274 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 19

The NIMA Method for Improved Moment Estimation from Doppler Spectra
CORINNE S. MORSE
National Center for Atmospheric Research, Boulder, Colorado

ROBERT K. GOODRICH
National Center for Atmospheric Research, and Department of Mathematics, University of Colorado, Boulder, Colorado

LARRY B. CORNMAN
National Center for Atmospheric Research, Boulder, Colorado

(Manuscript received 4 April 2000, in final form 13 September 2001)

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.

1. Introduction of fuzzy logic modules, was first described by Cornman

et al. (1998). This paper gave a generalized overview
To fully utilize Doppler measurement devices for the of the fuzzy logic and image-processing techniques
real-time measurement of wind, wind shear, and tur- used, particularly as applied to the problem of distin-
bulence, data from these devices must be quality con- guishing ground clutter and atmospheric signal in the
trolled. When rapidly updated measurements are re- spectra from the Sha Lo Wan, China, profiler. NIMA
quired, such as for an airport wind hazard warning sys- was shown to have several advantages over traditional
tem, the traditional method of long-term moment av- peak picking methods. Although NIMA was developed
eraging or ‘‘consensus’’ methods (Weber et al. 1993) for the processing of profiler data and has been applied
cannot be used. Analysis of the Doppler spectra by hu- to boundary layer, tropospheric, and 50-MHz devices,
man experts provides reliable quality control (Schu- it should be noted that the method is also applicable to
mann et al. 1999) but becomes impractical when con- any device that measures spectra as a function of range,1
tinuous real-time processing is required. The National and it has been successfully applied to both Doppler
Center for Atmospheric Research (NCAR) Improved lidar and weather radar spectral data.
Moments Algorithm (NIMA) spectral processing al- Since the original publication, the NIMA algorithm
gorithm produces a set of quality-controlled moments has been further developed and a relatively mature ver-
from each beam of spectral data at a skill level ap- sion has been employed operationally for over three
proaching that of a human expert. years as part of a terrain-induced turbulence research
The NIMA algorithm, composed of an integrated set

Corresponding author address: Corinne S. Morse, National Center 1

U.S. Patent Number 5940523 issued 17 August 1999. Contact the
for Atmospheric Research, P.O. Box 3000, Boulder, CO 80307-3000. author for information concerning no-cost license agreements to ob-
E-mail: morse@ucar.edu tain NIMA for research and educational purposes.

q 2002 American Meteorological Society

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
MARCH 2002 MORSE ET AL. 275

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.

study at the Juneau International Airport (Alaska). This denoted by an asterisk and the spectral width (twice the
paper describes the overall NIMA processing with an square root of the second moment) is indicated by the
emphasis on the new components of the algorithm. length of the line segment centered at the first moment.
These include a new fuzzy logic membership combi- Figure 1a shows the 30-s average spectral data and the
nation method, an algorithmic component to handle moments calculated by POP. This data has several con-
time-varying radio frequency interference (RFI) con- taminants in addition to the atmospheric signal. The
tamination, and perhaps most significantly, estimation strong narrow signal near zero velocity and below 600
of a quality control index (confidence) in the resulting m is characteristic of ground clutter. The weak signal
moments. These confidence estimates can be used in near 8 m s 21 Doppler velocity is RFI with characteristic
subsequent processing, that is, wind and turbulence cal- constant Doppler shift and near constant intensity over
culations based on the moments, to produce more ac- all range gates. The region of localized spectral intensity
curate and reliable wind and turbulence estimates. For around 440-m range and 29 m s 21 Doppler velocity is
an operational warning system, confidence estimates can characteristic of a transient contaminant such as a bird.
reduce false alarms, that is, a hazard warning may be POP is based on locating the peak intensity of the signal
suppressed in the case of calculated moments that in- at each range gate and produces good moments for clean
dicate a wind-related hazard but have a low confidence. spectra. However, when contaminants have spectral
The details of the wind calculations, the confidence es- peak intensities greater than that of the atmospheric sig-
timates for these values, and the resulting performance nal, erroneous moments can be generated as seen in Fig.
are the subject of related papers (Goodrich et al. 2002; 1a. In traditional applications, such erroneous moments
Cohn et al. 2001). are not problematic as they are eliminated through con-
An example of the spectral moments calculated by sensus averaging over time periods of 30 min or more.
NIMA compared with those calculated by the Profiler However, to support airport operations with the require-
On-line Program (POP; Carter et al. 1995) are shown ment for rapidly updated and highly accurate winds,
in Fig. 1 for a beam of data from the Lemon Creek 915- quality control must be applied on each beam of data.
MHz profiler in Juneau, Alaska. The plots show the This was the motivation for the development of NIMA.
contoured spectral intensity, or spectral surface, as a Figure 1b shows the NIMA moments overlaid on the
function of Doppler velocity (horizontal axis) and range spectra (median-filtered along the Doppler velocity axis)
(vertical axis). The first moment at each range gate is with the associated moment confidences in the right

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
276 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 19

sidebar. Note that the NIMA moments are smooth as a possible combination rule with this property would be
function of range and that the first moment confidence to take the minimum of the membership values; that is,
is lowered for ranges with a low signal-to-noise ratio MMT (x) 5 min j [M Mj(x)] (Zimmerman 1996). This meth-
(SNR). This example dataset will be used to illustrate od does not, however, allow for the use of weights. An
some of the components of NIMA processing described alternative method that does allow for weighting is to
in section 4 and will be used as the basis of the sim- create a total membership value MGT by combining
ulation study presented in section 5. membership values M Gj using a geometric weighted
The following sections will discuss the fuzzy logic mean,
combination methodology used in NIMA and how the

[P ] Oa
1/a GT
fuzzy logic parameters can be tuned, followed by a de- MGT (x) 5 [MGj (x)] a Gj
a GT 5 Gj
scription of the components of NIMA processing in- j j
cluding confidence estimation. In the results section,
NIMA performance is compared with that of a simple 0 # MGj (x) # 1. (2)
peak picking algorithm using a simulation study and the Note that the membership values M Gj must all be pos-
performance of the confidence estimates analyzed using itive. This combination method is particularly useful for
human-truthed data. membership characteristics that have high skill primar-
ily when the membership value is zero or near zero. By
2. Enhanced membership combination assigning a small a G weight to this characteristic, its
membership value will have a small effect on the total
Fuzzy logic (Yager et al. 1987) is used in NIMA membership value if it is close to unity (and has low
because of its strength in addressing the natural ambi- skill), but will have a large impact if it is near zero (and
guities in measurement data, as well as in classification has high skill).
and pattern recognition. The basis of the fuzzy logic Each membership characteristic is assigned a com-
approach is to avoid strict thresholding until all infor- bination method, that is, either algebraic or geometric.
mation has been combined. That is, in contrast to a The membership combination paradigm used by NIMA
Boolean logic algorithm where strict thresholds for each for the total membership function M T includes both al-
test are used, fuzzy logic algorithms delay this thres- gebraic and geometric combination as
holding until the final step. This leads to a much more
robust and accurate result. To achieve this end, each of MT (x) 5 [MAT (x) a AT · MGT (x) a GT ]1/(a AT1a GT) , (3)
the disparate measurable quantities or characteristics to where MAT and MGT are the total membership values of
be used in the decision is translated into a common the algebraic and geometric characteristics, respectively,
quantity, a membership value, using a membership func- and aAT and aGT are the sums of the weights of the
tion, which is specific to that characteristic. The original algebraic and geometric characteristics, respectively.
paper described the fuzzy logic combination of mem- Typically each of the NIMA fuzzy modules use some
bership values at each point x in the radial velocity- combination of algebraic and geometric characteristics.
range coordinate space of the spectral surface illustrated After the fuzzy combination step is the final defuz-
in Fig. 1. In that implementation, individual membership zification step, where a decision is made by comparing
values M Ai are combined using the center of gravity the total membership value M T (x) against the member-
(Zimmerman 1996) method to create a total membership ship threshold value. If M T (x) exceeds the threshold then
value MAT ; that is, an algebraic weighted mean is cal- the data at x is a member otherwise, it is not a member.
culated Section 4g gives a detailed example of the application

[O Oa
of membership functions and combinations in NIMA.
MAT (x) 5
i ]@
a Ai MAi (x) a AT

21 # MAi (x) # 1.
a AT 5
i
Ai

(1)
3. NIMA configuration and tuning
Spectral characteristics can vary from site to site de-
If a particular membership characteristic, i 5 p, shows pending on local conditions including topography and
more skill than others, a larger weight a Ap can be as- the radar-processing parameters used. Thus it is desir-
signed in Eq. (1). In this way, the relative skill of the able to be able to modify NIMA processing to optimize
membership characteristics can be taken into account. performance at a given site. For this reason, the pro-
In some cases the skill of the membership characteristic cessing in each of the NIMA fuzzy logic modules is
is dependent upon its associated membership value, for controlled by a configuration file. This configuration file
example, a membership characteristic that has high skill defines the membership function and the a A or a G
when its membership value is negative, but low skill weight associated with each membership characteristic
when the membership value is positive. It would be as well as the final membership threshold. The offline
desirable to combine membership values of this type in adjustment of the threshold, weights, and membership
such a way that the total membership would be low if functions to improve performance is known as tuning
any such individual membership value were low. One and generally proceeds as follows. The initial generation

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
MARCH 2002 MORSE ET AL. 277

FIG. 2. This chart shows the primary external input and output data and important intermediate data. Each circle, a–j, indicates processing
that is described in section 4.

of a membership function for a membership character- at each point of the spectral surface to determine which
istic is performed based on intuition and a sample of are likely to be associated with clutter or atmospheric
empirical evidence after examining ‘‘typical’’ spectra. signal. Contiguous points on the spectral surface that
The membership function is then adjusted based on exceed the membership threshold are grouped together
viewing how well the calculated membership values to form a feature. The fuzzy logic module for classi-
compare with desired results for a larger and more var- fying clutter results in a set of clutter features. The fuzzy
ied set of spectra. The weights are adjusted based on logic module for classifying atmospheric signal results
how well the membership values compare with the de- in a set of candidate atmospheric features. Another fuzzy
sired result relative to each other. Once an initial set of logic module scores each of these candidate atmospheric
membership functions and weights is chosen, a large features. The ‘‘best’’ atmospheric feature is used as a
number of spectra are processed and examined for cases guide to which spectral points should be used in cal-
where the resulting wind values look suspect, indicating
culating the moments. Figure 2 shows the overall flow
the possibility of incorrectly calculated moments. These
of NIMA processing as currently implemented and as
cases are then examined in detail to determine which
processing module(s), if any, produced errors. The described in the following sections.
membership functions and weights are then modified The fuzzy logic approach used in NIMA is not limited
until the error is mitigated. The large spectral sample to applications to points on the spectral surface. It can
is then reevaluated to ensure that no new errors were be applied to spectra at individual range gates, for ex-
introduced by the modifications. ample, to determine whether velocity folding is occur-
ring. It can also be applied to a set of generated features,
for example, to identify atmospheric or RFI features
4. NIMA processing overview based on feature characteristics. A detailed example of
The original paper described the NIMA processing the fuzzy logic processing as applied to RFI feature
philosophy and discussed the application of fuzzy logic identification is given in section 4g.

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
278 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 19

a. Noise calculation tral intensities near the positive and negative Nyquist
velocities relative to each other and relative to the noise
The method of Hilderbrand and Sekhon (1974) is used level. Intensities that are above the noise level and close
to calculate the noise level and its variance at each range in value are indicative of velocity folding. When ve-
gate of the Doppler spectra. The method also generates locity folding is detected, the spectrum is unfolded by
two statistical parameters that are saved and used in effectively doubling the Nyquist velocity and the num-
computing the moment confidence estimate. If the pa- ber of velocity bins and copying the spectral data to
rameters indicate that there was a problem in determin- satisfy the boundary conditions. Figure 4a shows a pro-
ing the noise level from the spectra, confidence in the file of 2D median-filtered spectra with velocity folding
resulting moments will be decreased. For example, if above 464 m. Because the spectral unfolding introduces
the Hilderbrand and Sekhon chi-square test fails, im- some ambiguity itself, for example, the two strong sig-
plying that the noise distribution is not as expected, then nals that appear in Fig. 4b, unfolding is performed only
the confidence value is lowered. at range gates where there is evidence to suggest that
the spectral data are folded. The decision to unfold the
b. Median filtering spectra is made on a per range basis. Range gates that
are not unfolded use the noise level to fill the spectral
At several steps in the analysis, the spectra are median image region outside their Nyquist velocity. In Fig. 4b,
filtered. Prior to performing the 2D surface analysis de- the range gates below 464 m were not unfolded, re-
scribed in section 4d, outliers are removed from the sulting in less ambiguity as only one of the resulting
averaged spectra by using a 2D (range and Doppler signal images covers the entire range.
velocity) median filter. Filtering results in a more con-
sistent and useful local least squares fit; that is, it high-
lights the larger-scale features of the spectral intensity d. 2D surface analysis
surface rather than the higher-frequency noise on the After unfolding the spectra, where necessary, the re-
surface. The velocity and range halfwidths of the median sulting spectral surface is subjected to the 2D analysis
filter are configurable in order to accommodate profilers described in Cornman et al. (1998). That is, local 2D
with differing Doppler velocity and range resolution. quadratic least squares fits to the spectra are employed
Figure 3a shows the spectra from Fig. 1a after the ap- to determine the curvature and gradient. These values
plication of the 2D median filter using halfwidths of two can then be used as membership characteristics for de-
velocity bins and three range gates. Note that small point termining the location of clutter and atmospheric signal.
target contaminants are effectively removed by the filter. The analysis is performed on several scalings of the
Prior to performing the Gaussian fits to spectra de- spectral intensity: linear, dB, range-normalized linear,
scribed in section 4h, 1D median filtering of the aver- and range-normalized dB. In the normalized scalings,
aged spectra along the velocity axis is performed to the linear or dB scaled data are normalized at each range
improve the performance of the fit. Although these 1D to a value between zero and one (one being the maxi-
median-filtered spectra, as shown in Fig. 1b for the ex- mum value for that range). This minimizes the effects
ample spectra, have given reasonable results, some in of reduced intensity at the higher range gates as illus-
the profiler community have expressed a desire for this trated by the contrast between the dB-scaled spectral
choice to be configurable. NIMA can be configured to surface in Fig. 3a and the range-normalized dB-scaled
use 1D median-filtered, 2D median-filtered, or unfiltered surface shown in Fig. 3b.
spectra for the final moments calculation. The unfiltered
spectra can be used if there is particular interest in de-
tecting small-scale shears along range. Conversely, the e. Detecting ground clutter features
2D median-filtered spectra, which are smoothed in range The detection of ground clutter was discussed in some
as well as in velocity, can be used if there is a large detail in the original paper. The clutter detection algo-
penalty for false reports of shear or if the desire is to rithm consists of a fuzzy logic module to determine
find shear over a larger scale. spectral points with high ground clutter membership,
followed by global image processing to combine con-
c. Velocity folding tiguous points with high membership values into distinct
clutter features. Figure 3c shows the clutter features
Ideally, the Nyquist velocity is chosen large enough found in the example spectra. The locations of the clutter
during data collection to ensure that no radial velocities features are saved for use in later processing.
will exceed it. However, even if the Nyquist velocity is In the original paper, the curvature and gradient char-
exceeded and the observed velocity is aliased or folded acteristics of the ground clutter spectral signal along
at some ranges, NIMA can determine the correct mo- with the expectation that clutter should be symmetric
ments. NIMA detects the presence of velocity folding around zero velocity were used to distinguish it from
at a given range gate using a fuzzy logic module that the atmospheric signal. The current implementation of
considers membership characteristics such as the spec- NIMA has added some additional characteristics based

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
MARCH 2002 MORSE ET AL. 279

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.

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
280 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 19

TABLE 1. RFI feature fuzzy logic example.

Feature 1 Feature 2 Feature 3 Feature 4

C M C M C M C M
Parameter aA aG value value value value value value value value
a. Feature width (m s21 ) 0.15 2.088 20.09 1.25 0.94 0.729 0.88 1.268 0.92
b. Width variance (m 2
s22 ) 0.08 0.395 0.61 0.000 1.00 0.119 0.88 0.042 0.96
c. Width variance per
range gate (m 2 s22
gate21 ) 0.35 0.015 0.79 0.000 1.00 0.040 0.35 0.001 1.00
d. Log intensity variance 0.08 6.49 21.00 1.49 0.30 2.23 1.00 3.34 0.97
e. Midpoint variance (m 2
s22 ) 0.12 0.134 0.79 0.000 1.00 0.030 0.95 0.006 0.99
f. Midpoint variance per
range gate (m 2 s22
gate21 3 10 3 ) 0.08 4.8 0.84 0.0 1.00 1.0 0.67 0.2 1.00
g. Midpoint slope (m s21
gate21 ) 0.14 20.036 0.63 0.000 1.00 0.16 20.56 0.001 0.99
h. Range gates covered 0.28 28 1.00 2 0.00 3 0.00 36 1.00
MAT 0.482 0.935 0.494 0.980
MGT 1.000 0.000 0.000 1.000
MT 0.565 0.000 0.000 0.984

on the expectation that the clutter signal has a narrow sulting from velocity unfolding. This is done by in-
Gaussian-like shape and/or corresponds to a user-con- creasing the membership of points in the vicinity of the
figurable intensity template. (The Gaussian fit analysis Doppler velocity of those moments. Clutter features de-
is discussed further in section 4h.) When clutter and tected in the spectra are used as a membership char-
atmospheric signal are expected to appear in the same acteristic to explicitly downgrade atmospheric signal
Doppler velocity bins, based on the calculated moments membership in regions where ground clutter was de-
from previous beams in the same direction, total clutter tected.
membership can be downgraded for those bins by gen- The membership threshold used to extract the at-
erating a negative clutter membership value for velocity mospheric features is typically set low enough to ensure
bins near those first moment velocities. that a mostly continuous atmospheric feature is ob-
The specific characteristics of clutter signals appear tained, even under low SNR conditions. In the presence
to vary widely between sites and typically a significant of an unusually strong signal, as illustrated by the pre-
portion of a tuning effort is focused on this fuzzy mod- cipitation spectra shown in Fig. 5a, this threshold may
ule. For example, in applying NIMA to data from a be too low to adequately separate the atmospheric fea-
profiler observing sea clutter in one of its beams, the ture, as shown in Fig. 5b (around range 500 m). An
membership function for the location characteristic was additional fuzzy logic module is used to determine
modified to detect the sea clutter signal that was con- whether the membership threshold should be raised.
sistently located slightly off zero at 0.5 m s 21 Doppler This module considers membership characteristics of
velocity. the largest feature extracted. These characteristics in-
clude the average width of the feature and the existence
of regions within the feature that did not exceed the
f. Detecting atmospheric features
threshold. If this test is positive, the set of features is
The detection of atmospheric signals proceeds simi- discarded and a new set is created using a higher thresh-
larly to that of the detection of ground clutter signal. A old. This process is repeated until the feature charac-
fuzzy logic module to determine spectral points with teristics indicate that the threshold used was reasonable.
high ‘‘Doppler peak’’ membership is followed by global Figure 5c illustrates the separated features generated
image processing to combine contiguous points with when the higher threshold is applied.
high membership values into distinct features. As de- After a set of well-separated candidate features are
scribed previously, the characteristics from the 2D sur- generated (illustrated in Fig. 3d for the example spectra
face analysis are combined with characteristics based of Fig. 1), they are then analyzed to identify which are
on the spectral intensity, for example, the ratio of spec- most likely to correspond to atmospheric signal and
tral intensity to the noise level or the range-normalized which may correspond to RFI. Characteristics consid-
spectral intensity, to determine locations of atmospheric ered for atmospheric features are the sum of the at-
signal. Temporal history, that is, moments calculated mospheric membership of the points included in the
from the previous beams in the same and opposite di- feature, the variance of the peak spectral intensity over
rections, is used to resolve potential ambiguities re- range, the feature average width, and the feature average

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
MARCH 2002 MORSE ET AL. 281

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 21 . (b) The unfolded 2D median-filtered spectra with an effective Nyquist velocity of 17.1 m s 21 . 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.

SNR. Also considered are the proximity of the feature In this dataset, the RFI appeared at a constant Doppler
to the Nyquist velocity, predicted RFI, and the signal velocity over range and time, and the RFI signals were
from the previous beam in the same direction. The fea- easily characterized and identified. Application of
ture with the highest atmospheric feature membership NIMA to the data from the Juneau profilers showed that
is selected as the ‘‘best’’ candidate for the atmospheric this characterization was overly restrictive and the orig-
signal. inal RFI algorithm was evolved into a more sophisti-
In the event that this best feature does not cover all cated fuzzy logic module. This fuzzy logic module an-
the range gates, two strategies are employed to attempt alyzes various characteristics of the atmospheric fea-
to remedy the situation. First, a fuzzy logic module is tures to determine whether they should be classified as
used to analyze the other features, which exceed the RFI. RFI features in a single beam tend to be narrow
atmospheric feature membership threshold to determine and constant over range in width, intensity, and Doppler
which, if any, could be connected with the existing best velocity. RFI may shift in Doppler velocity from beam
feature to extend the feature over additional range gates. to beam over time. If the RFI is observed consistently
Membership characteristics for these disjoint (not con- from beam to beam, this movement can be tracked and
nected) features include the difference in Doppler ve- subsequently predicted. If an RFI feature does not cover
locity between it and the best feature as well as the all the ranges, a fuzzy logic module is applied to identify
continuity in range of a combined feature. In Fig. 3d, disjoint features, which may be associated with the RFI
feature #2 is a disjoint feature that can be connected to signal. This processing is the same as that used for de-
feature #1. Second, a strategy of lowering the atmo- tecting disjoint atmospheric features, except that the
spheric membership threshold in the vicinity of the fea- membership functions and relative weights are tuned to
ture top or bottom edges is used to attempt to extend the more specific characteristics of RFI features. Like
the feature into the missing ranges. the ground clutter features, the location of an identified
RFI feature is saved for use in the subsequent calculation
g. Radio frequency interference (RFI) of spectral moments.
RFI contamination was observed during the analysis The following detailed explanation of the RFI feature
of the Point Loma dataset described in the original paper. identification fuzzy logic module is given to illustrate

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
282 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 19

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.

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
MARCH 2002 MORSE ET AL. 283

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 val-
ues (M) in columns C and M, respectively, in Table 1,
are shown in Fig. 6. The RFI identification module in-
cludes both algebraic and geometric membership char-
acteristics. The algebraic characteristics are associated
with measurable properties of an RFI feature and have
membership values in the range [21, 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 fea-
ture 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 sim-
plicity, 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 charac-
teristic 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 func-
tions for these characteristics are shown in Figs. 6a–c.
The true RFI feature 4 (Fig. 3d) shows high RFI mem-
bership values, M, for each of these characteristics (Ta-
ble 1). Membership based on width variance drops con-
siderably 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 char-
acteristic 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; mem-
bership 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 FIG. 6. Examples of membership functions for RFI feature detec-
tion. Membership characteristic values that lie outside the ranges
range. The variance of the peak signal intensity over shown receive the membership value associated with the closest value
the range gates within the feature is calculated and the within the range shown. Note that the membership value is positive
RFI membership function is shown in Fig. 6d. The at- for an RFI feature; zero or negative for non-RFI features.

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
284 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 19

mospheric feature 1 has a very negative membership ments calculation, as specified in the configuration file.
value because of its highly varying intensity. The atmospheric feature can be considered as providing
In some instances, RFI may be observed consistently a ‘‘first guess’’ as to the range of Doppler velocities
at a specific Doppler velocity, or it may change position within which lies the desired atmospheric signal. The
over time. The proximity of an atmospheric feature to velocities that mark the limits of this signal are referred
the expected position of RFI is used as an algebraic to as the spectral cutoffs. For clean spectra, the spectral
membership characteristic in the identification of RFI cutoffs are simply those points where the signal power
features. In order to determine this expected position, drops to the noise level on either side of the signal peak
the average Doppler velocity of an identified RFI signal for a given range gate, that is, the ‘‘standard’’ method.
is tracked over time and successive beams using the This simple method may produce erroneous moments
adaptive polynomial technique of Morel and Passi if there are overlapping signals, for example, overlap-
(1986). In this computationally efficient technique, a ping atmospheric signal and ground clutter. For over-
quadratic model is used for the RFI position after ac- lapping signals that have distinct peaks, the desired
counting for velocity folding. The adaptive nature of spectral cutoff is at the local minimum between the two
the algorithm allows it to track both stationary and non- signal peaks. The fuzzy logic module for spectral cutoff
stationary RFI signals. The model is used to predict the determination thus utilizes membership characteristics
expected RFI Doppler velocity position in the current of the spectral surface, which can identify that local
beam. After the initial detection of RFI, the uncertainty minimum, for example, curvature and slopes from the
in this prediction is large, but tends to decrease as more local 2D fits. This process results in the identification
data are included in the model coefficient calculations. of points on the spectral surface whose ‘‘spectral cut-
After a specified number of beams, for example, con- off’’ total membership value exceeds the spectral cutoff
figured by default as 5, in which the RFI signal is not membership threshold. As shown in Fig. 7, these points
observed, tracking and position prediction are discon- tend to be grouped together in spectral cutoff regions.
tinued and removed from consideration in detecting RFI Because the entire spectral surface is analyzed, many
features in subsequent beams. such spectral cutoff regions may be generated, but only
Two geometric membership characteristics are in- those that are adjacent to the atmospheric signal are
cluded with the goal of suppressing the classification of actually used in the moment determination.
a feature as RFI under specific conditions. The first such The spectral cutoff regions are used in conjunction
characteristic is the number of range gates included in with the atmospheric feature and the location of clutter
the feature. If there are too few, the width, midpoint, and RFI features to select the spectral points that are
and intensity variances may be artificially low and could used to calculate the moments at each range gate. In the
result in an erroneous classification. The membership absence of contaminants, the spectral data from the peak
function for this characteristic is shown in Fig. 6h. Both to the spectral cutoffs can be used directly for this cal-
features 2 and 3 get a zero membership value for this culation. However, when the atmospheric signal over-
characteristic keeping them from being classified as RFI. laps clutter, RFI, or other contaminants, the resulting
The other geometric characteristic is the average (over moments are likely to be biased if contaminated points
feature range gates) of the difference between the feature are included and also if they are simply excluded in the
midpoint Doppler velocity and the expected atmospheric calculation. To minimize this bias, NIMA replaces the
first moment. The expected first moment is predicted contaminated data by an estimate of the uncontaminated
based on persistence from the previous beam in the same atmospheric spectral intensity. These estimates are made
direction. RFI membership is low near the predicted first using the expectation that the atmospheric signal above
moment and is high away from it. This membership the noise should be Gaussian in shape,
characteristic can protect the atmospheric signal from
s(x) 5 s 0 e [2(x2x ) 2]/2s 2 , (4)
being classified as RFI if there is significant overlap of
the two signals. Features whose total RFI membership where s is the spectral intensity less the noise level
value exceeds the membership threshold are classified intensity, x is the Doppler velocity, x and s 2 are the first
as RFI. In the example in Table 1, feature 4 is classified and second Doppler moments, respectively.
as RFI based on its total membership value of 0.984, At each range gate, the atmospheric feature is used
which exceeds the 0.90 threshold value. 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 incremen-
h. Calculating Doppler moments
tally expanding the number of points included around
After the best candidate atmospheric feature is se- that peak. The confidence in each of these fits is cal-
lected, the portion of the Doppler spectra associated with culated as described below and the fit with the highest
this feature is used to determine the data from which confidence is selected as the Gaussian model. If points
the zeroth, first, and second moments are calculated. away from the peak are contaminated by noise or clutter,
Either the 2D median-filtered, the 1D median-filtered, the confidence in the fit will tend to be lowered, thus a
or the unfiltered spectral data can be used for the mo- fit containing these contaminants will tend not to be

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
MARCH 2002 MORSE ET AL. 285

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 x 0 and the resulting first and second
moments estimated from the Gaussian are indicated.

selected. To minimize noise contamination, points that A confidence in the Gaussian model is determined by
are less than one noise sigma above the noise floor are a fuzzy logic module using characteristics including the
excluded from the fits. Spectral points in the cutoff re- chi-square probability that the data used fit the Gaussian
gions as well as those within the clutter features or RFI model (Freund 1992). This is the probability that chi-
features are also excluded from the fits. The spectral square, if the model were correct, would be larger than
intensities of these excluded points are replaced with the value calculated for the data used. Other character-
those predicted by the Gaussian model before calculat- istics used in determining confidence in the Gaussian
ing the moments by the standard method (van de Kamp model are the number of points used in the fit, the con-
1988) using the linearly scaled spectral data. Figure 7 dition number of the SVD, the root-mean-square (rms)
shows a portion of the example spectral curve at 440 difference between the data and the model, the differ-
m overlaid with the position of the spectral cutoff re- ence between the estimated midpoint x 0 and the first
gions, the Gaussian curve fit, and the resulting moments. moment estimated from the Gaussian model x and the
The Gaussian model is applied by first applying a dB coefficient of the quadratic term C, which must be neg-
scaling, 10 log10 (s), to the spectral data and expanding ative for the Gaussian width s to be real. The confidence
around the spectral midpoint velocity x 0 . This reduces in the Gaussian model fit to the spectral data is used
the exponential form of Eq. (4) to the quadratic rela- later in assessing the confidence in the calculated mo-
tionship, ments.
10 log10 (s) 5 A 1 B(x 2 x 0 ) 1 C(x 2 x 0 ) 2
where i. Continuity of moments
A 5 10 log10 (s) 1 C(x 0 2 x ) 2
After the first and second moments have been cal-
culated at each range gate, they are evaluated with re-
B 5 2C(x 0 2 x ) C 5 210 log10 (e)/2s . 2
(5)
spect to the assumption that the atmosphere is contin-
Although fitting the Gaussian model in the linear spec- uous and there should be no sharp discontinuities in the
tral field would provide for more accuracy (May et al. Doppler moments as a function of range. Their conti-
1989), the logarithmic scaling is used for computa- nuity is tested using characteristics including the chi-
tional simplicity. A linear least squares quadratic fit to square probabilities from local linear and quadratic fits
the dB-scaled spectra data is performed to solve for A, of the moments as a function of range. By ‘‘local’’ it
B, and C. The singular value decomposition (SVD) is meant over 6n range gates, centered at the given
(Press et al. 1988) method is used here because it is range. The value of n is typically 2, but is configurable.
robust and provides a measure of how well-behaved Fits are calculated centered at each range gate where
the system of equations is, that is, the condition num- there are sufficient surrounding range gates to satisfy
ber. Gaussian parameter estimates for the first and sec- the 6n requirement. A moment is judged based on all
ond moments, x and s 2 , respectively, are then calcu- the fits in which it is included; that is, 2n 1 1 fits except
lated using Eq. (6): for the top and bottom n range gates. For each moment
B 210 log10 (e) a ‘‘badness of fit’’ total membership value is calculated
x 5 x0 2 s2 5 . (6)
2C 2C and the moment is classified as ‘‘bad’’ if that member-

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
286 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 19

ship value exceeds the membership threshold. Missing when the Gaussian model is good, the moments esti-
moments are always classified as bad. mated from the Gaussian model and the calculated mo-
When bad first moment values are found at one or ment agree, the signal-to-noise ratio is high, the mo-
more range gates, an attempt is made to estimate those ments are continuous, and the statistical parameters from
first moments based on continuity considerations. Be- the noise level determination are as expected.
cause no one method is clearly the best, several methods Different characteristics are used for moments esti-
are used to make an estimate, including interpolation mated from continuity considerations rather than cal-
based on quadratic and linear fits to moments in sur- culated from the spectral data. Recall that this estimate
rounding range gates. In some cases, the various esti- is the median of estimates made using several models
mation methods produce widely varying results, so the to interpolate or extrapolate based on moments in nearby
median rather than an average of the results of the var- range gates. The characteristics used for estimated mo-
ious methods is used as the first moment estimate. This ments include the variance between the different esti-
estimate is now used instead of the atmospheric feature mates calculated using the different methods, the dis-
to select a new set of spectral points as described in tance in range gates from the closest calculated moment,
section 4h. The first and second moments are recalcu- and the continuity badness score of the estimate. For
lated using this new set of spectral points. The Gaussian both calculated and estimated moments, the confidence
model fit to these points and its associated moments is lowered if the moment is near the predicted position
estimate and confidence are also calculated. for RFI signal or near ground clutter, or if the feature
Using the ‘‘good’’ first moments and the recalculated upon which the moments were based was overly wide
moments, continuity is retested and a new set of bad in the velocity dimension.
moments may be identified. If there are any bad mo- The moments confidence for the example spectra are
ments remaining, the first moment is reestimated using shown in Fig. 1b. Note that the confidence is lowered
the continuity considerations and this estimated value
in the region of weak signal, 1100–1265 m, and at the
is reported as the first moment. Using this final set of
range gates, 1595–1925 m, where the moments were
first moments, the continuity is tested again. The re-
estimated rather than calculated.
sulting 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 5. Results
NIMA moments for the example spectra are shown in
Fig. 1b. The first and second moments for the 1540– Two methods are used to demonstrate NIMA’s per-
1925-m range gates were estimated values. formance in spectral moment and moment confidence
estimation. The first is to perform a simulation where
j. Moment confidence exact truth is known. The computed moments can be
compared with the true values. The second is to compare
There are cases where even a human expert has trou- computed moments with those determined by human
ble locating all or part of the atmospheric signal region. experts. This requires the assumption that the human
Such cases can include data with a low SNR, or when
experts’ moments may be used as a true value. The
ground clutter or RFI features are located near the at-
advantage of this second approach is that the ambiguity
mospheric signal. A human expert may have lower con-
of real data may be studied. It will be shown using both
fidence in the moment estimates when the atmospheric
methods that NIMA produces first moments with small-
signal has discontinuities in range or when multiple fea-
tures are present. A confidence index has been devel- er errors than those of a peak-picking algorithm. In order
oped to mimic the methods that a human expert might to use the moment confidence estimates to identify poor
use in estimating the quality of the data. A confidence quality moments, it is necessary to relate the confidence
estimate for the calculated moments is useful for down- value to the measurement error. It will be shown using
stream processing, that is, wind, wind shear, or turbu- both methods that most of the large first moment errors
lence estimation. Confidence values are in the range of are assigned a low confidence by NIMA. To be useful,
zero to one, zero indicating no confidence and one in- a confidence value should decrease with increasing dif-
dicating complete confidence in the moment value. ference between the calculated moment and the true
The confidence in the moments is estimated based on moment (as determined by a human expert or simula-
a variety of characteristics generated during the NIMA tion). A practical way to determine whether this rela-
processing. The characteristics used for calculated mo- tionship holds is to consider the average difference for
ments include the confidence in the Gaussian model of a set of calculated moments each with a confidence val-
the estimated atmospheric signal region, the difference ue that exceeds some confidence threshold. That average
between the calculated moment (van de Kamp 1988), difference should decrease as the confidence threshold
and the moment estimated from that Gaussian model, is increased. This property of the NIMA first moment
the signal-to-noise ratio, the continuity badness score, confidence will be demonstrated using the human-
and the noise statistical measures. Confidence is high truthed data.

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
MARCH 2002 MORSE ET AL. 287

a. Simulated data ber of spectral averages used in the Juneau system) were
averaged to generate one simulated profile. NIMA was
1) METHODOLOGY
applied to the profile, and the first moments and first
Zrnić (1975) suggests that a power spectrum may be moment confidences were calculated and the NIMA first
simulated by using an exponential distribution with moments were compared to the true first moments given
mean S k 1 N where S k is the idealized power spectrum by the atmospheric model. A thousand realizations were
to be simulated in spectral bin k, and N is the noise generated and at each range gate, the average error and
floor. The use of the exponential distribution is justified its variance were calculated as were the average absolute
by Zrnić (1980) in the case of an atmospheric signal. error, its variance, and the average confidence from
The assumptions are that I and Q data measured by a NIMA.
profiler are independent zero mean Gaussian processes If a comparison of NIMA and POP is desired, an
with identical variances. Here I and Q are assumed in- algorithm similar to POP must be implemented. The
dependent. The times series for I (or Q) is allowed to simple peak algorithm (SPA) finds the peak signal at
be correlated with itself. This allows for windowing each range gate and integrates to the noise floor to find
effects in the times series processing. Under these as- the moments. This is similar to what was done in a
sumptions, the Doppler spectrum will have an expo- simulation by May and Strauch (1998). This algorithm
nential distribution. May and Strauch (1998) suggest is different from POP as no I and Q processing is done
modeling ground clutter in a similar fashion. This is in the simulation and no attempt has been made to im-
based on the observation that ground clutter has similar plement the POP ground clutter rejection algorithm. The
statistics to the atmospheric signal. To include RFI and Radian system performs some I and Q averaging, a win-
point targets, a similar methodology is assumed. That dow function is used in converting from I and Q to
is, each of these signals are assumed to be independent Doppler spectra, and a notch filter is applied to reduce
zero mean Gaussian processes, and the I and Q signals the effects of ground clutter at certain range gates. This
are the sum of the resulting I and Q values for each of I and Q processing gives spectra similar to the reali-
the corresponding contaminants plus the noise plus the zations produced here, thus SPA should have similar
atmospheric signal. This is based on the fact that the performance characteristics to that of POP. SPA is also
Doppler spectrum of RFI has an appearance similar to run on each simulated profile, and comparisons are made
a narrow atmospheric signal, and this similarity was the to the true first moment values. Although errors were
reason a special module had to be created in NIMA to calculated for both first and second moments, the dis-
distinguish RFI from the atmospheric signal. Other cussion emphasizes the first moment performance as the
models are possible such as a deterministic model for second moment analysis is more problematic. Second
RFI, but treating the RFI in a way similar to atmospheric moment error comparisons are meaningful only when
signal gives a simple way to generate realizations that the first moments are correctly identified and in the real
have similar statistics to atmospheric signals. Under data there is greater ambiguity and human expert dis-
these assumptions, it follows by the analysis in Zrnić agreement regarding the true second moment values be-
(1980) that an appropriate model will be to make S k the cause many of the spectra are not Gaussian.
sum of several Gaussian models, one each for atmo-
sphere, ground clutter, RFI, and point targets.
2) ATMOSPHERIC SIGNAL
The model for the simulation is the data in Fig. 1.
The stacked spectra from this example were examined The first simulation includes only the atmospheric
by eye to determine the approximate mean m, half-width signal, that is, without ground clutter, RFI, or point tar-
s, and peak intensity S 0 , of the component signals at gets. Figure 9 presents the NIMA and SPA first and
each range gate. Models of the form second moment errors as a function of range and the
associated theoretical SNR values from the idealized
Sk 5 S 0 e [2(xk2m) 2 ]/2s 2 (7)
spectra compared with average SNR values calculated
were produced for each component signal. The final S k by SPA and NIMA. NIMA and SPA both perform well
was a linear combination of these component models. on signals with a high SNR when no other contaminants
The noise N was set based on noise estimates from the are present. However, notice that at ranges between 1500
example spectra. The theoretical SNR was calculated as and 2000 m, where the SNR is low, SPA has much
[(2p)1/2S 0 s]/(2NVN), where S 0 is the amplitude of the larger first moment errors than NIMA. The variances of
atmospheric signal amplitude and VN is the Nyquist ve- the SPA estimates ranged from 20 m 2 s 22 to 110 m 2 s 22
locity. Figure 8 shows the idealized simulated spectral in these ranges. The NIMA confidences were low in
surface and the true moments of the atmospheric signal. these ranges (all were less than 0.4, where 0.67 is con-
At each range gate and each bin an exponential random sidered good), and the variances were all smaller than
number generator was simulated by calculating (S k 1 0.5 m 2 s 22 , indicating NIMA’s superior results in low
N) ln(R[0, 1]) where R[0, 1] is a uniformly distributed SNR conditions. If all data with true SNR values smaller
random variable on the interval [0, 1]. This produced than 210 dB are removed, then both SPA and NIMA
one set of unaveraged spectra. Fifty such sets (the num- have average absolute errors smaller than 0.1 m s 21 .

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
288 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 19

FIG. 8. The idealized (without random noise) spectral profile used in the simulation study is based on the profile shown in
Fig. 1.

For low SNR signals, NIMA tends to track the signal Fig. 11 with those in Fig. 9 shows that for these ranges
better than SPA, and furthermore low first moment con- both SPA and NIMA show increased errors. The average
fidence is given to these moments. Figure 9 also shows bias for SPA in these ranges is 1.44 m s 21 , while the
that the second moment errors (measured as errors in average bias for NIMA is 0.13 m s 21 . This shows that
the spectral half-width) are comparable for NIMA and NIMA reduces the bias due to the clutter when the at-
SPA for ranges where the first moment errors are small. mospheric signal is well separated from or only partially
Both algorithms tend to underestimate the second mo- overlaps the clutter. This is because of the way NIMA
ments. 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 in-
3) GROUND CLUTTER
tegrates to the noise floor thus including the clutter in
The second simulation involves the addition of the calculation of the first moment. SPA also selects the
ground clutter to the atmospheric signal. Figure 10 ground clutter over the atmosphere where the clutter has
shows the stacked spectra from the lowest four ranges a higher peak than the atmospheric signal, that is, 1155
where the atmospheric signal partially overlaps with the m. Note that in the presence of clutter, the SPA-cal-
ground clutter. Comparing the performance results in culated SNR is no longer well correlated with the mag-

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
MARCH 2002 MORSE ET AL. 289

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.

nitude 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 re-
jection 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 mit-
igation 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 mo-
ment estimate could be larger depending on the amount
of overlap between the atmospheric signal and the clut-
FIG. 10. The lowest four gates of the idealized simulated spectra
including atmospheric signal and ground clutter. The asterisks indi- ter, which in turn depends on the locations and second
cate the true first moment (center point) and the true spectral width moments of the two signals. Such cases do cause prob-
(outer points). lems for NIMA and this is a topic of current investi-

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
290 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 19

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.

gation. This problem has a particular impact in vertical the RFI as the atmospheric signal. SPA also selects RFI
beam estimates of the vertical wind. whenever the RFI signal has a higher peak than the
atmospheric signal. To analyze this, the atmospheric sig-
nal intensity was returned to its original value and Fig.
4) RADIO FREQUENCY INTERFERENCE
13 shows a comparison of NIMA and SPA average er-
In the next simulation, RFI was added to the ground rors in the presence of ground clutter and RFI. Notice
clutter and the atmospheric signals. The RFI problem that there are additional large errors when SPA selects
is acute in the Juneau system since two of the profilers the RFI instead of the atmospheric signal and that these
are in direct line of sight of each other. The RFI can errors are consistent with the errors in the POP moments
have a similar appearance to a narrow atmospheric sig- shown in Fig. 1a. Notice that NIMA does have some
nal. This was the motivation for adding an RFI detection average errors around 0.5 m s 21 , but the moments as-
module to NIMA. Before this module was added to sociated with these errors have a very small average
NIMA, RFI was sometimes selected as the atmospheric confidence. These errors occur at heights where the SNR
signal especially in cases where the atmospheric signal is low. All data with an average confidence above 0.4
was weak. To illustrate this point, the simulated atmo- have an average absolute error smaller than 0.3 m s 21 .
spheric signal was attenuated by a factor of 0.15 to
produce a weak signal. Figure 12a shows a realization 5) POINT TARGETS
in which NIMA selects the weak atmospheric signal in
preference to the stronger RFI signal, while Fig. 12b The final simulation adds a point target to the previous
shows NIMA with the RFI module disabled selecting signals. SPA selects this point target in the range gates

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
MARCH 2002 MORSE ET AL. 291

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.

where the point target has a higher peak than the other mitigates ground clutter biases (in cases where well-
signals (around 500 m). NIMA does not select the point defined separate atmospheric and clutter peaks are seen
target. The point target is often removed by the median in the stacked spectra) and tends not to select RFI nor
filter (see Fig. 1a vs Fig. 3a). In cases with larger point point targets. The first moment confidence may be used
targets, a feature might be formed, but such a feature to identify poor quality data. The next section will show
would receive a low score as it covers few ranges. Such that these conclusions also hold for the human-truthed
isolated targets usually do not pose much of a problem data.
for NIMA. Figure 14 compares NIMA and SPA average
errors in the case when the simulation contains the at-
mospheric signal, ground clutter, RFI, and a point target. b. Human-truthed data
Applying a confidence threshold of 0.4 to these data The version of NIMA described above was installed
results in an average absolute error of less than 0.3 m in December 1998 for use on three 915-MHz wind pro-
s 21 . Notice the large number of errors for SPA as com- filers located in Juneau, Alaska. In 1999, the 10-min
pared to the case when only an atmospheric signal was confidence-weighted average winds (Goodrich et al.
present. If contaminants are present, a peak finding al- 2002) were more reliable and consistent than those re-
gorithm will select a contaminant if the contaminant has ported during the previous year when an earlier version
a higher peak value than the atmospheric signal. If a of NIMA without the enhanced RFI processing had been
contaminant is selected by the peak finding algorithm, used. This empirical evidence is corroborated by a more
then the computed SNR will tend to be larger than the formal verification effort, described in Cohn et al.
true SNR. So although SNR can be useful in assessing (2001). In this verification study, the first moments from
data quality (Angevine et al. 1993), it can be misleading NIMA were compared to first moments estimated by
if used alone as a confidence index for a peak finding human experts. The dataset consisted of 172 profiles,
algorithm. that is, average spectra from single beams, each con-
taining 36 range gates. The profiles were selected to test
6) SUMMARY OF SIMULATION RESULTS NIMA under a variety of atmospheric and contaminant
In these realistic, though limited, simulations, NIMA conditions, with the greatest number selected to be near
tracks weak signals with less error than SPA. NIMA times when a research aircraft was flying near the pro-

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
292 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 19

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.

filer. In addition, some of the spectra were chosen at reported. There were a fair number of outliers in the
times when NIMA winds showed large temporal dis- NIMA first moments as signified by the rms error of
continuities. This was a challenging dataset containing 1.32 m s 21 . The largest outliers were caused by weak
spectra with weak signals, ground clutter, RFI, and point signal cases as well as cases where NIMA erroneously
targets. To evaluate these profiles, a graphical interface selected a part of the ground clutter as the atmospheric
was developed (Cohn et al. 2001) that allowed the hu- signal. The comparison of the POP first moments with
man expert to insert the spectral cutoffs. Estimates of the human-generated ones resulted in errors roughly
the first and second moments were then computed using three times larger than those from NIMA. The largest
a quadratic fit to the dB-scaled spectral data, Eqs. (5)– POP outliers were caused by weak signal cases, ground
(6). The human expert had the additional option of se- clutter, and cases where RFI was selected instead of the
lecting the location of the first moment if the quadratic atmospheric signal. RFI is an unavoidable problem in
fit did not give an adequate estimate. The human expert Juneau since two of the profilers are in a direct line of
also entered a confidence value for each moment. sight of each other.
This dataset of over 6000 spectra was processed by The first moment confidence values were also eval-
NIMA and POP. The first moments output from each uated. It should be expected that the average absolute
algorithm were then compared to the first moments giv- error and the rms error should become smaller as data
en by the human experts. A summary of the performance with low NIMA confidence are removed. In Fig. 15a,
of the two algorithms is given in Table 2, where the the horizontal axis represents the fraction of the data
average error, average absolute error, and rms errors are removed via confidence thresholding. The curve rep-

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
MARCH 2002 MORSE ET AL. 293

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.

resents the average absolute error as a function of the further reduce the absolute error. Thus NIMA first mo-
fraction of data removed. Thus the average absolute ment confidence has the property that increasing the
error of the data is reduced to about 0.3 m s 21 (roughly confidence threshold decreases the average absolute er-
the velocity resolution of the Juneau profilers) if the ror. A good quality control index should have this prop-
lowest 20% of the data as ranked by NIMA confidence erty.
is removed. This corresponds to using only data above Figure 15b is an analysis of the rms error of NIMA
the NIMA first moment confidence threshold of 0.676. first moments as a function of the percent of data re-
If SNR is used as a surrogate confidence for POP, nearly moved as ranked by NIMA first moment confidence.
60% of the data with lowest SNR must be removed to The entire dataset has an rms error of around 1.3 m s 21 .
obtain the same level of performance. The average ab- If 20% of the data are removed, this rms error is reduced
solute error for NIMA first moments levels off at about to around 0.75 m s 21 . This indicates some significant
0.2 m s 21 when about half of the data are removed. This outliers have been removed by applying a NIMA first
is less than the velocity resolution and so increasing the moment confidence threshold of 0.676. Similarly, if the
first moment confidence threshold does not significantly 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 21 .
There are still significant outliers because POP often
TABLE 2. Performance comparison for human-truthed data.
selected the RFI or ground clutter whose resulting SNR
Average was not low. If 30% of the data are removed based on
Average error absolute error Rms error NIMA confidence, the rms error for NIMA first mo-
Algorithm m s21 m s21 m s21
ments drops below 0.5 m s 21 . The NIMA first moment
NIMA 20.10 0.49 1.32 confidence values show skill in removing outliers, as is
POP 20.42 1.81 4.71
shown in Fig. 15b. This is further confirmed by Cohn

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
294 JOURNAL OF ATMOSPHERIC AND OCEANIC TECHNOLOGY VOLUME 19

et al. (2001) where the distribution of errors as a func-

tion 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 con-
taminants, 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 dif-
ficulty in estimating the moment. First moment esti-
mates 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 tur-
bulence and can contribute to confidence estimates for
those derived products (Goodrich et al. 2002). These
confidence estimates can be useful in improving oper-
ational 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 pro-
duce 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 es-
timates as well as turbulence estimates. Each of these FIG. 15. Plots of first moment errors as a function of the fraction
products will require a confidence value to help reduce of 6000 sample spectra removed based on NIMA first moment con-
false alarms. When the confidence value is low for a fidence. The top x axis shows the confidence thresholds used to dis-
card the fraction of the data indicated on the lower x axis. (a) A plot
product, no warning will be produced based on that of average absolute error. (b) A plot of rms error.
product. Where these products are produced from pro-
filer data, a factor in each of these confidence values
will be the first moment confidence, because a poor first for improvement, NIMA has reached a fairly stable lev-
moment confidence might indicate the derived product el. In applying NIMA to data from a new site, reasonable
quality is poor. An important first step in this process levels of performance can be obtained with minor pa-
is to obtain high-quality moment and moment confi- rameter tuning. One current topic of interest is an au-
dence values. It has been shown that the NIMA first tomated tuning process when installing NIMA at a new
moment values have excellent agreement with human- site. Recent modifications to the NIMA software infra-
produced first moments when the NIMA first moment structure now allow for different tunings when several
confidence value is high. This was also confirmed in scan strategies run concurrently on the same profiler.
the limited simulation exercise. One reviewer suggested that time continuity be used
Additional details have been given concerning NIMA in processing to connect disjoint features. This is an
processing with an emphasis on a new fuzzy logic mem- excellent suggestion and will be incorporated into the
bership combination method and a component to handle algorithm by using the moments from the previous beam
time-varying RFI contamination. RFI is a significant in the same direction as a membership characteristic.
problem in Juneau since two of the profilers are in line The first moments in cases where the overlap between
of sight to each other and it has been shown that a atmospheric signal and clutter is nearly complete also
traditional peak-finding algorithm often selects RFI in- need improvement. The effects of ground clutter on
stead of atmospheric signal. near-zero winds is especially important in the analysis
NIMA has been applied to a limited amount of data of vertical beams. This has received limited attention in
from profilers around the world. Each new site has pro- the current work because of our particular interest in
vided new challenges and opportunities to improve the high wind cases and the lack of a vertical beam in the
algorithm. Although there are still some areas targeted Juneau scan sequence. Studying and improving the per-

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC
MARCH 2002 MORSE ET AL. 295

formance of NIMA second moments is also of interest timation and aircraft measurements. J. Appl. Meteor., 40, 704–
719.
as these second moments are used in turbulence esti- Cornman, L. B., R. K. Goodrich, C. S. Morse, and W. L. Ecklund,
mation. 1998: A fuzzy logic method for improved moment estimation
Identification of precipitation is another topic of cur- from Doppler spectra. J. Atmos. Oceanic Technol., 15, 1287–
rent research. At operating frequencies where clear-air 1305.
Freund, J. E., 1992: Mathematical Statistics. 5th ed. Prentice Hall,
and precipitation signals are separate and distinct, it 658 pp.
would be desirable to be able to report moments from Goodrich, R. K., C. S. Morse, L. B. Cornman, and S. A. Cohn, 2002:
both signals; where the signals are not separable, it A horizontal wind and wind confidence algorithm for Doppler
would be useful to identify whether or not the moments wind profilers. J. Atmos. Oceanic Technol., 19, 257–273.
Hilderbrand, P. H., and R. S. Sekhon, 1974: Objective determination
are associated with a precipitation signal or not as this of the noise level in Doppler spectra. J. Appl. Meteor., 13, 808–
could affect the confidence in the resulting winds and 811.
turbulence calculations. Jordan, J. R., R. J. Lataitis, and D. A. Carter, 1997: Removing ground
and intermittent clutter contamination from wind profiler signals
using wavelet transforms. J. Atmos. Oceanic Technol., 14, 1280–
Acknowledgments. We thank Mercedes Maruri Ma- 1297.
chado of Universidad del Pais Vasco for allowing us to May, P. T., and R. G. Strauch, 1998: Reducing the effect of ground
apply NIMA to her wind profiler data containing sea clutter on wind profiler velocity measurements. J. Atmos. Oce-
clutter and Volker Lehmann of the Deutsche Wetter- anic Technol., 15, 579–586.
——, T. Sato, M. Yamamoto, S. Kato, T. Tsuda, and S. Fukao, 1989:
dienst for valuable feedback and suggestions regarding Errors in the determination of wind speed by Doppler radar. J.
NIMA processing. We would like to thank Dr. Stephen Atmos. Oceanic Technol., 6, 235–242.
Cohn for our valuable discussions on profiler technol- Morel, C., and R. Passi, 1986: Use of adaptive polynomials for real-
ogy, Dr. Rod Frehlich for our discussions on simulation, time data editing. J. Atmos. Oceanic Technol., 3, 494–498.
Press, W. H., B. P. Flannery, S. A. Tenkolsky, and W. T. Vetterling,
and also Andrew Weekley and Jeremiah Stone for their 1988: Numerical Recipes in C. Cambridge University Press, 735
contributions to the data analysis effort. This research pp.
is in response to requirements and funding by the Fed- Schumann, R. S., G. E. Taylor, F. J. Merceret, and T. L. Wilfong,
eral Aviation Administration (FAA). The views ex- 1999: Performance characteristics of the Kennedy Space Center
50-MHz Doppler radar wind profiler using the median filter/first
pressed are those of the authors and do not necessarily guess data reduction algorithm. J. Atmos. Oceanic Technol., 16,
represent the official policy or position of the FAA. 532–549.
van de Kamp, D. W., 1988: Profiler Training Manual #1: Principles
of Wind Profiler Operation. National Weather Service, 49 pp.
REFERENCES Weber, B. L., D. B. Wuertz, D. C. Welsh, and R. McPeek, 1993:
Quality controls for profiler measurements of winds and RASS
temperatures. J. Atmos. Oceanic Technol., 10, 452–464.
Angevine, W. M., S. K. Avery, and G. L. Kok, 1993: Virtual heat Yager, R. R., S. Ovchinnikov, R. M. Tong, and H. T. Nguyen, 1987:
flux measurements from a boundary-layer profiler-RASS com- Fuzzy Sets and Applications, Selected Papers by L. A. Zadeh.
pared to aircraft measurements. J. Appl. Meteor., 32, 1901–1907. John Wiley and Sons, 685 pp.
Carter, D. A., K. S. Gage, W. L. Ecklund, W. M. Angevine, P. E. Zimmerman, H. J., 1996: Fuzzy Set Theory and its Applications. 3d
Johnston, A. C. Riddle, J. Wilson, and C. R. Williams, 1995: ed. Kluwer Academic Publishers, 435 pp.
Developments in UHF lower tropospheric wind profiling at Zrnić, D. S., 1975: Simulation of weather like Doppler spectra and
NOAA’s Aeronomy Laboratory. Radio Sci., 30, 977. signals. J. Appl. Meteor., 14, 619–620.
Cohn, S. A., R. K. Goodrich, C. S. Morse, E. Karplus, S. Mueller, ——, 1980: Spectral statistics for complex colored discrete-time se-
L. B. Cornman, and R. A. Weekley, 2001: Radial velocity and quence. IEEE Trans. Acoust. Speech Signal Process., 28, 596–
wind measurement with NIMA: Comparisons with human es- 599.

Brought to you by NATIONAL ATMOSPHERIC RESEARCH LAB | Unauthenticated | Downloaded 06/12/24 11:27 AM UTC