Вы находитесь на странице: 1из 10

An Overview of Recent Advances for the TEMPER

Radar Propagation Model


G. Daniel Dockery, Ra'id S. Awadallah, David E. Freund, Jonathan Z. Gehman, Michael H. Newkirk
The Johns Hopkins University Applied Physics Laboratory
1 1100 Johns Hopkins Road
Laurel, MD 20723
Abstract - This paper is a survey of recent improvements and outlined here, as they are relevant to TEMPER developments
development efforts associated with the Tropospheric described in this paper.
Electromagnetic Parabolic Equation Routine (TEMPER). These Begin with Maxwell's Equations for a linear transmission
improvements include hybridized, multiple-method approaches medium in Cartesian coordinates. Assume a harmonic time
to terrain and grazing angle modeling, and a more robust dcq
impedance boundary. Ongoing development efforts include dependence e , where Si
a) the radial frequency of the radar.
improving the representation of ocean waves as small-scale This removes explicit time dependence from the equations.
roughness, range bias estimation based on phase, two- versus Later in this paper, an approach will be discussed for re-
three-dimensional model comparisons, and time-dependent pulse introducing linear time dependence into the model.
modeling. Next, ignore transverse (y) variations in the environment.
With this assumption and some transformation of variables,
I. INTRODUCTION the original system of equations reduces to a 2D scalar wave
equation given by (1). The 2D assumption is very accurate
A. TEMPER with respect to the atmosphere's index of refraction n; vertical
The Johns Hopkins University Applied Physics Laboratory gradients tend to be much larger than horizontal gradients.
(JHU/APL) began investigating Parabolic Equation (PE) Terrain, however, violates the 2D assumption. Comparisons of
methods for radar propagation modeling around 1980 [1]. 2D and 3D results later in this paper, however, will show this
These efforts lead to a Fourier-based, split-step algorithm PE assumption is reasonably good for moderate terrain slopes.
solver called the Electromagnetic Parabolic Equation Routine
(EMPE). EMPE was subsequently upgraded and renamed the a2 a2
Tropospheric Electromagnetic Parabolic Equation Routine + 2 +k2n2 Jf=o (1)
(TEMPER). TEMPER has been continually improved to better Ax az
model impedance boundaries, rough surfaces, and irregular
terrain. This paper summarizes the more recent improvements In (1), ij is related to the y-component of either the electric
and describes ongoing efforts to increase TEMPER's fidelity. or magnetic field (horizontal or vertical polarization,
Because many of these efforts focus on refining or amending respectively), x and z are the down-range and vertical
approximations made in deriving the PE, an overview of the coordinates, and k is the radar wavenumber.
basic PE derivation will now be presented. Putting j into paraxial form V ue'k, (1) may be factored'
=

B. The Electromagnetic Parabolic Equation into forward and backward propagating terms. If forward
Through a series of approximations summarized below, scatter dominates backscatter, as is typically the case in low-
Maxwell's four-dimensional (t,x,y,z) vector equations are elevation radar beams, one may retain only the forward term:
reduced to a two-dimensional (2D) (x,z) scalar Parabolic Wave
Equation (PWE or just PE). This is essentially an initial-value ( 2 2 _2
problem that, like the heat equation, can be solved by
"marching" a solution forward in range using either Fourier or
ox'kfn+e 2,JU=O (2)
finite-difference techniques. FFT-based PE solvers provide an
excellent mix of numerical efficiency and accuracy for Equation (2) is the EM Parabolic Equation. The only
terrestrial, low-elevation-angle radar propagation. The PE remaining complication is the square root; because n is, in
accurately models multipath, spherical-Earth diffraction, general, a function of both x and z, the square root prevents
atmospheric refraction, and ground impedance effects (e.g., separation of variables. Once separated, z derivatives can be
HF surface wave). The PE can also model terrain diffraction, handled with Fourier techniques, and the x dependence is a
terrain forward-scatter and blockage [2]. For these reasons, PE simple differential equation with exponential solutions.
solvers have been popular in radar modeling for over two Methods of approximating the radical will not be discussed
decades. References [3] and [4] chronicle the history of PE- here; it is merely noted that such approximations limit PE
based modeling in acoustics and radar. References [5] and [6]
provide a full derivation of the Electromagnetic (EM) PE. The 1 The factorization is exact when n is independent of x. For realisticx
major assumptions involved in the derivation will be merely dependence in n, the factorization is relatively accurate but does introduce a
small error [5].

1-4244-0284-0/07/$20.OO ©2007 IEEE 896


accuracy to a cone of elevation angles about the horizontal x -2
axis. The standard "narrow angle" approximation is limited to .0
roughly ±10 degrees of horizontal [3]. Various "wide angle"
approximations are available in the literature; [7] describes the I
particular wide-angle approximation used in TEMPER.5 10 5 20 5(a
PE codes compute the auxiliary function u(x,z) by 150
marching an initial source field u(0,z) forward in range.
Results are usually expressed in terms of propagation factor, 50
F, which is defined as the ratio of field strength at a position 0 5 10 15 20 25 (b)
(x,z) to the field strength that would exist at (x,z) in free space 2
with the antenna pointed in that direction. 5

It. TERRAIN AND GRAzING ANGLES


Recent TEMPER development has focused on improved 0 5 10 ranne km 15 20 25 (c)
littoral capabilities, both as a model of radar coverage and as a Figure
propagation "driver" for empirical clutter (radar backscatter) Fig. 1. Narrow sin(x)/x beam at 3 GHz over perfectly conducting wedges at
models. Key improvements include hybrid techniques for ranges of 5, 10, 15 & 20 km,. Wedge face slopes are 5, 10, 15 & 200
respectively. TEMPER results shown for a 600 problem angle, 213 FFT, Az = 5
propagation
propagation
terrain and
over terrain
over
and grazing
grazing angle
angle estimation. cm, Ar = 10 m, using the (a) KE, (b) LSM & (c) hybrid terrain methods.
A. Hybrid Terrain Methods
words, the fact that "hybrid" only models low-angle forward
TEMPER uses two methods to model propagation over scatter
Before a significant
is notmoving on, onelimitation TEMPER. detail of
additionalforimplementation
terrain. The simpler method propagates the field forward over th hyrdapoc swrhmninn.FrayP
a flat boundary then, before proceeding to the next range step,
forces field values to zero at and below the terrain height. This geometry - flat or irregular terrain - truncating the field in
method is called "knife edge" (KE) because it is analogous to altitude creates an effective reflecting boundary. Reflections
propagation over a series of perfectly conducting knife edges off this truncation boundary will contaminate the physical
whose peaks follow the terrain profile. Within the PE's solution unless the field is smoothly forced to zero before
elevation-angle limits, KE exactly matches the analytic reaching that boundary. This can be done by filtering or,
solution for single knife edge diffraction [7]. KE also agrees analogously,filters
adding a loss term that increases with height.
the field in both altitude and the Fourier
fairly well with geometric theory of diffraction (GTD) TEMPER
solutions for a perfectly conducting wedge; KE significantly (angle) domain, but does so sparingly to prevent loss of
over-predicts diffracted power in the deepest shadows behind physical energy. Not every PE step is filtered; instead, filters
are applied at range intervals designed to catch the widest
the wedge, but comes within a few dB elsewhere [8]:
TEMPER also implements *
* the more powerful Linear *r
Shift ~~~~~~~~angle
g of pro p pr aganation lust before it reflects and contaminates
method of [8] (a generalized, wide-angle version the physical solution. When computing hybrid terrain,
ls however, additional filtering may be required. Swapping
ofrrain
Beofilsand
terrain profile asTappieert'selark
piecewise-linear within eacProates
. the
each PE range step. between LSM and KE involves shifting the PE domain in
aliueInsmcss,opttononsthtadbnfr
a
Each sloped boundary is mapped to a flat domain using
vertical shifts, phase steering and surface impedance away from the truncation boundary on the LSM step will be
modifications. This method accurately models forward- shifted closer to that boundary on the KE step. Filtering prior
to ti
the KE prevents the
the step (revely) upward-shifted energy from
lowerd-trnctio
reflections and wedge diffraction for slopes less than -15' [8]. bndry.
reflecting off the (relatively) lowered truncation boundary.
Both KE and LSM
BothKE disadvantages: KE
have disdvanages
ad LS hav KE under-predicts
nderpredcts
diffraction loss and neglects forward-scatter, whereas LSM B. Multiple-Method Grazing Angle Estimation
becomes inaccurate for large slopes. Fig. 1 illustrates these Initial versions of TEMPER did not compute grazing
limitations for KE (a) and LSM (b). In particular, note the angles; ocean surfaces were approximated by a smooth, good-
field distortion caused by LSM above the 15° and 20° wedges. conductor Leontovich boundary, which is independent of
To overcome these limitations, TEMPER employs a grazing angle. Many improvements to TEMPER, however,
"hybrid" approach in which LSM is used whenever possible such as LSM, rough surface modeling, and a more general
(moderate slopes), and KE is used to get past slopes that impedance boundary condition, require knowledge of the
would be problematic for LSM. Fig. 1 (c) shows a TEMPER grazing angle at the next PE step in order to advance the PE
hybrid result: the 5° and 10° wedges were modeled using solution to that step. Furthermore, external radar clutter
LSM, whereas the 15° and 20° wedges were approximated as models usually require grazing angles in addition to
a series of knife edges. Note that, even if the forward-scattered propagation factors.
energy from the last two wedges had been properly modeled, TEMPER has three grazing angle estimation methods:
the reflection angle is too steep to matter at the low-elevation- 1. analytically compute angles based on geometry (free
angle geometries for which the PE was designed. In other space or standard atmosphere conditions);

897
2. run the PE on a scenario identical to the desired case,
but with a Leontovich boundary condition, then apply A
the spectral estimation (SE) technique described in m. or f)re
[10] to determine a dominant grazing angle at each pc

range step; or, no


3. run an internal geometric optics (GO) ray-trace code close
to estimate grazing angles at each range step. ante height
Each approach can be problematic in certain conditions. no es
Geometry only works for simple atmospheres or for large
grazing angles at close ranges. The SE method is very
powerful; however, strong refractive gradients in evaporation evap-
A 'n evap.
ducts cause SE to underestimate angles by as much as a factor GO only duct?
of two unless prohibitively small range and altitude increments
are used [11]. Furthermore, SE is very time-consuming
because it necessitates an additional PE calculation. GO Fig. 2. Logic tree for TEMPER's "automatic" grazing angle calculation.
methods handle evaporation ducts well; the accuracy of GO-
estimated angles was demonstrated via comparison with
normal mode calculations in [11]. However, GO cannot easily SE
determine grazing angles in terrain shadows. And although-GO
- Auto
GO is usually faster than SE, it still adds significant run time2 2------___-___---------------I--------__I____________
in some scenarios.
For a robust capability, TEMPER combines all three
methods into a hybridized, "automatic" grazing angle method. 0 20 40 60 80 100
In "automatic" mode, TEMPER uses a simple logic tree to (b)
determine which of the three A+ * 1>* 1>v+1>+1>
+1> A
methods 1>
should1A 1
be used
400 A on ,a l l4 , _~~~~~~~~~~~~~~~~~~~~~~~~~~~~~400I -Ocean
range-step by range-step basis. This logic is shown Fig. 2. in B 300 - .--s-land X
The "automatic" logic attempts to balance pros and cons of 200----------- --- --I - _
all three methods. For a typical calculation, the "close ranges" X.
where geometry applies corresponds to grazing angles greater H
than 1.5J. This is because even the most extreme atmospheric 0 20 40 60 80 100
refraction has minimal effect on angles > 1.50, and simple Range [kin]
geometry is sufficient. At longer ranges, SE or GO is used: Fig. 3. Illustration of how the automated grazing angle algorithm handles
is best for evaporation ducting the ocean, whereas SE mixed terrain and refractivity conditions. The resultant grazing angle profile is
GO evaporat1on duchng over tne ocean, wnereas re given by the blue line in (a); the terrain boundary is shown in (b).
1S best for
is better suited for terrain. To balance runtime and accuracy,
TEMPER avoids running both SE and GO whenever possible. On the back slope of the island, SE becomes somewhat
When both evaporation ducting and terrain are present, noisy; it is attempting to estimate a dominant diffracted angle
however, TEMPER must run both SE and GO; this is often the based on very little energy. In some cases, SE can be
case when modeling shipboard radars in the littorals. Over extremely noisy in low-power regions, however this has little
terrain, the SE angles are always selected. Over ocean, it was practical consequence as the angle errors are merely
found that taking the maximum of the two estimates, modulating a field level that is already negligibly small.
max(SE,GO), naturally selects an appropriate value. This is Behind the island, diffraction from the island tip is the
best illustrated with a simple example, shown in Fig. 3. This primary source of energy. SE correctly picks up the expected
notional scenario is a wedge "island" 40 km downrange, with 30 angle of this diffraction field, as well as the smooth
a height of 300 m and smooth faces sloped at 30, as illustrated decrease in angles that occurs as the tip becomes more distant
in Fig. 3 (b). An omni-directional 1-GHz source is radiating downrange. Eventually, evaporation duct refraction dominates
from a position 8 m above the ocean surface, and a 20-m the grazing angles again, rendering SE inaccurate. Using
evaporation duct is present on either side of the island. max(SE,GO), TEMPER catches this diffraction-to-refraction
Fig. 3 (a) shows both the SE and GO angle estimates, as transition at -75 km in Fig. 3 (a).
well as the result of TEMPER's "automatic" logic (Auto). In TEMPER's internal ray tracing code (GO) typically
the first few range steps, for grazing angles > 1.50, the "auto" launches -500 rays between +1° and -1.5° in elevation.
curve has been computed via geometry. After that, the TEMPER records the range and grazing angle each time a ray
evaporation duct begins to dominate angles. Note that SE strikes the surface. After tracing all rays, the resulting grazing
significantly underestimates these angles, hence max(SE,GO) angles are sorted in range order and then interpolated onto the
naturally selects the correct method (GO) at range steps TEMPER range grid. Interpolation is required because
leading up to the island. SE then correctly picks up the TEMPER, as well as external clutter models, requires a single-
expected ~3° grazing angle at the front of the island. valued grazing angle estimate at each TEMPER range step.

898
To do the required interpolation, a custom processing and The GAP method handles "skip zones" by interpolating
interpolation method was developed. This method, termed through regions where no rays struck the surface. While this is
greatest angle path (GAP), was designed with the following in a bold interpolation, it is adequate given that little or no radar
mind: 1) angle estimates should be relatively smooth and, 2) energy reaches the surface in these regions.
erring towards larger angles tends to be the conservative Two key points about TEMPER grazing angles are worth
approach for most TEMPER applications (e.g., external clutter reiterating. First, output angles are merely estimates, and can
models). Fig. 4 and Fig. 5 demonstrate GAP interpolation for be noisy (in the case of SE) or smoothed and interpolated (in
two ducting environments. Both figures display the grazing the case of GO); estimates tend to be good in high-power
angles of all ray surface bounces as blue dots (i.e., "raw" GO regions, but may not be accurate in low-power regions.
angles). GAP output is overlaid as a continuous red curve. Fig. Second, as a PE model, TEMPER's solution at any given
4 is GO output from the same scenario as Fig. 3, except that range represents a spectrum of propagating angles, whereas
the island has been removed and the 20-m evaporation duct is grazing angle output only provides the dominant angle. These
now capped by a 50-m bilinear/attached surface duct. Note two limitations only become severe when TEMPER's grazing
that at some ranges (e.g., 80 km) four ray paths are angle output are divorced from the corresponding propagation
intersecting the surface at four different angles. At such factor output. Thus, external models that rely on TEMPER
ranges, GAP interpolation attempts to follow the highest grazing angles should also take into account propagation
angles without creating severe discontinuities in the output. factors. The accuracy of TEMPER's solutions when using
Fig. 5 shows GAP output for a 20 m evaporation duct capped these grazing angles internally was demonstrated in [11].
by a larger, 250-m trilinear/detached surface duct. This
environment yields a very different grazing angle profile, due
to "hot spots" and "skip zones" created by a detached surface Recent and ongoing work has also improved TEMPER's
duct. boundary condition for rough, imperfectly conducting
surfaces. A major advance in the most recent release of
0.5 TEMPER, version 3.1, is a more robust impedance boundary
Raw GO Angles condition (IBC) [12]. This new implementation alleviates a
0.45 -__Interpolated Angles numerical instability that had affected TEMPER and other PE
0.4 -------------------------------- ---------------------- ----------- routines. In ongoing work, significantprogress
hasalso been
-035 L . made towards improving the small-scale surface roughness
03 portion of TEMPER's IBC. TEMPER currently uses the
0.25 Miller-Brown method [13] to model small-scale roughness
0.2 -|-------- A/---/z-t---<(e.g.,
-- 7 ocean waves). The Miller-Brown method is currently
0 0.15 ~~~~~~~~~~being
compared to stochastic averages over ensembles of
exact integral-equation calculations for a variety of sea states
0.1------------ ----------- -------------------- moreaccurately
([14] & [15]). This work will help TEMPER
0.5 Lmodel the effect of ocean waves on radar propagation.
00 ! ! ! ! !Before proceeding, it should be mentioned that TEMPER
l3 l20 40 60 80 100 120 can also ingest profiles of ocean-wave height versus range to
Range [kin] compute wave blockage, scattering and diffraction in a manner
Fig. 4. TEMPER's GAP interpolation of geometric optics results; output for a
. . duct.
50-meter attached surface .- ' ~~~~analogous to terrain calculations. However, this approach
necessitates very small range steps to sample the ocean-wave
surface, and ensemble averaging of TEMPER runs over
different realizations of an ocean wave spectrum. Thus, it is
Raw GO Angles often advantageous to represent ocean-wave effects as an
0588 gE - - ---------,-- ----- -Interpolated Angles
nt r o a e AnlIs impedance boundary on a mean-flat surface.
A. A More Stable IBC Implementation
@0.52 In solving the PE (2), a boundary condition must be
-@ lt |enforced at the surface. TEMPER uses the following IBC:

KD ~4 ----------- -----------
*,* +au(xO)=O (3)

0.42 7- In (3), the parameter a is related to the surface impedance and


Ais function of grazing angle, surface conductivity,
150l04 200
a
0 50 100 250 300 permittivity and small-scale roughness (see [11]).
Range [kin]X
Fig. 5. GAP interpolating output for a 250 meter detached surface duct. Noeta h Bsmerc ae
hra
perfectly conducting boundary condition leads to symmetry

899
about z = 0. The advantage of symmetry is that the boundary v3. 1 final testFRIe9_POPA1.OISO-SCIOO-SPIOOORS3.0
condition is matched by either a sine or cosine transform
(horizontal or vertical polarization). The IBC algorithm 233

initially proposed by Kuttler and Dockery [5] required a mix


of sine and cosine transforms. This "mixed Fourier transform" 117
(MFT) is essentially a spectral transform that has the IBC
embedded in its kernel. This implementation was a discretea ..
implementation of a continuous transform, and required twice Rg In

as many FFT calculations as a perfectly conducting boundary. h t i posilo


h is [dB]
o
F
0 _
This continuous MFT was found to be numerically unstable ! *-2
for some combinations of high sea state and high frequency. Fig. 6. Color plot of TEMPER F12 versus range and height. The two V-shaped
features are not physical - they are the result of instabilities in [1 1]'s DMFT.
Dockery and Kuttler [11] were later able to derive a
discrete-transform implementation that was more stable and (Letter lBand )
cut the number of transforms in half relative to [5] - i.e., the HF UHFVHF L S C X K. K
algorithm proposed in [I11] requires no more transforms than a 7at
perfectly conducting case. To illustrate how this is possible, fs 7t
note that if the left-hand side of (3) is used to define an B. 7

auxiliary function "w" for all z, not just z 0, then the function 1.5
"w" has negative symmetry and can therefore be propagated ac
I
using only sine transforms. The only additional burden of the the r
IBC then becomes back-solving a simple differential equation For 4
for u in terms In the discrete version of (3), [II] used a 3
centered, second-order finite difference for the z-derivative: 2
0.01 0.1 0.5 1.5 36 10 1520
un+ un -
Frequency (GHz) -oo
derivati(. ) natnFig. 7: Red indicates "bad alpha" problems in [ I]'s DMFT at the given
1e, 2r a Nive
insty Wne frequency and roughness; TEMPER is now stable in these regions.
Mca T,[1 1]'s
Despite its improvements over the continuous B. Improving The Miller-Bro,n Rough Surface Model
discrete MFT (DMFT) was still unstable for certain values of As mentioned above, TEMPER accounts for small-scale
a. Fig. 6 shows an example of this so-called "bad alpha" surface roughness via the parameter a in the IBC (3). More
instability. This problem was especially sensitive to frequency specifically, a "roughness reduction factor" modifies a via the
and small-scale surface roughness (a is a function of the reflection coefficient. Ongoing work is focused on improving
roughness-modified reflection coefficient; small scale the roughness reduction factor used in TEMPER.
roughness will be discussed in more detail in the next section). For simplicity, consider a smooth, flat ocean surface. In
Fig. 7 illustrates the frequency/roughness combinations that this case, propagation factor can be written in closed form as:
were problematic for [it]'s DMFT algorithm.
The "bad alpha" problem was solved in [12] by using an F 4i R)2 +4Rco2 ( c/2(5
alternate discretization for (4). Specifically, [12] uses either a Fflat I- II+4R O / 5

first-order forward (U+ - Un) or backward (un - Un-]


difference instead of a second-order difference for the z In (5), R is the complex Fresnel reflection coefficient, RD =k
derivative. These new discretizations still have regions of + j is the total phase difference between direct and reflected
instability; however, for a given a one can always find a stable paths, yi is the reflection coefficient phase, k =r21Tl is the
solution using either the forward or backward first-order wavenumber for radar wavelength 2, and 8 is the difference
difference. Hence, by using a judicious combination of the between the direct and reflected paths. Evident in (5) is the
two, TEMPER now avoids all of the "bad alpha" problems of well-known multipath effect; i.e., variations in 8 at a fixed
Fig. 7. The new first-order approximations to a/az also have a range generate maxima (peaks) and minima (nulls) in Fflat.
fringe benefit of faster back-solve (w h u) equations. On the This phenomenon also exists over a realistic rough ocean
other hand, there are larger discretization errors relative to (4). surface, however the multipath effects will differ from (5). No
Nevertheless, a comparison of the old and new DMFTs closed-form solution exists for a general rough surface.
showed nearly identical results down to 100 NMHz. Because Instead, a common approximation is to treat ocean waves as a
"bad alpha" instabilities were not a significant problem at "4small scale" roughness on a mean flat surface. This scheme
lower frequencies, TEMPER can safely revert to the more amounts to defining an effective reflection coefficient, Reff:
where p is the "roughness reduction factor." This parameter accordance with seawater at 20° C and 3.6% salinity. Because
p is a function of the ocean wave spectrum and radar grazing atmospheric refraction would only complicate the goal of
angle, and affects the field via the parameter a in (3). pis validating p, and because this is difficult to account for in
formally defined as an ensemble average of exp[2ik4siny] over method of moments, calculations assumed free space above
ocean surface heights ~;[15]: the surface.
For the MGIA calculations, a Monte Carlo approach was
oc used. 100 realizations of a rough sea surface were generated
p(2, y) = Fexp [2ik4sin y]P(4d; (7) using the 1-D Bjerkaas-Riedel surface-wave spectrum [17] at
_ wind speeds of 5, 10, and 15 m/s. Each surface was 5.05 km
long and discretized at 8 points per radar wavelength (i.e., 1.25
cm). Consequently, 2.5 cm is the smallest horizontal scale of
In (7), y is the grazing angle of the incident radiation and P(4) the wave spJectrum that iS included in the MGIA simulations.
is the probability density function (PDF) of surface heights.
9 Propagi factrs
~~Propagation factors FGa were then
FMGIA wr tae by
then obtained
s
taiong a
by taking
Through the effective reflection coefficient Re11, the coherent average over the 100 realizations at each wind speed.
magnitude of p (0 < p < 1) will modulate the multipath Quantitative comparisons were made between FMGJA and FMB
envelope; instead of varying between zero and two, the (propagation factors computed over a flat surface using the
propagation factor's variations will taper with increasing Miller-Brown approximation with an equivalent hrms). Both
altitude. That is, null depth and peak levels will diminish with depth and physical location of multipath nulls were compared.
increasing altitude due to an increase in grazing angle and Table I displays percent difference in vertical location of
correspondingly increased incoherent scattering. Furthermore, the FMGJA nulls versus the corresponding FMB nulls for three
if p is complex, its phase relates to changes in the spatial wind speeds and two ranges. Hyphens "-" mean "no null,"
location of peaks and nulls. Most well-known approximation, indicating that incoherent scattering was dominant in FMGJA -
including the two discussed here (Miller-Brown and Ament), i.e., the coherent interference pattern has been washed out to
evaluate to a purely real p in (7). Hence, these models account the point where there is no discernable FMGJA null. 0%
for reduced specular reflection, but they cannot account for corresponds to perfect null alignment between FMGJA and FMB;
shifts in null/peak locations. 100% corresponds to a null in FMGJA aligning with a peak in
TEMPER currently uses the Miller-Brown approximation FMB (or vice-versa). Null #1 is the first null above the surface
[13]. This model was developed assuming a collection of in altitude.
sinusoidal waves with Gaussian amplitudes and uniform phase TABLE I
distribution. That is, the surface height distribution was taken
to be = A*sin(x) where8the * .- 1 * * ~~~~~~PERCENTFOR
wave amplitude, A, is Gaussian
DIFFERENCE BETWEEN FMGIA AND FMBNULL LOCATIONS
A HORIZONTALLY POLARIZED LINE SOURCE
distributed with zero mean and standard deviation (hrms)/(12),
and a is uniformly distributed on the interval [-pT/2, ;T/2]. It U 5 m/s U= 10 m/s U= 15 m/s
was also assumed that the sea is sufficiently calm such that no Null # Rng Rng Rng Rng Rng Rng
significant superposition of sea waves occurs. With these 1 km 5 km 1 km 5 km 1 km 5 km
assumptions, [13] derived a PDF P() that, when inserted into 1 3.0 2.1 31 11 90 33
(7), resulted in the following roughness reduction factor: 2 4.8 2.8 33 18 - 48
3 6.4 3.0 36 23 - 64
~ ~~\14 5.7 4.4 43 29 75
(2kh,ms siny) (2khmssiny)j (8)
-
PMB (A,Y; hT )= expL-
2
10 5 7.8 5.2 41 33 -

6 6.6 4.4 - 36
In (8), Jo is a modified Bessel function of the first kind of order 7 6.2 5.3 - -
zero and hrms is the rms wave height. While the assumptions 8 6.0 7.4
used to derive (8) do not follow any realistic wave model, [13] 9 6.8 5.4
showed improved agreement with measured data using (8), 10 13.0 6.3
relative to alternate models for p circa 1984.
For the purpose of assessing the accuracy of the Miller- Table I shows that the Miller-Brown approximation
Brown approximation, a rigorous method of moments code becomes less accurate as wind speed increases. Physically, this
was used to perform rough sea surface scattering calculations. is because scattering becomes more incoherent as roughness
To facilitate the implementation of method of moments increases. Since the Miller-Brown model is an approximation
computations, we combined an accelerated iterative method for the coherent field, one would expect it to become less
and multigrid technique. For brevity, we will call this the accurate with increased wind speed. Comparing the two range
multigrid iterative approach (MGIA). All numerical cuts (1 km and 5 km) also shows that FMB is more accurate at
calculations were performed at a frequency of 3 GHz using a longer ranges. This is because, for a fixed altitude, longer
horizontally polarized line source (cylindrical wave) for the ranges translate to smaller reflection angles and, therefore, a
incident field. The source is fixed at a height of 15 m above more coherent reflected field.
the surface. Surface dielectric properties were assigned in

901
In addition to null location, taper of the interference solution. In this approach, field magnitudes and phases are
envelope was also compared to FMGIA. The accuracy of this computed for two atmospheres: the actual atmosphere and a
taper corresponds to how well the approximate method reference atmosphere. Phase differences between these two
predicts the amount of coherent scattering. Both the Miller- fields, AO, are then computed at every range and height, and
Brown and the well known Ament approximation were phase ambiguities are unwrapped. Range bias is then simply
considered; that is, tapers of FMB and FAment were compared to (AO / 2;T )*, where A is the radar wavelength in the reference
FMGIA's taper to determine which approximation was more atmosphere (usually 2O). AO between two fields is used, rather
accurate. The comparison was broken down into different than a single field, in part because this approach "normalizes
regimes of the Rayleigh roughness parameter, 2khrmssiny. For out" some of the PE phase errors at wider angles.
small Rayleigh values (< 0.88), both approximations tend to The primary difficulty in this approach is the 2D phase
over-predict the coherence scattering relative to FMGIA. For unwrapping, and is analogous to the phase unwrapping
moderate Rayleigh values (0.88 to 2.5), Ament tends to match problem typically encountered in interferometry. The first of
the FMGIA taper slightly better than Miller-Brown; Miller- two simple unwrapping algorithm developed by Hock Lim, et
Brown still has a slight coherent bias. At larger Rayleigh al. [19] was found to give good results for the TEMPER range
values (> 2.5), Ament begins to under-predict the coherence, bias application, provided that 1) multiple pixel-unwrapping
whereas FMB's taper aligns relatively well with FMGIA. Overall, passes were applied, with successively relaxed algorithm
Miller-Brown performs well, considering that a coherent bias tolerances, 2) physical insights (e.g., strength of the field,
generates deeper null fades, and thus a more conservative number of unwrapped neighbor pixels) were used to help
estimate of communication margins and target detectability. determine the order in which pixels were unwrapped.
Additional details of this comparison can be found in [15]. This PE-based method is extremely flexible. Furthermore,
More recent work has focused on developing improved it naturally computes measurement biases relative to any
approximate models that yield a complex-valued p. As alluded assumed atmosphere, not just free space. Many radar systems
to above, a non-zero phase on p can result in more accurate perform table-lookup corrections that are based on a nominal
null-location predictions. One model under investigation is refractivity profile. Using this nominal profile as the reference
somewhat ad hoc, but nevertheless motivated by the physical atmosphere in TEMPER, range biases can be computed for a
notion of a raised mean scattering-center plane due to wave specific radar system under specific atmospheric conditions.
shadowing. This ad hoc approach involves translating the And because the approach is based on full-wave methods, it is
mean plane by an amount Az until a value for p is obtained potentially more accurate and general than ray tracing.
that results in the first null of Feff being exactly aligned with The disadvantages of the approach are that both the
the first null of FMGIA. Interestingly, it was found that aligning TEMPER solution and the phase unwrapping are numerically
this first null tended to bring several higher-altitude nulls into intensive and, unlike ray tracing, elevation angle errors are not
alignment as well. A second, more physics-based method uses naturally provided along with the range biases. Angle errors
an expression for P() in (7) that incorporates shadowing. for specific radar systems can, however, be determined using
Specifically, we employ a "shadowed" PDF that is valid for TEMPER and monopulse processing.
low grazing angles. Initial results using this shadowed PDF in B. 3D Versus 2D Propagation Modeling
(7) show significantly improved null alignment [14]. JHU/APL also recently developed and tested a 3D
IV. ADVANCED TEMPER APPLICATIONS propagation model based on the vector PE (VPE):
All the improvements discussed above are related to ( 2 2
improving TEMPER capability. In addition to these efforts, - k2 + + |E (9)
TEMPER's new capabilities have also been applied as a ox a
2 oz2)
research tool in a wide range of radar propagation studies.
A. Refractive Range Bias where E is the vector electric field [20]. The 3D model solves
TEMPER was recently applied to the problem of (9) subject to an IBC at the terrain surface (10).
predicting refractive range biases for high resolution radars
[18]. Due to slowed propagation speed and refractive bending, nxE = Z fix(fixVxE) (10)
radar pulses take longer to return from an object than one V E= o
would expect based on free-space propagation time. Thus, the
actual range of a scatterer will be closer than the range In (10), Zis the terrain impedance and n is the outward unit
computed assuming a free-space time delay. The resulting normal vector to theterrain surface.
range measurement bias is often negligible, but can become Note that the terrain boundary conditions (10) couple the
significant for some geometries when radar accuracy is three field components, which precludes independent solutions
crucial. for the components. To overcome this difficulty, each range
Range biases are typically computed using look-up tables step of the VPB solution is carried out in two parts. The first
based either on analytic solutions for a simplified atmosphere part propagates fields without any terrain boundary, which
or ray tracing. One may also infer range biases, however, by uncouples the E components into three scalar finite-difference
counting the phase rotations in TBMPBR's full-forward-wave

902
equations. The second part of every range step corrects E for the numerical calculation. For visualization of the 3D field, the
boundary effects in close proximity to the terrain. This is done one-way propagation factor is plotted in three orthogonal
by discretizing (9) and (10) near the terrain, resulting in a planes. Fig. 9 provides a more quantitative comparison of the
sparse linear system of the form: two results at x = 22 km. In this bivariate histogram, perfect
2D-3D agreement at all power levels would correspond to all
Ax = b (I1) data clumped along a horizontal line in the middle of the plot.
The observed agreement is, in fact, surprisingly good,
The linear system (l1) is then solved iteratively to compute especially at higher power levels. Its only when 1F2 drops
the field values near the terrain boundary, and these corrected below -20 dB that 2D errors exceed ±5 dB. More pronounced
field values replace the values computed in absence of a 2D-3D differences were observed, however, in notional urban
boundary during part 1 of the range step. scenarios, where vertical building edges reflect and diffract
The primary motivation for developing a 3D model was to significant energy out of the 2D range/height plane [20].
assess the aforementioned 2D assumption made in deriving One of the hurdles to performing a broader 2D-vs-3D
TEMPER. TEMPER, along with virtually every long-range analysis is that memory and processor capabilities had limited
propagation model in the radar community, relies on this the comparisons in [20] to relatively small domains ("small"
assumption due to the numerical efficiency it affords. A 2D by radar standards, though quite large by 3D modeling
assumption is known to be quite accurate for open-ocean standards!). The primary bottleneck in the 3D solution scheme
propagation, and was assumed to be accurate for moderately- is the iterative solution of (I1). The convergence rate of this
sloped terrain. Using the 3D model, this assumption is now iterative scheme depends on the height and cross-range width
being assessed quantitatively over digital terrain. While an of the computation domain, as well as the steepness of terrain
exhaustive set of terrain has not yet been examined, initial slopes.
comparisons show relatively good 2D-vs-3D agreement [20]. Case 07, V-pol @ x=22000 (n=188221) 10
Fig. 8 compares TEMPER and the 3D VPE model for 15
vertically-polarized radar propagation over a small hill at the ll
southern tip of Ni'ihau Island in Hawaii. Digital terrain 0.8
resolution was -4O0 meters, linearly resampled to -I meter for 5
~~~~~~w0.6
()
_liJM,, ......
1_I~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ . . . . . . .

* ~ ~ ~ >rlE a .t]> >esvll


0A

X ~~~~~~~~~~~~~~~~-00 -0 -20D -1 0oewtitez=22


104 ln nFg .Te2 acltoswr

200 4ey(}=
arranged such that 2D and 3D numerical grids aligned perfectly at this plane;
(a) hence, no interpolations were performed in the creation of this histogram

~~z ~ ~ ~ ~~~~ P (in)

1.
{0 4 0 3|l
,! o
w
y (; * 1.
~~~~~~~~~~~~~~~~~~~~Jacobi

jle 44 w y (m) ~~~(b)


Fig. 8. Color plots of one-way power propagation factor lFl2 in dB on three _____ ___
slices within a 3D volume. (a) was rigorously computed by the 3D VPE, while -5 200 400 600 800 1000
(b) was assembled from multiple 2D TEMPER runs at various azimuths. Both Range (in)
calculations used the same digital terrain (hill -150-meter high, --600 m wide Fig. 10. Comparison of iteration count per range step for the 3D VPE routine's
& --800 m deep, centered at a range of --18.5 km from a 1 GHz, V-pol radar). new Zero Fill-in ILU preconditioner versus a Jacobi preconditioner. Range is
measured from the front edge of the hill in Fig. 8.

903
One may significantly enhance the numerical efficiency of the the environment's transfer function. In modeling realistic radar
iterative procedure, however, by designing a good systems, care must be taken to properly account for the
preconditioning matrix [21]. Recently, a Zero Fill-in ILU antenna beam and frequency-response characteristics at both
preconditioner was implemented in the iterative solution of the transmitter and receiver. The results shown below in Fig.
(11). Fig. 10 shows the dramatic improvement achieved by the 11 and Fig. 12 have assumed that transmitter and receiver
new preconditioner, relative to a Jacobi preconditioner, in responses are perfectly matched.
solving the problem shown in Fig. 8. Note that the Zero Fill-in Of particular interest to this application are errors related
ILU preconditioner reduces the number of iterations by three to PE assumptions. For example, the PE is known to have
orders of magnitude at certain range steps. This dramatic errors at angles away from horizontal. TEMPER's wide-angle
improvement will enable future efforts to examine larger and propagator has a tunable "problem angle" parameter; within
more varied swaths of digital terrain [22]. one "problem angle" above and below the horizontal,
TEMPER's solution is accurate [7]. It is worth taking a closer
tim-enn
In addition to 3D modeling, Ppaoodeling,
nt efforts henalo
look at TEMPER's accuracy here, however, because pulse
synthesis is sensitive to phase errors, and additive errors in the
explored time-dependent pulse modeling [23]. Asumined summation in (14) are of particular concern.
abve TEPE is a tim ineedn
moe.Sumn,
linear combination of chromatic solutions, however, allows
Fig. 11 investigates TEMPER phase errors for a 200-MHz
FFT b ndwithgeter 300 Mhz, usings probleMag
FFT bandwidth centered on 300 MHz, using a problem angle
TEMPER-based modeling of pulsed propagation. This
of 25' in TEMPER (marked with a 25° diagonal line on the
frequency-domain synthesis method has been routinely
employed in seismology and ocean acoustics (see, e.g., [24]). plot). The emitter is assumed to be point-source-like at all
The method is based on linearity. For typical radars in the frequencies. Within the 25° problem angle, Fig. 11 (a) exhibits
lower atmosphere, a linear-media assumption is appropriate the correct pulse shape, curvature, and magnitude (down to at
and, as mentioned above, used to derive the PE. Hence, least -30 dB). Above 25°, however, PE phase errors begin to
TEMPER is a linear operator that takes a source field with cause numerical dispersion; this can been seen as an artificial
frequency f at range x and returns the field at an advanced spreading of the pulse at higher angles in Fig. 11 (a).
x+Ax,
rangeaccounting for atmospheric and boundary effects. Curvature Check 200 to400 megahertz. 25 degree line Random error-O5to 05 by frequency
10

TEMPER(yi(x, z, f) q,i(x + Ax, z, f) (12) 200 0 200


ISO 150
A time-dependent field T' at range x can be expressed as a |-20
linear sum of k =[0,1,2,.. .,N-1] chromatic fields at fixed -3......0
frequencies cok= 2Tcfk. 0 0

N-1
50 100
150 200
250 300 350 s0 100 150 200
250 300 350

~~~~~~~~~~~~~~~~~~~~~~~~~~
T(x,z,t)
k=O
PkVI(XIZ,
( fk)e -i 2;7fkt
(13) (a)
Fig. 11: Effect of phase errors on (14): (a) invalid PE phase above the
(b)
"problem angle" (25°, noted by diagonal line) cause numerical dispersion,
however pulse is correct within the problem angle; (b) for comparison, the
In (I13), Pk are complex linear coefficients. By letting x 4 . . with artificial ±0.5° phase noise intentionally introduced into
same calculation
x+Ax in (13), substituting (12), then swapping the summation (14). Both plots show I]2up to -300 m in altitude and to G400 m in range.
and linear operator, one obtains (14).
Note that, at most ranges and heights outside the pulse, the
N-1 sum of TEMPER solutions accurately cancels to at least 1F2 =
T(x + Ax,z,t)= pkTEMPER(VJ(x, z,fk))e (14) -50 dB (the lowest value on Fig. 11's color scale). This
k=O cancellation indicates that TEMPER is computing a
sufficiently accurate ephase. For comparison, Fig. 11 (b) was
Equation (14) states that, given a time-dependent field at scmpted forat psce ario p nd phas error
range x, use TEMPER to range-advance each fk frequency cofp+0.5 was ithenial intriodcedt ' sum.
component of that field through the environment separately, Phase errors atideales andu all th other s ret
then sumthenthese
sum these
togeerateThe
solutions to generate the time
TEMPER solutions tie- to approximations made angles,
Phase errors at wide and all
in deriving thethe
PE,other errors
do not related
affect the
dependent solution at x ±Ax. The sum in (14) is efficiently more rigorous finite difference time domain (FDTD) method.
computed via FFT, and the coefficients Pk can be related to Hwvr Efeunydmi us oeighsahg
time dependence of the initial transmitted pulse. Discretization numeverical fefi
numerical efficiency advantage over FDTD, ming
ency-dvan ove FdT makingthe the Pe
PE
in time
in At and ndfreueny
imeAt frequency Af fk+] - fk are subject
f ==fk+ re ubjcttoto Nyquist
yqust attractive for long-range radar calculations. This advantage is
and aliasing constraints (see, e.g., chapter 8 of [24]). In exemplified by the time snapshots in Fig. 12, where a 20 ts "

practice, At and Af are also constrained by factors such as pulse has been propagated over a domain that extends 120 km
computational burden and desired output resolution. in range and 600 m in altitude. Fig. 12's scenario also
This frequency-domain approach is directly analogous to demonstrates the ability to incorporate realistic atmosphere
system transfer function analysis, with TEMPER representing

904
caliFbrniat stepfl9 t69423e6OQs , 1T8126km caIifornit: stpO78t2.4396eiO04s r73A264km daiif6rnia: stepll18#
I St6O-04s 1 M626km
500 5~~~~~~~~~~~00 501

400 400 400

300 300 300

200 200 200

100 I00 100 45


0 0 01|1|1 -201

20 40 60 s0 100 I2N 20 40 60 00 100 12( 20 40 60 so 100 120


(a) km (b) kmn (c) kmh
Fig. 12: Three snapshots of a modeled pulse propagating in a measured surface duct over realistic digital terrain. Equation (14) was evaluated using 128 frequency
samples over 0.3 MHz, centered at 1 GHz. The pulse itself is -20 ,us wide with a smooth, rectangular envelope.

and terrain effects. The refractivity profile is one measured via [10] R.O. Schmidt, "Multiple emitter location and signal parameter
rocketsonde off the coast of southern California, and the estimation," IEEE Trans. Ant. and Prop., vol. 34, no. 3, pp. 276-280,
terrain profile iS also from southern California, based on the
terrain sohrClfMarch
poiesasfo[11]
1986.
G.D. Dockery and J.R. Kuttler, "An improved impedance boundary
Digital Terrain Elevation Database (DTED). Fig. 12 (b) and algorithm for Fourier split-step solutions of the parabolic wave
(c), respectively, show terrain diffraction and atmospheric equation," IEEE Trans. Ant. and Prop., vol. 44, no. 12, pp. 1592-
refraction effects on the pulse for this realistic scenario refraction 1599, December 1996.
ee[12] J.R. Kuttler and R. Janaswamy, "Improved Fourier transform
V. CONCLUSION methods for solving the parabolic wave equation," Radio Science,
vol. 37, no. 2, pp. 5.1-5.11, March-April 2002.
TEMPER is one of the first successful applications of [13] A.R. Miller, R.M. Brown, and E. Vegh, "New derivation for the
Fourier split-step PE methods to the problem of long-range rough-surface reflection coefficient and for the distribution of sea-
wave elevations," IEE Proc., vol. 13 1(H), no. 2, pp. 114-116, 1984.
electromagnetic propagation. Through the years, many [14] D.E. Freund, N.E. Woods, H. Ku, and R.S. Awadallah, "The effects
improvements have been made to the code's numerical of shadowing on multipath radar propagation modeling," IEEE AP-S
algorithms, and significant advanced capabilities have been International Symposium and USNCIURSI National Radio Science
added. This paper has briefly described the current status of [15]
Meeting, Washington, DC, July 2005.
D.E. Freund, N.E. Woods, H. Ku, and R.S. Awadallah, "Forward
TEMPER along with recent developments to Improve radar propagation over a rough sea surface: a numerical assessment
robustness and to extend applicability. of the Miller-Brown approximation using a horizontally polarized 3-
GHz line source," IEEE Trans. Ant. and Prop., vol. 54, no. 4, pp.
REFERENCES 1292-1304, April 2006.
[16] J.Z. Gehman, J.R. Kuttler, and M.H. Newkirk, "A more robust
[1] H.W. Ko, J.W. Sari, and J.P. Skura, "Anomalous microwave formulation of TEMPER's rough-surface boundary condition,"
propagation through atmospheric ducts," Johns Hopkins APL Tech. National Radio Science Meeting, Boulder, CO, January 2001.
Digest, vol. 4, no. 1, pp. 12-26, 1983. [17] A.W. Bjerkaas and F.W. Riedel, Proposed Modelfor the Elevation
[2] M.H. Newkirk, J.Z. Gehman, and G.D. Dockery, "Advances
calculating electromagnetic field propagation near the earth's
in
Spectrum of a Wind-Roughened Sea Surface. Laurel,
HopkinsAPLTechnicalMemorandumTG-1328
MD:
Johns
December1979.
surface," Johns Hopkins APL Tech. Digest, vol. 22, no. 4, pp. 462-
H
ehmans GD. DckerandA.K Kocha, "Timedlyrn
472, 2001. [18] J.Z. Gehman, G.D. Dockery, and A.K. Kochhar, "Time delay range
[3]
472,
2001.
Tappert, "The parabolic
F.D. Tappert,"The parabolic approximation Wavebias
approximation method," in WaveSceeMetn,BudrCOJaay205
in
estimation using the parabolic equation," National Radio
S
Propagation and Underwater Acoustics, chap. 5, J.B. Keller and J.S. [19] H. Lim, W. Xu, and X. ' Huang,
' ' new practical methods for
"Two
Papadakis, Eds. New York, NY: Springer-Verlag, 1977, pp. 224-287. phase unwrapping," IGARSS Proc., vol. 1, pp. 196-198, 1995.
[4] G.D. Dockery, "Development and use of electromagnetic parabolic [20] R.S. Awadallah, J.Z. Gehman, J.R. Kuttler, and M.H. Newkirk,
equation propagation models for U.S. Navy applications", Johns "Effects of lateral terrain variations on tropospheric radar
Hopkins APL Tech. Digest, vol. 19, no. 3, pp. 283-292,1998. propagation," IEEE Trans. Ant. and Prop., vol. 53, no. 1, pp. 420-
[5] J.R. Kuttler and G.D. Dockery, "Theoretical description of the 434, January 2005.
parabolic approximation/Fourier split-step method of representing [21] Y. Saad, Iterative Methods for Sparse Linear Systems. Philadelphia:
electromagnetic propagation in the troposphere," Radio Science, vol. SIAM Press, 2003.
26, no. 2, pp.38a1-393, March-Apinl 1991. [22] R.S. Awadallah, J.Z. Gehman, J.R. Kuttler, and M.H. Newkirk,
[6] M. Levy, Parabolic Equation Methods for Electromagnetic Wave "Preconditioning schemes for vector parabolic wave equation
Propagation.VLondon,
Propagation. London, UK: IEEE,
IEEE, 2000. propagation modeling," National Radio Science Meeting,
[7] J.R. Kuttler, "Differences between the narrow-angle and wide-angle Washington, DC, July 2005.
propagators in the split-step Fourier solution of the parabolic wave 123l J.Z. Gehman, D.C. Scheinerman, J.R. Kuttler, and G.D. Dockery,
equation," IEEE Trans. Ant. and Prop, vol. 47, no. 7, pp. 1131-1140, "Modeling pulse propagation by frequency-domain synthesis of
July 1999. parabolic equation solutions," National Radio Science Meeting,
[8] D.J. Donohue and J.R. Kuttler, "Propagation modeling over terrain Washington DC July 2005.
using the parabolic wave equation," IEEE Trans. Ant. and Prop., vol. [24] F.B.Jensen,W.A.Kuperman,M.B.Porte
48, no. 2, p. 260-277 (2000).[4 ..Jne,WA uemn,MB otr n .Sh1t
[9] A. Beilis and F.D. Tappert, "Coupled mode analysis of multipleCoptinaOcnAouis.NwYr,Y:APre,194
rough surface scattering," J. Acoust. Soc. Amer., vol. 66, pp. 811-
826, 1979.

905

Вам также может понравиться