1

Application of Path-based Multipath Component

Tracking on Air-Ground Channel Measurement Data
Daniel M. Mielke, Michael Walter, Senior Member, IEEE, Dennis Becker, Miguel A. Bellido-Manganell,
Uwe-Carsten Fiebig, Senior Member, IEEE,

Abstract—The operation of unmanned aircraft is unthink- the UA, is necessary. Wireless channels are characterized by
able without reliable wireless communication links: Despite a reflection, diffraction, and scattering; these effects are summa-
comparatively high expected level of autonomy of unmanned rized as multi-path propagation [4]. Channel measurements are
aircraft, monitoring and remote controlling are required to some
degree. As the design of every communication link requires a common procedure to gain knowledge on wireless channels.
good knowledge on the characteristics of the communication This knowledge can later be used for the development and val-
channel, we have performed a measurement campaign to collect idation of a channel model. The resulting channel model may
channel sounding data for the wireless air-ground channel at cover several scenarios – in case of the air-ground channel,
C-band. While we have focused on the campaign description those scenarios can be the different phases of a flight like take-
and the analysis of the dominant signal component in previous
publications, we now concentrate on the detection and tracking off, en-route, and landing. A further differentiation, e.g. based
of multipath components. In this paper, we present our data on the ground surface, is also considered in literature, e.g. [5],
processing chain that allows a fast parallel processing of the [6]. The findings from the channel measurements and the
measurement data, as data dependencies are reduced as much resulting channel model can be used to design and evaluate
as possible. We furthermore introduce a path-based multipath waveforms for a new wireless communication system.
component tracking approach and apply it to our measurement
data. This tracking allows us to estimate the location of reflectors Inherently, channel measurements for aeronautical channels
causing multipath component signals. We apply our processing are quite complex and costly as they require at least one air
chain to data recorded during take-off at a small airport and vehicle. Both costs and effort may rise significantly when
compare the results of the reflector localization to a satellite jet aircraft operating at high speeds at several kilometers of
image of the airport to successfully verify our approach. altitude are involved. Additionally, there was no high demand
for sophisticated models for aeronautical channels in the past
I. I NTRODUCTION since modern digital communication links did not play a role
in civil aviation for decades. These two aspects might be
D URING the next years, more and more Unmanned Air-
craft (UAs) are expected to enter the skies. According to
[1], the market for UA is expected to have a compound annual
the reasons, why aeronautical channels in general and the
aeronautical air-ground channel in particular have not received
as much attention as other channels like those used in cellular
growth rate of 12.23 % until the year of 2027. The anticipated
networks, e.g. LTE and 5G. However, we want to provide
fields of application are transportation or other logistic tasks,
a brief overview of the available literature in the field of
as well as surveillance, reconnaissance, exploration, and tasks
aeronautical (air-ground) channel models. A comprehensive
in the agronomy. While UAs are already routinely used in
review on air-ground channels, with a special focus on UAs
military environments, civil applications, like large unmanned
can be found in [7].
freight aircraft, only emerge. The main task is the integration
In [8], [9], the application of simple tap-delay line models
of civil UAs into non-segregated airspace [2].
on aeronautical channel modeling has been discussed. The
One key aspect of this integration is the implementation of a
authors have already stated, that the different phases of a flight
reliable communication link that allows the exchange of Com-
require different models or at least an individual parametriza-
mand and Control (C2) data and telemetry data, respectively,
tion.
between the air vehicle and the remote pilot among other
The development of digital data links for civil aviation
potential communication partners. The type and amount of
motivated the development of more sophisticated channel
data that needs to be transmitted has been widely discussed in
models. For example, a Wide Sense Stationary Uncorrelated
[3], where both satellite-based and terrestrial (i.e. air-ground)
Scattering (WSSUS) based channel model for C-band has been
data links are investigated.
developed as part of the adaption of IEEE 802.16e to an airport
For the development of reliable, high-throughput wireless
surface data link called Aeronautical Mobile Airport Commu-
data links, a good understanding of the physical characteristics
nication System (AeroMACS) [10]; however, according to the
of the wireless channel between the communication partners,
application of AeroMACS, the model focuses on the airport
in our case of a terrestrial system the Ground Station (GS) and
surface only. An alternative approach for modeling the airport
D. M. Mielke, M. Walter, D. Becker, M. A. Bellido-Manganell, and surface channel, also in C-band, has been introduced in [11].
U.-C. Fiebig are with the German Aerospace Center (DLR), Insti- In contrast to AeroMACS, the L-Band Digital Aeronautical
tute of Communications and Navigation, 82234 Wessling, Germany (e-
mail: {daniel.mielke, m.walter, dennis.becker, miguel.bellidomanganell, uwe- Communication System (LDACS) is a communication link de-
carsten.fiebig}@dlr.de) signed for all flight phases of manned aircraft, predominantly
 2

developed by the German Aerospace Center (DLR). As part A detailed description of the campaign can be found in [16];
of its development, a channel measurement campaign in L- this section just intends to give a brief overview on the
band was performed and a new approach of channel modeling campaign setup.
was proposed [12], [13]. The channel model is based on the
evaluation of the tracked Multipath Components (MPCs) and A. Hardware Setup
does not only incorporate statistical element but also integrates
The campaign setup consisted of a Ground Station (GS) and
geometric aspects.
an Airborne Station (AS) aboard DLR’s Falcon 20E aircraft.
In [14], air-ground channel measurements with a bandwidth
Through all experiments, the GS was the transmitter of the
of 2 MHz in the Ultra High Frequency (UHF) band have been
channel sounding signal and the AS was the receiver of the
performed. The authors also provide a model on the path loss
signal.
and discuss the observed effects of multipath propagation by
The most relevant devices of the GS are the Arbitrary
a statistical evaluation of sets of Channel Impulse Responses
Waveform Generator (AWG) and the High Power Amplifier
(CIRs).
(HPA) that amplifies the output signal of the AWG. The GS
Comprehensive channel measurements in both C- and L-
also contains a Global Navigation Satellite System (GNSS)
band have been performed by a team of the University of
receiver and an atomic clock. The latter two devices act as a
South Carolina and NASA. The evaluation has been published
GNSS-disciplined oscillator that is used as time base for the
in a series of articles: While [5] described the campaign
AWG.
setup and proposed models for the air-ground channel in over-
The most important part of the AS setup is the IQ-
water scenarios, the team focuses on the modeling of the air-
recorder that samples the received signal, and counts and
ground channel in hilly and mountainous terrain in [6]. In [15],
stores these IQ-samples. Additionally, the setup contains an
the focus lies on the modeling of the air-ground channel in
Inertial Measurement Unit (IMU) and also a GNSS-disciplined
suburban and near-urban environments. The described models
oscillator that is used as a time base for the IQ-recorder. As
concentrate on the evaluation of sets of CIRs and not on the
the applied IQ-recorder is not capable of handling the applied
tracking of individual MPCs over time.
carrier frequency, a frequency mixer (”downconverter”) is used
The DLR also performed a C-band channel measurement
to shift the signal to an intermediate frequency. This mixer
campaign with a jet aircraft that covered multiple flight
also provides an Adaptive Gain Control (AGC), that allows
scenarios. This measurement campaign was described in detail
an adaptive amplification or attenuation, respectively, of the
in [16], where we have also provided a close analysis of the
received signal. It is therefore possible to adjust the receiver’s
dominant signal component of the received channel sounding
operating point during recording e.g. depending on the Line
signal. In our present paper we focus on the signal components
of Sight (LOS) distance to the transmitter.
besides the dominant component, the MPCs, since they, as
stated above, strongly define the physical properties of a
wireless channel. We address the task of identifying and B. Measurement Signal and Procedure
tracking the MPCs in our measurement data by the application The channel sounding signal consists of a so called channel
of a multi-stage processing chain. In contrast to [17] we do not sounding sequence of length 40.96 µs that is gaplessly repeated
apply a filter-based algorithm but a concept, where the MPCs in an infinite loop on the GS’s AWG. The channel sounding
are first identified and are later tracked by a novel path-based sequence is an iteratively filtered multitone signal whose initial
algorithm in a second step. This two-staged approach allows phases are distributed as Newman Phases. A more detailed
a parallelization of the computational expensive processing of description of the signal generation can be found in Section
the raw measurement data. II-B of [16].
The paper is structured as follows: In Section II we briefly As the receiver samples the incoming signal at fsr =
describe the setup of the measurement campaign the processed 50 MHz, it is assured that N = 2048 consecutive samples of
data has been collected. The pre-processing of this data is the received signal consist exactly one complete instance of
explained in Section III. Sections IV and V provide infor- the channel sounding sequence.
mation on how the individual MPCs are extracted from the Before and after all measurement flights, the output of the
measurement data. In Section VI, we introduce our approach GS’s HPA is directly connected to the receiving hardware
used for the tracking of the detected MPCs over time before we aboard the AS using attenuators and a cable to perform a
apply it to our measurement data in Section VII. We conclude reference measurement of the channel sounding sequence. The
our paper in Section VIII. transmission and the recording of the channel sounding signal
Fig. 1 also provides a graphical overview on how the part is then started for a few seconds. During post-processing, N
of the paper explaining the processing is structured. consecutive samples of this recorded signal are cut out and
called reference signal xref ∈ CN . We also store the IQ-
recorder’s sample counter value ρref for the first sample of
II. M EASUREMENT C AMPAIGN xref . Additionally, the GNSS-disciplined oscillators of the GS
The DLR performed a measurement campaign in 2018 and the AS are synchronized during this process.
where the air-ground/ground-air channel in C-band was mea- During the measurement flight, the GS is in its transmission
sured. The applied channel sounding signal had a bandwidth of location on the rooftop of DLR’s IKN building and the output
about 50 MHz and used a carrier frequency of fc = 5.2 GHz. of the HPA is connected to the transmitting antenna. The
 3

sRx
information on the aircraft’s position, orientation, and heading
Pre-Processing at a certain time instant. This allows us to estimate the

Section III
1) Cut out data vector LOS component’s Free Space Path Loss (FSPL), the LOS
2) Compensate shifts
according to LOS component’s delay, denoted by τLOS , and the LOS component’s
3) Reshape data to block
Doppler shift, denoted by νLOS .
X (b)

Coarse MPC Detection B. Block Processing

1) Compute either
Section IV

- Impulse response, or The processing of the measurement data is performed block

- Power Delay Profile, or
- Doppler-Delay Spreading Function wise. A block b of BN consecutive samples is cut out of the
2) Compute noise power received signal:
3) Detect peaks, convert indices
0
M(b,crs) x(b) = sRx [ρ(b) , ..., ρ(b) + BN ]
yes no = sRx [ρ(b) , ..., ρ(b+1) − 1], (1)
optimize
where ρ(b) denotes the IQ-recorder’s sample counter corre-
Fine MPC Detection sponding to block b. We call B ∈ N the block size, as it defines
Section V

1) Initialize optimization with how many consecutive channel sounding sequence instances
coarse MPCs
2) Optimize within bounds of length N are processed together in one block. A larger B
given by resolution
increases the Doppler resolution during processing, however,
M(b,fne) M(b,crs) this increase comes at the cost of time resolution: If B is
chosen large, it is possible that relevant fast fading channel
Section VI

MPC Tracking
1) Create graph from detected MPC data effects are vanished out and thus are not modeled appropriately
2) Compute metrics later. This effect becomes even more likely considering the
3) Detect paths
4) Apply path improvement strategies comparatively high velocity of an aircraft.
0
Based on the ground truth data, x(b) is shifted such that the
P current LOS delay and LOS Doppler shift are compensated:
(b)
Fig. 1. Block diagram of the processing described in this paper. The section −ν  (b)0
where the corresponding processing step is discussed in detail is given next x(b) = F LOS
(b) x , (2)
−τLOS
to the box.
where Fντ {.} denotes a function shifting a signal by delay τ
and frequency ν.
AS’s receiving hardware is connected to the receiving antenna
In a last step, we reshape the vector x(b) ∈ CBN to a
mounted at the bottom fuselage of the Falcon aircraft. During
matrix X (b) ∈ CN ×B , where each column corresponds to
the flight, the AS’s IQ-recorder stores the received samples in
one instance of the channel sounding sequence, thus the items
sRx .
of X (b) are set according to:
It is ensured, that the GS’s AWG and the AS’s IQ-recorder
are continuously running during and between the pre-flight X (b) [n, m] =x(b) [mB + n],
reference measurement, the actual measurement flight, and the ∀ n ∈ {0, ..., N − 1},
post-flight reference measurement. Thus, we can simply use ∀ m ∈ {0, ..., B − 1}. (3)
the sample counter ρ of the IQ-recorder as a reliable time
base during offline processing, when the measurement data In the following, we assume that the power levels of xref
recorded during flight are evaluated using the reference signal and x(b) are matched, such that the attenuation used during
xref , see Sections III to V. the recording of the reference signal and the FSPL of the LOS
component during the recording of x(b) are compensated as
C. Flight Tracks and Maneuvers suggested in [16, (11)].
All flights started and ended at the EDMO airport close to
Munich, Germany. The campaign involved four flights where IV. C OARSE M ULTIPATH C OMPONENT D ETECTION
different flight scenarios were covered. The flight scenarios not In this section, we describe our approach for detecting
only included typical en-route flight patterns at cruising speed Multipath Components (MPCs) in our measurement data. As
at multiple cruising altitudes, but also take-offs, landings, go- shown in [17], an MPC l can be described by a quadruple
arounds, and flights with extreme banking. The data collected
during the different flight scenarios allows us to develop ξl = ( τl , νl , |αl | , arg {αl } ) , (4)
individual channel models for each scenario. where τl denotes the delay shift, νl the Doppler shift, and αl
the complex weight of the MPC1 . The weight αl is split into
III. P RE -P ROCESSING its absolute value and its phase to achieve ξl ∈ R4 , which not
A. Ground Truth
1 Note that MPCs that cannot be resolved due to limited resolution are
Based on the data recorded by both GNSS-receivers and the handled as a single MPC, although the more correct term would be MPC
IMU, a ground truth is computed. The ground truth provides cluster.
 4

only allows a direct access to the MPCs’ amplitudes, but also 4) Index Translation: Based on the resolution of the IRs
a more memory efficient processing. If not denoted otherwise, given by (22), the indices found in the previous step can be
we assume that both τl and νl are given with respect to the translated to a delay according to
LOS component.
The set of MPCs that have been detected in a processing τl = nl ∆τ . (6)
block b is denoted by M(b) having cardinality L(b) . To avoid
ambiguities, the block index b is added to the MPC notation The complex weight is given by αl = y (b) [nl ]. In order to
(b)
according to ξl if necessary. complete the four elements of the MPC as defined in (4), the
We now describe the processing steps for the coarse detec- Doppler shift of MPC l is set to νl = 0. Then, ξl is added to
tion of MPCs or clusters of MPCs, respectively. The detected M(b,crs) .
set of MPCs is therefore denoted by M(b,crs) .
Please note that the described approaches have no data
dependencies on previously processed blocks, which allows B. Based on Doppler-Delay Spreading Function
a parallel processing of an arbitrary amount of blocks at the
same time. The basic idea of this approach is very similar to the previ-
ous one, however, instead of using the IR/PDP, the Doppler-
Delay Spreading Function (DDSF) is used. The DDSF pro-
A. Based on Impulse Response/Power Delay Profile vides not only information on the delay, but also on the
The basic idea of this approach is to detect peaks in the Doppler shift of an MPC. Again, peaks in the DDSF represent
absolute value of the Impulse Response (IR) or the Power either a single MPC or a cluster of MPCs.
Delay Profile (PDP), respectively, of a data block. Each of 1) Doppler-Delay Spreading Function: The detection pro-
these peaks represents either a single MPC or a cluster of cess starts with the computation of the DDSF according to
MPCs; the position of the peaks can be used to estimate the Appendix C of the current data block b as defined above and
underlying MPCs’ delay τl . However, no Doppler information take its logarithmic absolute value:
νl can be extracted (directly) following this approach. n o
f
1) Impulse Response/Power Delay Profile: The first step is A(b) = 10 log10 Sxupref X (b) , (7)
to compute either the IRs or the complex PDP of the current
data block X (b) by applying the processing described in where fup denotes the upsampling factor along the delay axis
Appendix A or Appendix B, respectively. While the (coherent) i.e. the columns of A(b) ; the rows correspond to the Doppler
PDP has the advantage of a lower noise floor, short term MPCs axis.
may vanish out if the block size B is chosen too large. On the
2) Noise Floor: The power of the noise floor for the current
other hand, if B is chosen too small, signals of some reflectors 2
block σ (b) is estimated based on areas of A(b) where the ap-
may vanish in the noise floor and will remain undetected.
pearance of measurable strong MPCs or measurement artifacts
As the further processing is the same for both cases, in the
is very unlikely. Thus, the average power for the areas where
following, we refer to both the individual IRs of the current
{τσ |τmax < τσ < N/fsr } and {νσ |νmin < |νσ | < νmax } is
block and the PDP, respectively, by y (b) ∈ CN fup . fup denotes
computed.
the upsampling factor.
We also define the corresponding logarithmic vector 3) Peak Detection: The next task is to detect local maxima
(”peaks”) in A(b) . A peak must fulfill the following two
(b)
y log = 10 log10 y (b) . (5) conditions:
2
• The peak’s value must exceed a value of σ (b) |dB +Pthresh .
2) Noise Floor: The power of the noise floor for the • The peak must have a certain prominence defined by an
2 (b)
current block σ (b) is estimated based on areas of y log where elliptical footprint. This allows a more flexible search
the appearance of measurable strong MPCs or measurement than a simple distance measure.
artifacts is very unlikely. Thus, the average power for the part
(b) All detected L(b) peaks’ coordinates inside of A(b) are stored
of y log where {τσ |τmax < τσ < N/fsr } is computed, where pairwise (nl , ml ) in P (b) , where nl denotes the row index and
τmax denotes the threshold delay up to which significant MPCs ml denotes the column index of the l-th peak.
are expected.
4) Index Translation: Based on the resolution of the DDSF
3) Peak Detection: The next task is to detect local maxima
(b) given by Eqs. (22) and (23), the index pairs found in the
(”peaks”) in y log . A peak must fulfill the following two
previous step can be translated to a delay and Doppler value.
conditions:
2 For the delay, this mapping is done according to (6), for the
• The peak’s value must exceed a value of σ (b) |dB +Pthresh . Doppler, the mapping is done according to
• The peak must have a certain prominence defined by a
 
minimum distance to other detected peaks. B
(b) νl = ml − ∆ν . (8)
All detected L(b) peaks’ positions nl inside of y log are stored 2
in P (b) , where nl denotes the index of the l-th peak.
 5

5) Least Squares Optimization: To find the complex 3) Dimensions: The optimization algorithm tries to achieve
weights for the MPCs, we first create a matrix describing an optimal result by adjusting the delay and Doppler shift of
a synthesized version of the signal based on the delays and every MPC. As the amount of assumed MPCs does not change
Doppler shifts of the MPCs detected above using the shift during the processing of block b, the optimization problem is
function Fντ {.} introduced in Section III: solved along 2|M(b,crs) | = 2L(b,crs) dimensions.
 n o  4) Optimization Bounds: The optimization bounds for each
Fντ00 xref
(B) MPC are given by the DDSF’s tiles’ dimensions. Thus, the
 n
(B)
o  lower bound (lb) and the upper bound (ub) for the delay
 Fντ11 xref
 
 , Y (b) ∈ CL(b) ×BN , optimization of MPC l are given by:
Y (b)

=

..  (9)
 .n   
 ν
(b) −1 (B)
o
({lb,ub}) ∆τ
FτLL(b) −1 xref τl = max 0, τl ∓ (12)
2
(B)
where xref denotes a vector with B concatenated instances and, correspondingly, for the Doppler optimization of MPC l,
of the reference signal. Thus, the l-th row of Y (b) contains a the bounds are given by
modified version of the reference signal, shifted in time and
frequency according to τl and νl , respectively. ({lb,ub}) ∆ν
(b)
Now, an approximation for the vector α(b) ∈ CL needs νl = νl ∓ . (13)
2
to be found for the equation
T
5) Optimization Process: The algorithm basically executes
Y (b) α(b) = x(b) . (10) the LS optimization described in Section IV-B5 in an iterative
loop indexed by i ∈ N:
We apply the Least Squares (LS) algorithm for this task. (b)
Besides the desired approximation for α(b) , the LS algorithm First, a matrix Ỹ i is generated using the delay and
also returns a residual (b) that we understand as a measure Doppler shifts of the MPCs given in the current (hypothetical)
(b)
of precision of the estimated MPCs’ parameters. The residual set of MPCs M̃i similar to (9). Then, an equation similar
is computed according to to (10)
T (b)T (b)
= x(b)
X
(b) = ||x(b) − Y (b) α(b) ||2 , (11) Ỹ i α̃i (14)
| {z }
BN (b)
∈RBN
is solved for α̃i – again using the LS approach.
which corresponds to the sum over all elements of the vector Thus, the overall objective of the optimization problem is
that results from the squared Euclidean distance between the (b)
to minimize the residual i returned by the LS algorithm
measured signal block and the weighted synthesized signal. applied to (14). Using the definition given in (11), this leads
By combining τl , νl , and the absolute value and phase of to
the l-th entry of α(b) , all elements of MPC ξl , as defined in
(4), are given. ξl is now added to M(b,crs) . ( )
n X (b)T (b) 2 o
(b,fne) (b)
M = arg min x − Ỹ i α̃i .
(b)
V. D ETECTION I MPROVEMENTS M̃i BN
| {z }
A. Fine Detection (b)
i
The potential problem of the approach described in Sec- (15)
tion IV-B is, that its precision is limited by the resolution of
the discrete DDSF. Therefore, the exact position of an MPC 6) Termination: The execution of the optimization is ter-
inside of a tile of the DDSF matrix – and consequently its minated when the result does not improve more than a certain
(b) (b)
exact delay and Doppler shift – cannot be determined any threshold from one iteration to the next: i−1 − i < thresh .
further.
We address the this issue by applying an optimization B. Cluster Resolution
algorithm that tries to adjust the delay and Doppler shift of
every detected MPC within the bounds defined by the size of Although the approach presented in Section V-A has the
a tile of the DDSF. ability to determine the parameters of the MPCs even below
the resolution of the DDSF, it cannot resolve multiple MPCs
1) Optimization Algorithm: As suggested in [17], we
whose mutual distances in delay and/or Doppler are lower than
apply the Bound Optimization by Quadratic Approxima-
the corresponding DDSF resolution. Thus, they can only be
tion (BOBYQA) algorithm [18] to the given problem. The
described as clusters and not get detected individually. In case
BOBYQA algorithm can be applied to multi-dimensional
this is not acceptable, we suggest the following approach to
problems, does not require a derivative of the optimization
resolve the individual MPCs.
objective, and allows the usage of optimization bounds.
It is assumed that the coarse detection process as defined in
2) Initialization: The optimization is initialized by the Section IV-B has been completed; thus there is a set M(b,crs) .
delay and the Doppler shifts of the MPCs given in M(b,crs) . For each item in M(b,crs) , Ψ ∈ N copies are created. A
 6

random delay and Doppler shift is added to each of these 1 # o r d e r e d l i s t where a l l d e t e c t e d p a t h s a r e s t o r e d

2 list all paths = ()
copies according to 3
4 # i t e r a t e over a l l blocks along time
5 f o r i n t b i n range ( 0 , num blocks −1) :
∀ l ∈ {0, 1, . . . , L(b,crs) − 1} : 6
7 # r u n u n t i l a l l MPCs o f c u r r e n t b l o c k a r e a s s i g n e d
to a path
∀ ψ ∈ {0, 1, . . . ,Ψ − 1} : 8 w h i l e n o t a l l ( Γ(b) == 1 ) :


τl,ψ = τl + U[0,1] − 0.5 ∆τ (16) 109 # i t e r a t e o v e r MPCs o f c u r r e n t b l o c k
f o r i n t l i n range ( 0 , L(b) −1) :


νl,ψ = νl + U[0,1] − 0.5 ∆ν , (17) 11
12
13 # c h e c k i f MPC i s a l r e a d y a s s i g n e d t o a p a t h
i f Γ(b) [ l ] == 1 :
where τl and νl correspond to the delay and Doppler shift of 14 15
ξl ∈ M(b,crs) and U[v,w] denotes a uniform distribution on the 16 17
# i t i s → c o n t i n u e w i t h n e x t MPC
continue
interval [v, w]. These newly created MPCs are denoted by ξl,ψ 18
(b,crs) 19 # c a l l subprocedure to d e t e c t path
and summarized in MΨ . 20 l i s t γ = detect path (
21 root block index = b ,
(b,crs)
So far, this approach alleges, that every item in M is 22 root mpc index = l
23 )
in fact a MPC cluster. However, the idea is that due to the 24
further processing, i.e. by applying the optimization algorithm 25 26
# append p a t h t o p a t h l i s t
a l l p a t h s . append (
from Section V-A, the weights of the ”unnecessary” MPCs 28 27 γ
)
copies will tend to 0: Listing 1. Pseudo code of the outer part of the path detection algorithm. See
The algorithm given in Section V-A is initialized using Listing 2 for the detect path subroutine.
(b,crs)
MΨ . For each ξl,ψ , the optimization bounds are the same
as for the corresponding ”parent” MPC ξl .
The amount of optimization dimensions is 2ΨL(b,crs) . Once B. Path Detection
the termination criterion is fulfilled (see above), the resulting We use a path, represented by a sequence
(b,fne) γ = ((b, l)i )i=0,1,...,β−1 , to describe the behavior of an
MPCs are given in MΨ .
MPC over time. Each element of a path is a tuple (b, l) that
sufficiently describes the position of each node inside of G.
VI. M ULTIPATH C OMPONENT T RACKING To avoid ambiguities, we add a superscript to our notation
(e.g. γ (a) ) when necessary. The set of all paths is denoted by
Tracking of MPCs describes the behavior of the components P.
over time. This is often realized by filter-based approaches, For the sake of simplification, we have split the description
e.g. in [17]. However, path-based approaches have also been of our path detection algorithm into three parts: the initializa-
discussed, e.g. in [19]. We now describe our path based tion, an inner part, and an outer part.
approach to track MPCs over time, i.e. multiple consecutive
1) Initialization: For each block b, a vector Γ(b) ∈ NLb
blocks b.
is defined and all of its items are initially set to zero. The
purpose of this vector is to indicate whether an MPC has been
assigned to a path: If MPC l of block b gets assigned to a path,
A. Graph Representation
Γ(b) [l] = 1 is set.
Based on the sets of MPCs M(b) ∀ b detected using one of A threshold dthresh ∈ R is defined that provides a maximum
the approaches given in Sections IV and V, a directed graph distance two MPCs are allowed to have within a path. If this
G = (M, D) as given in Fig. 2 is created. The nodes SM of the threshold is exceeded, the respective MPC cannot be part of
graph are given by the union of all MPC sets: M = b M(b) . the given path.
The edges D of the graph are given by the set of all distances For the sake of completeness, an ordered list (corresponding
between the MPCs of consecutive blocks. These distances are to P) is created in Listing 1, line 2 where all detected paths
defined as follows: The distance from the ls -th MPC (start) of are stored.
the b-th block to the le -th MPC (end) of the (b + 1)-th block 2) Outer Part: The purpose of the outer part (see Listing 1)
is given by of the algorithm is to detect the root note of a new path. Once
n o a root element has been found, the inner part, represented
(b→b+1) (b) (b+1)
dls →le = D ξls , ξle , (18) by the subroutine detect path, is called to find the remaining
elements of the path. In the following, we explain the outer
where D : R4 × R4 7→ R is a function expressing the distance part of the algorithm in detail:
between two MPCs by a real value. Possible realizations for The loop defined in line 5 iterates over the blocks starting
D are discussed in Appendix D. at the first block. Considering the graph in Fig. 2, this loop
The entry of the ls -th row and the le -th column of the iterates over the blocks from left to right.
(b) (b+1) (b→b+1) The loop in line 8 ensures that the algorithm keeps pro-
distance matrix D (b) ∈ RL ×L equals dls →le .
We furthermore assume, that all MPCs of block b are sorted cessing on the current block b, until all of its MPCs have
(b) been assigned to a path.
in descending order based on their absolute weight: αl ≥
The loop defined in line 11, finally, iterates over the MPCs
(b)
αl+1 . of the current block b. Unless the current MPC l has been
 7

(3→4)
d0→0
3)
(0) (1) (2 → (3) (4)
ξ0 ξ0 d 0→0 ξ0 ξ0
d (0→1
0→ )
1
2)
(1→ (2)
d 1→0 ξ0

(0) (1) (3) (4)

ξ1 ξ1 ξ1 ξ1

2) d1(2→3)
(1→ (2)
d 2→1 ξ1 →2
(0→1) 4)
d2→2 →
(3 →1
(0) (1) (3) d3 (4)
ξ2 ξ2 ξ2 ξ2
d (3→4
2→ )
3
2) d2(2→3)
(1→ (2)
d 3→2 ξ2 →3

(0) (0) ×L(1) (1) (1) ×L(2) (2) ×L(3) (3) (3) ×L(4) (4)
ξ3 D (0) ∈ RL ξ3 D (1) ∈ RL D (2) ∈ RL ξ3 D (3) ∈ RL ξ3
b=0 b=1 b=2 b=3 b=4

(b)
Fig. 2. Graph representation of the MPCs ξl detected in five consecutive blocks indexed by b; MPCs are represented by a black dot, blocks are represented
by a dashed-line rectangle: While only three MPCs have been detected in block b = 2, four MPCs have been detected in the other blocks. The arrows
(b→b+1)
connecting the MPCs represent the distance metrics dls →le between all MPCs of consecutive blocks; however, for the sake of clearness, only few distance
metrics are given in the figure. All distance metrics between block b and block b + 1 are given in D (b) .

assigned to a path yet (check in line 14), the actual path caused the loop’s termination, the subroutine returns the de-
detection subroutine detect path, corresponding to the inner tected path γ.
part of our approach, is called in line 20. The arguments passed
to this call are the current block b and MPC l as they are used C. Path Improvements
as the root of the new path. The algorithm presented in Section VI-B has properties
Once detect path returns a new path γ, it is appended to that may result in a degraded MPC tracking under certain
the global path pool in line 26. conditions. We suggest the following approaches to address
Please note that the update of Γ(b) is performed inside of these issues.
the called subroutine. Due to this side-effect we do not call 1) Path Merging: In case an MPC remains undetected for
detect path a function to highlight the lack of idempotence. the duration of a single (or more) block(s), e.g. because it does
3) Inner Part: The inner part of the algorithm is given in the not exceed the detection threshold, an actual path is detected
subroutine detect path in Listing 2. As the arguments passed as multiple individual paths. To detect and connect these path
to this subroutine describe the root of a new path inside of the segments, we evaluate the distances between the last element
graph, this tuple is added as a first element (line 10) to the of all paths and the root elements of all other paths:
list representing the path γ. We first set a threshold βmax ∈ N that defines the maximum
During this subroutine, the MPC index of the last element of length of interruptions given in blocks that we consider as
the path is represented by l; in the beginning, this corresponds acceptable. We then iterate over all permutations of all paths
to the MPC index of the root node (line 15). γ (m) and γ (n) to detect all path combinations, where
(n) (m)
Similar to the block loop in Listing 1, the block loop of 0 < broot − bend < βmax , with
the subroutine iterates from ”left” ro ”right” over the blocks, (n) (n) (n)
(broot , lroot ) := (b, l)0 ∈ γ (n)
starting with the root block (line 18).
(m) (m) (m)
Within the loop, the distances from the current MPC l to all (bend , lend ) := (b, l)β (m) −1 ∈ γ (m) (19)
MPCs of the next block that are not yet assigned to a path are holds. The paths γ (m) and γ (n) are merged if
computed and the minimum is determined (line 30). Please (m) (n)
(b →broot )
note, that the actual call of D can be substituted by a lookup d (m)
end
(n) < dthresh (20)
lend →lroot
in the corresponding distance matrix D (b) .
In case either no unassigned MPCs can be found (check is fulfilled. In case this condition holds for multiple path pairs,
in line 24) or the determined minimum distance exceeds the one with the lowest distance is chosen.
the threshold dthresh (check in line 33), the loop stops – So far, the stability of a path γ (m) was equal to its length
(m)
corresponding to a termination of the current path. β . However, the process of path merging motivates the
If none of these conditions is fulfilled, the index of the MPC definition of a new measure of the path stability, as the path
with the minimum distance is assigned to l (line 39) and the now may skip some blocks. We therefore define the path
new node is appended to the path γ (line 42). Finally, the stability η (m) as the delta of the block index of the last element
detected MPC is marked as assigned (line 47). of the path end the block index of the root element of a path:
(m) (m)
η (m) = bend − broot .
Once the loop has terminated, no matter what criterion
 8

1 sub d e t e c t p a t h ( VII. R ESULTS AND D ISCUSSION

2 int root block index ,
3 int root mpc index In this section we present some results from the different
4 ) → list :
5 processing steps described above. Due to the power normal-
6 # i n i t i a l i z e o r d e r e d l i s t f o r new p a t h
7 list γ = () ization assumed in Section III, all amplitudes are given in dB
8
9 # add t u p l e d e f i n i n g p a t h r o o t with respect to the transmitted power and compensated FSPL.
10 γ . append (
11 ( root block index , root mpc index ) Delays and Doppler shifts are given with respect to LOS if
12 ) not denoted otherwise.
13
14 # i n i t MPC i n d e x f o r l o o p
15 int l = root mpc index
16 A. MPC Detection
17 # i t e r a t e over a l l subsequent blocks
18 f o r i n t b i n range ( r o o t b l o c k i n d e x , num blocks −2) :
19 1) Take-Off: Fig. 3 shows a color-coded representation of
20 # g e t l i s t o f u n a s s i g n e d MPCs i n n e x t b l o c k
21 l i s t a v a i l i n d i c e s = i n d e x ( Γ(b+1) == 0 )
the DDSFs of three consecutive blocks b ∈ {0, 1, 2} computed
22 according to the steps described in Section IV-B. The under-
23 # c h e c k i f a l l MPCs a r e a l r e a d y a s s i g n e d
24 i f empty ( a v a i l i n d i c e s ) : lying data were recorded during take-off and processed with
25
26 # they are → break loop a block size of B = 480, corresponding to a block duration
27 break of BN/fsr = 19.6 ms. The detected MPCs are highlighted by
28
29 # f i n d minimum d i s t a n c e byn a pn p l y i n g (18) oo
(b) (b+1)
red markers.
30 f l o a t dmin = min∀ l̃∈avail indices D ξl , ξ
l̃ All three figures show two main reflector areas: One that
31
32 # c h e c k i f min . d i s t a n c e i s a b o v e t h r e s h o l d corresponds to the LOS path and its close environment and
33 i f dmin > dthresh :
34 one that consists of multiple MPCs with a Doppler shift of
35 # i t i s → break loop around −1.1 kHz and a delay of 2.1 µs to 3.5 µs. A third, less
36 break
37 intense reflector area can be seen for a Doppler shift of around
38 # g e t i n d e x o f MPC w i t h min . d i s t a n c e and u s e
i t a s new MPC i n d e x f o r n e x t i t e r a t i o n 0.8 kHz and a delay of 4.0 µs to 4.3 µs. Only very few MPCs
39 i n t l = arg {dmin }
40 with a positive Doppler shift can be observed.
41 # add d e t e c t e d p a t h s e g m e n t
42 γ . append ( It can be observed, that most of the MPCs detected in block
43 ( b+1 , l ) b = 0 are also detected for b = 1 and b = 2 and vice versa.
44 )
45 However, this does not apply for some MPC: Especially those
46 # mark MPC a s a s s i g n e d
47 Γ(b+1) [ l ] = 1 having a delay of more than 5.8 µs tend to be missed in some
48 blocks during detection. At the same time, it can be observed
49 # return the l i s t d es cr i bi n g the path
50 return γ that these MPCs not only have a comparatively huge delay, but
Listing 2. Pseudo code of the detect path subroutine called by Listing 1. are also comparatively weak in power. This behavior motivates
the concept of Path Merging (see Section VI-C) during the
MPC tracking process.
2) Flyover: Fig. 4 also shows a color-coded representation
2) Short Path Elimination: MPCs that appear for the du-
of the DDSFs of three consecutive blocks b ∈ {0, 1, 2}. The
ration of just one block are likely to be the results of a mis-
underlying data were recorded just seconds before a direct
detection. In this case, the algorithm in Section VI-B assigns
flyover of the AS over the GS in an altitude of around 3.2 km.
these single MPCs to individual paths of length one. To only
All processing parameters were the same as for the take-off
keep track of MPCs that have been detected for a longer
scenario described above.
duration, a threshold βmin ∈ N can be set, that defines a
First, it can be observed that the noise floor is about 7.3 dB
minimum length of paths. Consequently, all paths that do not
stronger compared to the take-off scenario. We explain this by
fulfill
the applied AGC setting of the mixer, see Section II-A, during
recording: As the LOS distance – and consequently the FSPL
η (m) ≥ βmin (21) – was higher than during the recording of the data shown in
Fig. 3, a higher pre-amplification of the received signal was
required, including an increase of the noise floor.
are then eliminated. It can also be observed, that the LOS component is blurred
If the proposed MPC tracking approach is used for channel along the Doppler axis. We explain this by the fact, that the
modeling, those MPCs that have been assigned to distinct aircraft was moving, while the received data are processed
paths which have not been eliminated can be modeled by a in blocks. Theoretically, block processing as applied here
process that generates paths. However, the MPCs assigned to requires a perfect snapshot of the received data where all
one of the eliminated paths should not be ignored: We interpret parameters that affect the measurement remain exactly the
these MPCs as a result of diffuse scattering and suggest to same for the length of one block. For obvious reasons this
model them by a separate process. Thus, βmin is understood as is not possible in a real world scenario, especially when an
a parameter that has an influence on the amount of MPCs that aircraft is moving at cruising speed. However, to extract any
are modeled by the process that generates paths. The higher kind of Doppler information from measurement data, some
this amount, the smaller the amount of MPCs that are modeled sort of block processing is required as the Doppler can be
by the separate process. described as the time derivative of the delay. Thus, a trade-off
 9

2 2 2
b = 0, σ = −85.66 dB b = 1, σ = −85.90 dB b = 2, σ = −85.88 dB
0 0

2 −10

−20
4

Amplitude [dB]
−30
Delay [µs]

6
−40
8
−50

10
−60

12 −70

14 −80
−2 −1 0 1 2 −2 −1 0 1 2 −2 −1 0 1 2
Doppler [kHz] Doppler [kHz] Doppler [kHz]

Fig. 3. Color-coded representation of the logarithmic DDSFs of three consecutive blocks recorded during take-off. The strongest 50 MPCs that also exceed
a power 15 dB above the respective noise floor σ 2 are highlighted by red crosses. All blocks have been processed individually without any time dependent
data dependency.

2 2 2
b = 0, σ = −72.39 dB b = 1, σ = −72.39 dB b = 2, σ = −72.35 dB
0 0

2 −10

−20
4

Amplitude [dB]
−30
Delay [µs]

6
−40
8
−50

10
−60

12 −70

14 −80
−2 −1 0 1 2 −2 −1 0 1 2 −2 −1 0 1 2
Doppler [kHz] Doppler [kHz] Doppler [kHz]

Fig. 4. Color-coded representation of the logarithmic DDSFs of three consecutive blocks recorded during a flyover over the GS. The strongest 50 MPCs
that also exceed a power 8 dB above the respective noise floor σ 2 are highlighted by red crosses. All blocks have been processed individually without any
time dependent data dependency.

between available Doppler resolution and time precision has B. Path-based MPC Tracking
to be resolved.
Fig. 5 shows a 5 s cutout of the evolution of the MPC
The envelope of the detected MPCs shows a shape similar to components represented as paths. The MPCs were detected
a parabola. This is not a coincidence, but can be explained by using the approach shown in Section IV-B; the paths were
the fact, that in the air-ground channel, most of the reflectors found using the algorithm presented in Section VI using the
causing the MPCs are located on the ground comparatively delay-Doppler-only metric as defined in Appendix D. The
close to the GS. At the same time, each reflector is located underlying data were recorded while the aircraft was on the
on the surface of a prolate spheroid having the transmitter runway and accelerating for take-off. For the sake of clearness,
and receiver in its focal points [20]. The size of the spheroid short paths having a length less than 50 blocks have been
a given reflector is located on is determined by the corre- excluded from the plots.
sponding MPC’s delay. As most of the reflectors that cause 1) Interpolated Data Points: Fig. 5a) shows the paths as
the MPCs are distributed on the earth’s surface, which can be they were detected by the algorithm including the application
modeled by a plane, the intersection of the prolate spheroid of the path improvements presented in Section VI-C. Due
and the surface results in a parabola-like shape. to the path merging, paths with missing blocks are part of
 10

a) Interpolated b) Interpolated and Filtered

12.5

10.0
Delay [µs]

7.5

5.0

2.5

0.0

1.0

0.5
Doppler [kHz]

0.0

−0.5

−1.0

−10
Amplitude [dB]

−20

−30

−40

−50
0 1 2 3 4 5 0 1 2 3 4 5
Observation Time [s] Observation Time [s]

Fig. 5. A set of MPC-paths showing the evolution of delay (top), Doppler (middle), and the absolute value of the weight (bottom) of the detected MPCs
during take-off. In both a) and b) missing values were interpolated using a cubic spline interpolation where possible. In b), an additional low-pass filter has
been applied to the data of a) to smooth the evolution and to reduce the impact of outliers. For the sake of clearness, the plots show only paths with a
minimum length of 50 blocks.

the detected path set. Here, a cubic spline interpolation was either constructively or destructively.
applied to estimate the values of these missing blocks.
Most of the Doppler shifts increase in value over time by 2) Filtered Data Points: Fig. 5b) shows low-pass filtered
becoming more negative.. This is caused by the increasing versions of the paths presented in Fig. 5a). The purpose of
speed (and therefore higher absolute Doppler shift) of the the filtering is to smooth the curves to compensate the effects
aircraft during take-off and by the relative movement of the introduced by the processing, e.g. mis-detection of MPCs in
aircraft with respect to the reflectors like airport buildings, the DDSF or, most prominent, the unsteady evolution of the
trees and fences. Doppler shifts due to the low resolution. Another cause could
The figure shows no paths with a positive Doppler shift. be erroneously assigned MPCs: If dthresh is chosen very large,
This matches our observations from Fig. 3, where no strong some MPCs may be assigned to the same path although they
MPCs with a positive Doppler shift were detected. In both are of different origin and therefore should be assigned to
cases, the underlying data was recorded during the same take- separate paths.
off; however, the data shown in Fig. 3 was recorded shortly
after the data presented in Fig. 5. As neither the aircraft nor the reflectors are expected to do
The Doppler shifts, especially the shifts of the paths with the sudden movements that would explain these effects, it is a
lowest Doppler shifts, show a step-wise unsteady evolution. reasonable assumption that these effects are just processing
From the plot this step size can be determined as 25.4 Hz artifacts in most of the cases.
which corresponds to the Doppler resolution of the DDSFs the The impact of the filtering can be clearly seen in the plot
underlying MPCs got extracted from. We therefore consider of the Doppler shifts: The step-wise leaps that dominated the
these steps not as a natural effect, but a processing artifact that plotted paths in the unfiltered case are now gone and the curves
will be addressed below. look much smoother.
The amplitudes of the detected components, including the
LOS component given in orange, show a typical small scale The effect on the delay and the amplitudes, respectively, is
fading behavior. We assume this fading of the individual not as strong as for the Doppler shifts; however, the paths now
components is caused by non-resolvable MPCs that add up look smoother than before.
 11

C. Reflector Localization approach. In a last step, we have applied our processing chain
To validate the processing described above, the results of to our measurement data and performed a reflector localization
the MPC processing and tracking have been used to perform a based on the detected and tracked MPCs. We have shown that
reflector localization according to Appendix E. The estimated our approach can be used to locate reflectors that are causing
reflectors’ locations are then compared to the actual location MPCs that have a significant impact on the wireless channel.
of potential reflectors given in a map. In a next step we want to evaluate more of our measurement
Fig. 6 shows a map of the EDMO airport. The position of data recorded during other flight scenarios and use the results
the transmitter of the channel sounding signal is marked by a for a reflector localization – depending on the scenario also
red cross. The track the aircraft has traveled while the data that in three rather than in just two dimensions – and finally for
got evaluated for the reflector localization has been recorded is channel modeling.
denoted by a blue line; the aircraft started heading south-west. The authors are also planning to apply the presented MPC
The path based reflector localization works by superimpos- tracking approach to measurement data collected during other
ing the results of the reflector localization for each individual campaigns, not necessarily limited to the aeronautical air-
MPC of an MPC path. The longer a path is, i.e. the more ground channel.
persistent an MPC is, the better will be the result of the
estimation, as outliers and even ambiguities are averaged out A PPENDIX
due to the superimposition. In the figure, the certainty about A. Calculation of the Impulse Response
the position of a reflector is represented by the data point’s
The Impulse Response (IR) is computed by correlating
opacity. A color code is used to let the reader distinguish, what
the transmitted signal, here denoted by the reference signal
data points belong to a common path. However, as the amount
xref ∈ CN , and one instance of the channel sounding sequence
of detected paths exceeds the amount of available colors by
in the measurement data, either denoted by y ∈ CBN
far, the colors are not exclusively mapped to a specific path.
or Y ∈ CN ×B . We implement the correlation in frequency
While the comparatively high delay resolution allows a
domain, and assume a block processing of B consecutive
pretty precise estimation of the ellipse the reflector is located
impulse responses2 . The computation of the IR either maps
on, the estimation of the angle suffers from the comparatively
CBN 7→ CN fup ×B or CN ×B 7→ CN fup ×B , depending on the
low Doppler resolution. Although the superimposition helps
input data. In case of a vector input, y ∈ CBN is converted to
to improve the precision of the angle estimation, the shape of
a matrix Y ∈ CN ×B where the item in the n-th row and the
the data point clusters shows the lack of precision of the angle
m-th column yn,m is set according to yn,m = y[mB + n].
estimation. However, in most of the cases the algorithm was
After this optional reshaping, the processing continues for
able to resolve the ambiguities caused by the Doppler-to-angle
both cases as follows.
conversion: The actual reflector’s position has been identified,
while the mirrored reflector’s position got averaged out. 1) A matrix X ref ∈ CN ×B which elements are set to
In the following, we want to discuss a few of the most xn,m = xref [n] ∀ m ∈ {0, 1, ..., B − 1} is created based
prominent reflectors highlighted in the map. on the reference signal; the Fast Fourier Transform
The reflector at A in the south has been identified as a shelter (FFT) is computed: X 0ref := FFTN,↓ {X ref }.
with a metal roof. The reflectors around B have been identified 2) The FFT along each column of Y is computed, resulting
as a fence that is part of the airfield barrier. in a matrix Y 0 := FFTN,↓ {Y }.
Many reflectors have also been found close to the aircraft 3) An element-wise multiplication K 0 := Y 0 · X 0ref is per-
along the runway: The reflectors both in C and D are very formed; the result is expanded by N (fup − 1) rows of
persistent and have been identified as airport service buildings. zeros in case upsampling is requested.
4) We then perform an Inverse Fast Fourier Transform
The reflectors in E, F, and G are most likely related to
(IFFT): K := IFFTN,↓ {K 0 }.
the railway lines along which they have been detected. Their
origin might be the power lines along the railway, the rails or The n-th column of K ∈ CN fup ×B now corresponds to the
even a train that passed by during recording. IR hn of the n-th instance of the channel sounding sequence
A few reflectors have been detected in the mining area north in the input data, upsampled by fup .
of the highway at H; the reflector in I is located between a Assuming that xref and y or Y , respectively, are sampled
parking lot and an intersection with traffic lights. at a rate of fsr , the delay resolution of the IRs is given by
The reflectors in J and K are located in the suburbs north ∆τ := (fsr fup )−1 . (22)
of the highway and are most likely some higher buildings.
East of the runway we have highlighted the building in L,
as it is also the origin of a quite persistent reflector. B. Calculation of the Power Delay Profile
We denote the function computing the Power Delay Pro-
f
VIII. C ONCLUSION AND O UTLOOK file (PDP) by PDPxupref . It either maps CBN 7→ CN fup or
N ×B N fup
C 7→ C , where xref denotes the reference signal and
In this paper we have described a processing chain for the
fup denotes the upsampling factor, depending on the input data.
evaluation of channel sounding data recorded during flight. We
have explained our pre-processing and the actual MPC extrac- 2 The computation of a single IR therefore corresponds to the special case
tion process before introducing a path-based MPC tracking of B = 1.
 12

Fig. 6. Map of the EDMO airport with estimated positions of reflectors: The red cross marks the position of the transmitter, the blue line on the runway
shows the track of the aircraft during the recording of the displayed data. Each reflectors’ position was estimated by evaluating the nodes of the corresponding
path as given in Fig. 5b) according to Appendix E. Only paths of a minimum length of 50 blocks were considered. The locations of a subset of the most
prominent reflectors are highlighted by red circles. Background image: © by Google.

The computation of the PDP is based on B consecutive D. MPC Distance Metrics

Impulse Responses (IRs), thus we assume the IR matrix K The function D : R4 × R4 7→ R returns the distance metric
has been computed according to Appendix A. between two Multipath Components (MPCs) ξn and ξm .
The complex-valued PDP is then computed by taking the The first step is to perform a feature scaling, as the typical
mean of matrix K along its rows: aC := K̄ → . values and units of the items of an MPC vary. If not denoted
The real-valued PDP is given by the complex-valued PDP’s otherwise, we apply the popular min-max scaling to all ele-
absolute value: aR := K̄ → . ments x of an MPC:
Please note that the definitions above correspond to the
∀x∈ξ:
coherent PDP. Its delay resolution is given by (22).
x − xmin
x0 = , (24)
xmax − xmin
C. Calculation of the Doppler-Delay Spreading Function where x{min,max} denote a minimum/maximum value, respec-
tively, that are used for the scaling of the specific item. They
We denote the discrete Doppler-Delay Spreading Function
are chosen such that all appearing values of the specific item
(DDSF), sometimes also called Cross Ambiguity Function
are mapped to a value within [0, 1).
(CAF), using reference signal xref and upsampling fup by
f 1) Delay-Only Metric: In this case, only the MPCs’ delay
Sxupref . It either maps CBN 7→ CN fup ×B or CN ×B 7→ CN fup ×B ,
is considered. This metric is useful in case no Doppler
depending on the input data.
information is available and huge variations in the MPCs’
First, the B Impulse Responses (IRs) are computed accord- amplitude are expected. We define it as
ing to Appendix A, resulting in the IR matrix K. p
Then, an FFT along the rows of K is performed: D {ξn , ξm } = (τn0 − τm 0 )2 , (25)
A := FFTB,→ {K}. We assume that this FFT includes an 0
where τ{n,m} are feature scaled (24) versions of τ{n,m} ∈
FFT-shift, such that the 0-th frequency is shifted to column ξ{n,m} .
B/2. 2) Delay-Doppler-Only Metric: In this case, the MPCs’
The columns of the resulting matrix A correspond to the complex weight is ignored. This metric is useful in case huge
delay axes, the rows correspond to the Doppler axes. While variations in the MPCs’ amplitude are expected. We define it
the resolution along the delay axis is the same as for the IRs as
as given in (22), the resolution along the Doppler axis is given p
by D {ξn , ξm } = (τn0 − τm0 )2 + (ν 0 − ν 0 )2 ,
n m (26)
0 0
fsr where {τ{n,m} , ν{n,m} } are feature scaled (24) versions of
∆ν := . (23) {τ{n,m} , ν{n,m} } ∈ ξ{n,m} .
NB
 13

E. Two-Dimensional Reflector Localization as given above. Instead of computing the intersection of the
The delay and Doppler information of an MPC can be used sets as given in the last step, the exploitation of the evolution
to estimate the corresponding reflector’s location. Here, we to reduce ambiguities is realized by computing the mean of
briefly describe the implementation of the reflector localization the matrices of each path node. This approach is more robust
for the two-dimensional case (i.e. no height information is against outliers as erroneous detections are averaged out.
provided).
1) Exploiting Delay Information: All (potential) reflec- R EFERENCES
tors with the same delay are located on an ellipse whose [1] Fortune Business Insights, “Unmanned Aerial Vehicle (UAV)
Market Size, Share & COVID-19 Impact Analysis,” Jul.
focal points are given by the position of the transmit- 2020. [Online]. Available: https://www.fortunebusinessinsights.com/
ter (Tx) and the receiver (Rx). It is straight forward to industry-reports/unmanned-aerial-vehicle-uav-market-101603
compute this set of points based on a given delay τ :
[2] M. Webster, M. Fisher, N. Cameron, and M. Jump, “Formal methods
for the certification of autonomous unmanned aircraft systems,” in 30th
Eτ = P | TxP + P Rx = cair τ , where cair denotes the speed International Conference, SAFECOMP, 09 2011, pp. 228–242.
of light in air. [3] “Report itu-r m.2171,” ITU, 12 2009.
If we assume an MPC-detection with a delay resolution of [4] M. Pätzold, Mobile Fading Channels. New York: Wiley, 2002.
[5] D. W. Matolak and R. Sun, “Air-ground channel characterization for
∆τ , we can expand this to an inner and an outer ellipse. The unmanned aircraft systems - part i: Methods, measurements, and models
reflector’s position is then lying in the set: for over-water settings,” IEEE Transactions on Vehicular Technology,
vol. 66, no. 1, pp. 26–44, Jan 2017.
   [6] R. Sun and D. W. Matolak, “Air-ground channel characterization for
∆τ unmanned aircraft systems - part ii: Hilly and mountainous settings,”
Eτ = P TxP + P Rx Q cair τ ± . (27) IEEE Transactions on Vehicular Technology, vol. 66, no. 3, pp. 1913–
2
1925, March 2017.
2) Exploiting Doppler Information: The angle of arrival α [7] D. W. Matolak, “Air-ground channels & models: Comprehensive review
and considerations for unmanned aircraft systems,” in 2012 IEEE
with respect to the receiver’s heading and speed ~vRx can be Aerospace Conference, 2012, pp. 1–17.
estimated based on the Doppler shift ν according to: [8] P. Hoeher and E. Haas, “Aeronautical channel modeling at vhf-band,”
  in Gateway to 21st Century Communications Village. VTC 1999-Fall.
cair ν IEEE VTS 50th Vehicular Technology Conference (Cat. No.99CH36324),
α = ± arccos . (28) vol. 4, 1999, pp. 1961–1966 vol.4.
fc |~vRx |
[9] E. Haas, “Aeronautical channel modeling,” IEEE Transactions on Ve-
The reader may note the non-resolvable ambiguity caused hicular Technology, vol. 51, no. 2, pp. 254–264, Mar 2002.
[10] P. Pulini, “Forward link performance analysis for the future ieee 802.16-
by the nature of the cosine function. In case of a Doppler based airport data link,” in 2010 IEEE International Conference on
resolution of ∆ν , the set of points that arrive at the receiver Communications, 2010, pp. 1–5.
within the given angle range is given by [11] I. Sen and D. W. Matolak, “The 5-ghz airport surface area channel -
part ii: Measurement and modeling results for small airports,” IEEE
Eν = E−ν ∪ E+ν , where Transactions on Vehicular Technology, vol. 57, no. 4, pp. 2027–2035,
( ( )) July 2008.
cair (ν ∓ ∆2ν ) [12] N. Schneckenburger, T. Jost, D. Shutin, M. Walter, G. del Galdo, and
E−ν = P ∠(P Rx, ~vRx ) Q − arccos U. Fiebig, “Reflector localization for geometrical modeling the airground
fc |~vRx | channel,” IEEE Transactions on Vehicular Technology, vol. 67, no. 9,
pp. 7994–8008, 2018.
and [13] N. Schneckenburger, T. Jost, M. Walter, G. del Galdo, D. W. Matolak,
( ( ))
and U. Fiebig, “Wideband airground channel model for a regional airport
cair (ν ± ∆2ν ) environment,” IEEE Transactions on Vehicular Technology, vol. 68,
E+ν = P ∠(P Rx, ~vRx ) Q + arccos .
fc |~vRx | no. 7, pp. 6243–6256, 2019.
[14] J. Kim and I. Lee, “Channel measurements and characterizations for
(29) long range air-to-ground communication systems in the uhf band,” IEEE
Access, vol. 10, pp. 101 880–101 888, 2022.
3) Localization: The final position of the reflector is lying [15] D. W. Matolak and R. Sun, “Air–ground channel characterization for
in the intersection of the sets E = Eτ ∩ Eν . By combining the unmanned aircraft systems—part iii: The suburban and near-urban
information gained through the processing of multiple blocks environments,” IEEE Transactions on Vehicular Technology, vol. 66,
no. 8, pp. 6607–6618, 2017.
recorded over time, the uncertainty over the reflector’s position [16] D. M. Mielke, N. Schneckenburger, U.-C. Fiebig, M. Walter, and M. A.
can be reduced. Bellido-Manganell, “Analysis of the dominant signal component of
the air-ground channel based on measurement data at c-band,” IEEE
4) Exploiting Evolution: As it is assumed that an MPC’s Transactions on Vehicular Technology, vol. 70, no. 4, pp. 2955–2968,
signal is caused by a reflector, its localization can be improved 2021.
by observing the evolution of the corresponding MPC repre- [17] N. Schneckenburger, “A wide-band air-ground channel model,” Ph.D.
dissertation, University of Ilmenau, 2017.
sented by a path γ. This can be done by performing the steps [18] M. Powell, “The bobyqa algorithm for bound constrained optimization
described above for each
T node of the path and finally joining without derivatives,” DAMTP 2009/NA06, 08 2009. [Online]. Available:
these results: Eγ = E (b) . https://www.damtp.cam.ac.uk/user/na/NA papers/NA2009 06.pdf
∀b∈γ [19] D. Shutin and G. Kubin, “Tracking and prediction of multipath
components in wireless mimo channels,” in Proceedings of the 65th
5) Implementation: In practice, the sets that are used to IEEE Vehicular Technology Conference VTC2007, 2007, pp. 229–333.
locate the reflector are represented by a two-dimensional [Online]. Available: papers/vtc2007.pdf
matrix, where the rows correspond to the east-west axis and the [20] M. Walter, D. Shutin, M. Schmidhammer, D. W. Matolak, and A. Zajic,
“Geometric analysis of the doppler frequency for general non-stationary
columns correspond to the north-south axis of an East-North- 3d mobile-to-mobile channels based on prolate spheroidal coordinates,”
Up (ENU) coordinate system with the transmitter in its origin. IEEE Transactions on Vehicular Technology, vol. 69, no. 10, pp. 10 419–
For each node of an MPC’s path, the matrix is determined 10 434, 2020.
 14

Daniel M. Mielke is with the Institute of Commu-

nications and Navigation of the German Aerospace
Center (DLR) since 2016. He received his B.Sc.
and his M.Sc. from the University of Kiel in 2014
and 2016, respectively. He is currently enrolled as
a PhD candidate at the University of Ulm. His
research interests include radio wave propagation
and channel modeling in the field of aeronautical
communications as well as signal processing and
robust communication systems.

Michael Walter received his Dipl.-Ing. and Dr.-Ing.

degree in electrical engineering from Ulm Univer-
sity, Germany, in 2008 and 2015. Since 2009 he has
been a member of the scientific staff at the Insti-
tute of Communications and Navigation at the Ger-
man Aerospace Center (DLR) in Oberpfaffenhofen,
Germany, where he gained extensive experience in
organizing and conducting numerous measurement
campaigns for vehicular and aeronautical research.

Dennis Becker received his Bachelor and Mas-

ter degrees in Electrical Engineering from Duale
Hochschule Baden-Württemberg in 2011 and from
Technische Universität Darmstadt in 2014 respec-
tively. He continued studying Information and Com-
munications Engineering and obtained his second
Master degree in 2017. In January 2018, Dennis
joined the Institute of Communication and Naviga-
tion at the German Aerospace Center (DLR) as a re-
search assistant in the Aeronautical Communications
group. His current research focus is on analyzing
signal propagation characteristics of the communications channel between
small sized unmanned aerial vehicles.

Miguel A. Bellido-Manganell received his B.Sc.

and M.Sc. degrees in telecommunication engineer-
ing from University of Granada and University of
Malaga, Spain, in 2014 and 2016. In 2016 he joined
the Institute of Communications and Navigation of
the German Aerospace Center (DLR) as a research
associate. He has participated in multiple measure-
ment campaigns and had a leading role in the
2019 LDACS flight campaign. His research interests
include channel modeling, physical layer design and
digital signal processing in wireless communica-
tions, as well as medium access control in ad hoc networks.

Uwe-Carsten Fiebig studied electrical engineering

at the Technical University of Munich, Germany, and
spent the last year of his studies at ESIEE in Paris,
France. In 1993 he received the Dr.-Ing. degree
from the University of Kaiserslautern, Germany.
He joined the Institute of Communications and
Navigation at DLR (German Aerospace Research),
Germany, in 1988. As head of department he is
responsible for research in aeronautical and vehic-
ular communications, multi-sensor navigation and
swarm exploration. A special focus is on propagation
and communications for aircraft, drones, ships and trains. Since 2011 he is
honorary professor at the University of Ulm, Germany, and teaches satellite
communication and navigation.