confidence, because a poor first moment confidence might indicate the derived product quality is poor. An important first step in this process is to obtain high-quality moment and moment confidence values. It has been shown that the NIMA first moment values have excellent agreement with human-produced first moments when the NIMA first moment confidence value is high. This was also confirmed in the limited simulation exercise.

Additional details have been given concerning NIMA processing with an emphasis on a new fuzzy logic membership combination method and a component to handle time-varying RFI contamination. RFI is a significant problem in Juneau since two of the profilers are in line of sight to each other and it has been shown that a traditional peak-finding algorithm often selects RFI instead of atmospheric signal.

NIMA has been applied to a limited amount of data from profilers around the world. Each new site has provided new challenges and opportunities to improve the algorithm. Although there are still some areas targeted for improvement, NIMA has reached a fairly stable level. In applying NIMA to data from a new site, reasonable levels of performance can be obtained with minor parameter tuning. One current topic of interest is an automated tuning process when installing NIMA at a new site. Recent modifications to the NIMA software infrastructure now allow for different tunings when several scan strategies run concurrently on the same profiler.

One reviewer suggested that time continuity be used in processing to connect disjoint features. This is an excellent suggestion and will be incorporated into the algorithm by using the moments from the previous beam in the same direction as a membership characteristic. The first moments in cases where the overlap between atmospheric signal and clutter is nearly complete also need improvement. The effects of ground clutter on near-zero winds is especially important in the analysis of vertical beams. This has received limited attention in the current work because of our particular interest in high wind cases and the lack of a vertical beam in the Juneau scan sequence. Studying and improving the performance of NIMA second moments is also of interest as these second moments are used in turbulence estimation.

Identification of precipitation is another topic of current research. At operating frequencies where clear-air and precipitation signals are separate and distinct, it would be desirable to be able to report moments from both signals; where the signals are not separable, it would be useful to identify whether or not the moments are associated with a precipitation signal or not as this could affect the confidence in the resulting winds and turbulence calculations.

Acknowledgments

We thank Mercedes Maruri Machado of Universidad del Pais Vasco for allowing us to apply NIMA to her wind profiler data containing sea clutter and Volker Lehmann of the Deutsche Wetterdienst for valuable feedback and suggestions regarding NIMA processing. We would like to thank Dr. Stephen Cohn for our valuable discussions on profiler technology, Dr. Rod Frehlich for our discussions on simulation, and also Andrew Weekley and Jeremiah Stone for their contributions to the data analysis effort. This research is in response to requirements and funding by the Federal Aviation Administration (FAA). The views expressed are those of the authors and do not necessarily represent the official policy or position of the FAA.

REFERENCES

  • Angevine, W. M., Avery S. K. , and Kok G. L. , 1993: Virtual heat flux measurements from a boundary-layer profiler-RASS compared to aircraft measurements. J. Appl. Meteor., 32 , 19011907.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Carter, D. A., Gage K. S. , Ecklund W. L. , Angevine W. M. , Johnston P. E. , Riddle A. C. , Wilson J. , and Williams C. R. , 1995: Developments in UHF lower tropospheric wind profiling at NOAA's Aeronomy Laboratory. Radio Sci., 30 , 977.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Cohn, S. A., Goodrich R. K. , Morse C. S. , Karplus E. , Mueller S. , Cornman L. B. , and Weekley R. A. , 2001: Radial velocity and wind measurement with NIMA: Comparisons with human estimation and aircraft measurements. J. Appl. Meteor., 40 , 704719.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Cornman, L. B., Goodrich R. K. , Morse C. S. , and Ecklund W. L. , 1998: A fuzzy logic method for improved moment estimation from Doppler spectra. J. Atmos. Oceanic Technol., 15 , 12871305.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Freund, J. E., 1992: Mathematical Statistics. 5th ed. Prentice Hall, 658 pp.

  • Goodrich, R. K., Morse C. S. , Cornman L. B. , and Cohn S. A. , 2002: A horizontal wind and wind confidence algorithm for Doppler wind profilers. J. Atmos. Oceanic Technol., 19 , 257273.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Hilderbrand, P. H., and Sekhon R. S. , 1974: Objective determination of the noise level in Doppler spectra. J. Appl. Meteor., 13 , 808811.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Jordan, J. R., Lataitis R. J. , and Carter D. A. , 1997: Removing ground and intermittent clutter contamination from wind profiler signals using wavelet transforms. J. Atmos. Oceanic Technol., 14 , 12801297.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • May, P. T., and Strauch R. G. , 1998: Reducing the effect of ground clutter on wind profiler velocity measurements. J. Atmos. Oceanic Technol., 15 , 579586.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • May, P. T., Sato T. , Yamamoto M. , Kato S. , Tsuda T. , and Fukao S. , . 1989: Errors in the determination of wind speed by Doppler radar. J. Atmos. Oceanic Technol., 6 , 235242.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Morel, C., and Passi R. , 1986: Use of adaptive polynomials for real-time data editing. J. Atmos. Oceanic Technol., 3 , 494498.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Press, W. H., Flannery B. P. , Tenkolsky S. A. , and Vetterling W. T. , 1988: Numerical Recipes in C. Cambridge University Press, 735 pp.

  • Schumann, R. S., Taylor G. E. , Merceret F. J. , and Wilfong T. L. , 1999: Performance characteristics of the Kennedy Space Center 50-MHz Doppler radar wind profiler using the median filter/first guess data reduction algorithm. J. Atmos. Oceanic Technol., 16 , 532549.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • van de Kamp, D. W., 1988: Profiler Training Manual #1: Principles of Wind Profiler Operation. National Weather Service, 49 pp.

  • Weber, B. L., Wuertz D. B. , Welsh D. C. , and McPeek R. , 1993: Quality controls for profiler measurements of winds and RASS temperatures. J. Atmos. Oceanic Technol., 10 , 452464.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Yager, R. R., Ovchinnikov S. , Tong R. M. , and Nguyen H. T. , 1987: Fuzzy Sets and Applications, Selected Papers by L. A. Zadeh. John Wiley and Sons, 685 pp.

    • Search Google Scholar
    • Export Citation
  • Zimmerman, H. J., 1996: Fuzzy Set Theory and its Applications. 3d ed. Kluwer Academic Publishers, 435 pp.

  • Zrnić, D. S., 1975: Simulation of weather like Doppler spectra and signals. J. Appl. Meteor., 14 , 619620.

  • Zrnić, D. S., . 1980: Spectral statistics for complex colored discrete-time sequence. IEEE Trans. Acoust. Speech Signal Process., 28 , 596599.

    • Crossref
    • Search Google Scholar
    • Export Citation

Fig. 1.
Fig. 1.

Contour plots of example Doppler spectra from the Juneau Lemon Creek 915-MHz profiler including atmospheric, clutter, point target, and RFI signals. (a) Unfiltered dB scaled spectra overlaid with POP moments, (b) 1D median-filtered dB scaled spectra overlaid with NIMA moments with moments confidence shown in the right sidebar, first moment confidence in red, second moment confidence in green. Note that confidence is lower in low SNR regions

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Fig. 2.
Fig. 2.

This chart shows the primary external input and output data and important intermediate data. Each circle, aj, indicates processing that is described in section 4

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Fig. 3.
Fig. 3.

Illustrations of the results of NIMA processing on spectra from Fig. 1. (a) 2D median-filtered dB scaled spectral data. (b) Range-normalized dB scaled spectral data. (c) Clutter features detected by NIMA. (d) Atmospheric features detected by NIMA and their classifications. Feature numbers correspond to columns in Table 1

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Fig. 4.
Fig. 4.

Example of velocity-folded spectra from the Sha Lo Wan profiler. (a) The velocity folded 2D median-filtered spectra with a Nyquist velocity of 8.55 m s−1. (b) The unfolded 2D median-filtered spectra with an effective Nyquist velocity of 17.1 m s−1. Range gates below 464 m were not detected as being folded, so the spectral intensities at the additional Doppler velocities are set to the noise level

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Fig. 5.
Fig. 5.

Example of Doppler peak membership threshold adjustment processing for spectra from the Lemon Creek profiler under precipitation conditions. (a) The spectra and NIMA moments. Note the strong signal and velocity folding at the melting layer around the 522-m range gate. (b) The NIMA atmospheric features at the initial Doppler peak membership threshold of 0.15. Each feature is shown in a distinct color. (c) The NIMA atmospheric features at the raised threshold of 0.36

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Fig. 6.
Fig. 6.

Examples of membership functions for RFI feature detection. Membership characteristic values that lie outside the ranges shown receive the membership value associated with the closest value within the range shown. Note that the membership value is positive for an RFI feature; zero or negative for non-RFI features

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Fig. 7.
Fig. 7.

Portion of example spectrum from Fig. 1 at 440 m. Dashed line is the 2D median-filtered spectrum, dotted line is the noise floor, shaded areas are the spectral cutoff regions, solid line shows the Gaussian fit. The Gaussian-estimated midpoint x0 and the resulting first and second moments estimated from the Gaussian are indicated

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Fig. 8.
Fig. 8.

The idealized (without random noise) spectral profile used in the simulation study is based on the profile shown in Fig. 1

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Fig. 9.
Fig. 9.

Simulation results for atmospheric signal without contaminants. Average errors in the first and second moments (second moments as spectral half-width) over 1000 trials are shown for NIMA and SPA algorithms. Also shown are the theoretical signal-to-noise ratio and the average SNR as calculated by SPA

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Fig. 10.
Fig. 10.

The lowest four gates of the idealized simulated spectra including atmospheric signal and ground clutter. The asterisks indicate the true first moment (center point) and the true spectral width (outer points)

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Fig. 11.
Fig. 11.

Simulation results for atmospheric signal contaminated with clutter. Average errors in the first moments over 1000 trials are shown for NIMA and SPA algorithms. Also shown are the theoretical SNR for the atmospheric signal and the average SNR calculated by SPA based on its picked peak

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Fig. 12.
Fig. 12.

Simulated example of the effect of NIMA's RFI processing in the case of weak atmospheric signal. (a) With RFI processing enabled, NIMA selects the weaker atmospheric signal in preference to the stronger RFI signal. (b) With RFI processing disabled, NIMA selects the RFI signal

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Fig. 13.
Fig. 13.

Simulation results for atmospheric signal contaminated with clutter and RFI. Average errors in the first moments over 1000 trials are shown for NIMA and SPA algorithms. Also shown are the NIMA average first moment confidence values

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Fig. 14.
Fig. 14.

Simulation results for atmospheric signal contaminated with clutter, RFI, and point targets. Average errors in the first moments over 1000 trials are shown for NIMA and SPA algorithms. Also shown are the NIMA average first moment confidence values and the SPA average SNR calculated at each height

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Fig. 15.
Fig. 15.

Plots of first moment errors as a function of the fraction of 6000 sample spectra removed based on NIMA first moment confidence. The top x axis shows the confidence thresholds used to discard the fraction of the data indicated on the lower x axis. (a) A plot of average absolute error. (b) A plot of rms error

Citation: Journal of Atmospheric and Oceanic Technology 19, 3; 10.1175/1520-0426-19.3.274

Table 1.

RFI feature fuzzy logic example

Table 1.
Table 2.

Performance comparison for human-truthed data

Table 2.

1

U.S. Patent Number 5940523 issued 17 August 1999. Contact the author for information concerning no-cost license agreements to obtain NIMA for research and educational purposes.

Save