Algorithm for Electromagnetic Power Estimation in Radio

Environment Map
ABSTRACT

Perhaps the most important part of building radio environment map; is the
estimation of the electromagnetic field strength. The more efficient estimation
algorithm used; the more accurate radio environment map reached. A new hybrid
algorithm to estimate the electromagnetic power - using sensing data gathered by
monitoring sensors - is proposed in this paper. A certain propagation model was
used considering all physical phenomenon take place along the whole path of the
electromagnetic wave (including: losses, attenuation, etc.).A theoretical variogram
based on a certain propagation model was used, and fitted using real variogram
through regression. Apparently this method simulates the real physical
phenomenon more accurately. The proposed algorithm mainly based on weighting
parameters (that relies in variogram method) and a factor taking the similarities
because of the neighbourhood into consideration. Experimental results showed
close similarity between given and computed electromagnetic power through both:
regression curves and objective evaluation indices. The main contribution of the
proposed method is gathering the merits of using suitable propagation model and a
theoretical variogram, besides getting merits of traditional methods(like Inverse
Distance Weight) using new vision.

1-INTRODUCTION

The radio environment map is an integrated and intelligent database system that
supports cognitive radio, Zhao [1].

This database contains different information from many different domains;

including:

I. Geographical features.
II. Services availability.
III. Spectral regulations, rules, and procedures.
IV. Past, present, and expected future (by studying past experience) situation of
equipment.
V. Locations and activities of radio equipment.

Page 1 of 18
VI. Policies of service providers and users through the environment.

The main objective of radio environment map is the efficient and optimized
management of the radio spectrum. Many researchers like Ojaniemi [2], proved
that , the most important part of constructing an accurate radio environment map ;
is to develop an algorithm to estimate the radio parameter (electromagnetic power)
since it is the main factor to be determined throughout the whole concept in order
to make and take the right decision concerning both ; primary and secondary
users. This estimation problem is called by Pesko [3] as radio frequency layer.

There have been extensive researches concerning this problem since 2012.Some
researchers used the inverse distance method (IDW) like J.Riihijavri, P.Mahonen,
W.Wellens and M.Gordziel [4], 2008 .and Two researches were held in 2012, one
of them was introduced by D.Denkovski, V.Atanasovski, L.Gavrilovska,
J.Riihijavri and P.Mhonen [5].The other one was proposed by C.Phillips, M.Ton,
D.Sicker, and D.Grunwald [6].the authors used Kriging method .In 2015, a paper
was proposed by S.Ulanganathan, D.Deschrijver, M.Pakparvar, J.Coukuyt, W.Liu,
D.Plets, W.Joseph, D.Dhaene, L.Martens and I.Moerman [7], used antiregressive
Co Kriging model. In 2017, K.Sato and T.Fujii [8], proposed a new method for the
concept of spatial spectrum sharing, and the occurring of radio environment map, it
was improved using Kriging interpolation. Although Kriging method thought of –
by many researchers- as the most accurate method, however many other
researchers- on the other side- overviewed the limitations of the method, as
Ojaniemi [2]. He proved the poorness of the method under certain conditions.

Apparently Kriging model treats the whole data as normally distributed; which is
not always the case. It does not take observations like sill and nuggets and
generally all the outliers into consideration properly. Hence, to overcome this
problem a new model introduced, using both, Kriging and propagation model.

The theoretical Kriging variogram is improved by using a certain propagation

model suitable for considering outliers, then regression was used to optimize the
variogram, resulting in in a modified Kriging model that works much better than
traditional one.

2- EXPERIMENTAL SECTION

Page 2 of 18
 2-1 THE ALGORITHM

For any random process S ( d )

{ S ( d ) :d ∈ D∁ Rn } (1)

Where
d : Is a location in D

Rn ( n=1,2,3 ) : Is an n dimensional Euclidean space.

Spatial data expressed as ( Pi , d i ) ,i=1,2,3 , … , N

Where Pi: is the i th observation of the phenomenon (here, it is the strength of

electromagnetic field) of interest in location d i .

For the process to be stationary random process:

i. E ( P ( d ) )=μ , ∀ d ∈ D
ii. Cov ( P ( di ) , P ( d j ) ) =C ( h )=C ( d i−d j ) < ∞
iii. Var ( P ( d i )−P ( d j ) ) =2 γ ( d i−d j )=2 γ ( h ) ∀ di , d j ∈ D
Where:
μ: is the expected value of the phenomena.
C ( h ) And 2 γ ( h )are the covariance and variogram for distant h pair of points,
respectively.
A block diagram for the algorithm is shown below in figure (1)

INPUT ALGORITHM OUPUT

1. Sample points’ coordinates

X n∗1 , Y n∗1 , Z n∗1 .
Estimated (expected) value
2. Phenomena observed
of the phenomena
values Pn∗1 .
P¿t
3. The coordinates of the
point of interest (to
estimate the phenomena in) Figure (1)

( X t ,Y t , Z t ) Page 3 of 18
 A block diagram shows inputs and outputs of the algorithm.

And all necessary steps are shown in figure(2)

1- Begin

2-Specify the pairs’ number, and configure the

distance distribution

3-Is the pairs’ number sufficient (in this lag category)?

4- Compute “lags” and

“max_distance”.

5- Calculate the experimental variogram.

6- Choose variogram model and parameters .

7- Use the selected theoretical variogram model to fit the

experimental variogram. Page 4 of 18
 8- Do theoretical and experimental
variograms agree?

9- Achieve modified (proposed) Kriging method using selected theoretical

variogram model and parameters.

10- End

Figure (2)

Shows detailed steps of the algorithm

2-2 THE VARIOGRAM

It is an autocorrelation measure used in geostatistics, given by:

N
γ ( h )=
1
N ( h) √∑ (
i=1
2
P ( d i ) −P ( d i+ h ) ) (2)

Where
γ ( h ) : is the variogram function.

N ( h ) : is the number of data pairs with lag distance h .

P ( d i ) : is the value of electromagnetic power at position d i .

Page 5 of 18
P ( d i +h ) : is the value of electromagnetic power at position d i +h .

N : is the total number of data points.

2-3 THE THEORETICAL VARIOGRAM

A variogram based on Longely – Rice model (sometimes called the irregular

terrain model) is introduced.

Since:

i. Heights of antennae are known.

ii. Loss of free space propagation considered.

Theoretical variogram will be:

γ ( h )=∝1 +∝2 d +∝3 ln ( h∅ + δ )+ ∝ log

4 10 ( h+ δ ) (3)

Where
h : is the distance between two data points.

δ : is a very small constant (to overcome division by zero problem)

∅ : used to simulate distance of sight (taken as 14000).

∝1 , ∝2 , ∝3 ,∧∝4 : are the coefficients to be determined.

2.4 Fitting algorithms

Three weight coefficients used to overcome certain three problems as follows:

a. Random distribution of sampling points :

N
A weight coefficient λ 1= N (4) ,is used to overcome the sampling randomness,
i

where:
N i : is the number of sample point pair correspond to certain lag.

N : is the total number of sample number pair.

Page 6 of 18
 b. Inconsistencies and Inaccuracies

This problem caused by electromagnetic shadowing and reflections of buildings,

multipath, and radio propagation diffractions(abnormally large or small sampled
values).to deal with this issue; a second weighting coefficient is used ,
γ̌ ( h )
λ 2= (5)
γ ( hi )

Where
γ̌ ( h ) : is the mean value of the variogram.

γ ( hi ) : Is the value of the variogram at lag distancehi .

c. Reflection of variability of regionalized variable

This is achieved by considering points of smaller lag distance. Hence, a third

ȟ
weighting coefficient used, which is λ 3= h (6) , where:
i

ȟ : is the mean value of the lag distance.

hi : is the corresponding lag distance.

Thus the total weight will be: λ total =λ1 . λ2 . λ3 (7)

Now Ꞃ and Ꞃt calculated using theoretical variogram model

γ 11 γ 12 γ 13 … γ 1 n 1 γ 1t λ1

[
Ꞃ = ……………
]
γ 21 γ 22 γ 23 γ 2 n 1

¿ γ n1 γ n 2 γ n 3 … γ nn 1
11 1 0
Ꞃ
[] []
γ 2t
(8) and t = … (9) and λ =
¿ γ nt
1
λ2
… (10)
¿ λn
µ

Ꞃ* λ =Ꞃt , hence λ = Ꞃ−1 * Ꞃt

n
P¿t =∑ λi Pi (11)
i=1

Where:
Piis the power at point i.

Page 7 of 18
Ꞃ is the augmented total histogram matrix (between the whole n points).
Ꞃt is histogram matrix (between every point and the point of interest t.
¿
Finally the expected power at point t = a * Pt + b*(neighbourhood average power).

Where: neighbourhood average power is the power in the nearest point or the
average in the powers of the nearest points (if there is more than one point) .

Given: a + b = 1.

In this paper: a =b 0.5 are used.

2.5 DATA

Two sets of data taken at two certain frequencies, 263.2148 MHz, and 600 MHz, at
a calibrated chamber.
2.6 SIMULATION AND IMPLEMENTATION

Python version 3.4.0 was used to implement the proposed algorithm.

3. RESULTS AND ANALYSIS

3.1 Two cases were studied : 263.2148 MHz , and 600 MHz
3.2 The resultant expected power at the point of interest is taken as a half of
the power using weighted parameters algorithm and the other half using
the average power of the nearest points.
3.3 The results were compared with the well-known two algorithms of
Inverse Distance Weight and Inverse Distance Weight Squared.
3.4 Comparisons are held using regression graphs similarities between each
of the three methods (Algo (proposed whole algorithm), IDW (Inverse
distance weight algorithm), and IDWSQU (Inverse Distance Weight
Squared) compared with the real points’ powers.
3.5 Also five objective evaluation indices were used :

3.5.1Average Estimation Error Percentage:

Page 8 of 18
 n
1
AEEP = ˇ ∑ ❑ ( P¿t −Pgiveni ). (12)
n∗Pgiven i=1

3.5.2Relative Mean Square Error:

n
1 2
RELMSE = 2∑
❑ ( P¿t −Pgiven i ) . (13)
n∗S i=1

3.5.3Root Mean Square Error:

√
RMSE = 1 ∑ ❑ ( P¿t −Pgiveni )2 . (14)
n i=1

3.5.4Maximum Error:
¿
Max_Err = Maximum Value of( Pt −Pgiveni ). (15)

¿
3.5.5 Av_Err = Average Value of( Pt −Pgiveni ). (16)

Where:
n : is the total number of sampling points.
ˇ
Pgiven : is the average value of the real sampling points’
powers.
P¿t : is the estimated power of the point at position i.
Pgiveni : is the real power of the point at positions i.
S2 : is the variance of the real sampling points’ powers.

3.6 figure (5) shows the curves of real powers and expected powers using Inverse
Distance Squared algorithm of the whole fifty points (treated as unknowns and
then estimated using Inverse Distance Squared algorithm) taken at frequency of
263.2148 MHz.
Figure (5)

Shows the curves of real powers and expected powers using IDWSQU at f=263.2148MHz.

Page 9 of 18
 10

0
1 4 7 1 0 13 16 1 9 22 2 5 2 8 3 1 34 3 7 40 4 3 46 4 9
-10

-20

-30 p given
-40 IDWSQU

-50

-60

-70

-80

3.7 figure (6) shows the curves of real powers and expected powers using Inverse
Distance algorithm of the whole fifty points(treated as unknowns and then
estimated using Inverse Distance algorithm) taken frequency of 263.2148 MHz.
Figure (6)

Shows the curves of real powers and expected powers using IDW at f=263.2148MHz.

10

0
1 4 7 10 13 16 1 9 2 2 25 28 31 3 4 3 7 4 0 43 46 49
-10

-20

-30 p given
-40 IDW

-50

-60

-70

-80

3.8 figure (7) shows the curves of real powers and expected powers using Hybrid
algorithm proposed in this paper of the whole fifty points (treated as unknowns and
then estimated using the proposed algorithm)taken at frequency of 263.2148 MHz.

Page 10 of 18
 Figure (7)

Shows the curves of real powers and expected powers using ALGO at f=263.2148MHz.

20

10

0
1 3 5 7 9 1 1 13 15 17 1 9 21 2 3 2 5 27 29 31 3 3 35 3 7 3 9 4 1 43 45 47 4 9
-10

-20

p given
-30
ALGO

-40

-50

-60

-70

-80

3.9 table (1) compares the five objective evaluation indices defined in 3.5 for
the three algorithms at frequency of 263.2148 MHz.
Table (1)

Compar ALGO IDW IDWSQU

es the
five
objectiv
e
evaluati
on
indices
defined
in 3.5
for the
three
algorith
ms at

Page 11 of 18
 frequen
cy of
263.214
8
MHz.X 
PAEE -7.979787889 -20.83185285 -20.56436201
       
RMSE 0.003658915 0.009551882 0.009429231
       
RMSPE 6.384915063 10.31628663 10.24983964
       
MAX-
ERR 34.04258976 63.62412487 63.35257744
       
AVG-
ERR 4.365355635 6.204105244 4.995278133

3.10 figure (8) shows the curves of real powers and expected powers using
Inverse Distance Squared algorithm of the whole fifty points (treated as
unknowns and then estimated using Inverse Distance Squared algorithm)taken
at frequency of 600 MHz.

Figure(8)

Shows the curves of real powers and expected powers using IDWSQU at f=600MHz.

Page 12 of 18
 0
1 4 7 10 13 16 19 22 25 28 31 34 37 40 43 46 49
-5

-10

-15

-20

-25
p given
-30 IDWSQ
U
-35

3.11figure (9) shows the curves of real powers and expected powers using
Inverse Distance algorithm of the whole fifty points (treated as unknowns and
then estimated using Inverse Distance algorithm)taken frequency of 600 MHz..
Figure (9)

Shows the curves of real powers and expected powers using IDW at f=600MHz.

0
1 4 7 10 13 16 1 9 2 2 25 28 31 3 4 3 7 4 0 43 46 49
-5

-10

-15
p given
IDW
-20

-25

-30

-35

Page 13 of 18
 3.12 figure (10) shows the curves of real powers and expected powers using
Hybrid algorithm proposed in this paper of the whole fifty points(treated as
unknowns and then estimated using the proposed algorithm) taken frequency
of 600 MHz.
Figure (10)

Shows the curves of real powers and expected powers using ALGO at f=600MHz.

0
1 3 5 7 9 1 1 13 15 1 7 19 21 23 2 5 2 7 2 9 3 1 3 3 3 5 37 39 41 43 45 47 4 9
-5

-10

-15

-20

-25

p
-30 given
ALGO
-35

3.13table (2) compares the five objective evaluation indices defined in 3.5 for the
three algorithms at frequency of 600 MHz.
Table (2)

Powers and three dimensional coordinates of fifty different points at frequency of 600 MHz

ALGO IDW IDWSQU

- -
AEEP -3.332002 3.773065 3.753388938
       
RELMSE 0.8918367 1.009891 1.004624179
3.14 Discussion and Analysis
        3.14.1 from 3.6 , it was noticed that : the
RMSE 5.9798869 6.363375 6.346760598
curve represents the estimated values of
       
Max_Er - -
r 1.8242065 10.10416 Page 14 of 18
10.05338129
       
- -
Avg_Err -0.448179 10.50261 10.53823175
powers using the IDWSQU algorithm compared with the curve of real powers at
frequency of 263.2148 MHz , does not resemble nor the pattern neither the values
in any way. The estimated values of powers curve using the IDWSQU algorithm is
more like a linear curve rather than a regression curve following a certain pattern
of distribution using a certain regression algorithm that depends on the inverse
relationship with the squared distance contribution, i.e., near points powers
contribute effectively into estimation of the unknown point’s power while far
distant points’ powers have neglectable contributions.

3.14.2 from 3.7 the same notice as in 3.14.1 is valid except that the curve of IDW
estimated powers is supposed to follow a certain distribution depending on the
inverse relationship with the distance (rather than the square of the distance as
above) contribution, i.e., far distant points’ powers have less more contribution
than above. However again the curve seems to be more linear and does not follow
the pattern of the real powers’ curve.

3.14.3 Now, from3.8 using the proposed Hybrid algorithm: the curve represents the
estimated powers using the Hybrid algorithm has exactly the same pattern of the
curve representing the real powers, and moreover even the values seem to be more
like their analogous values in the other curve. However it appears as if the Algo
curve needs a mathematical electromotive force to push it up to match the Pgiven
curve exactly.

Again we can notice that the similarity in the pattern appears clearly everywhere
except at the very beginning of the curve and at the end of the curve, i.e., at so far
distant points, although the similarity still exists.

Also it must be noticed that : for similarity of values , it appears to be very clear in
the range of powers between 10 and -10 ; and clearly the maximum error appears
to happen at the point where the power is near -70,which is far in the powers axis.

3.14.4 The table in section 3.9 shows clearly, that using those five objective
evaluation indices Algo has the lowest value for the whole five indices. Although
for the Av_Err, the IDWSQU algorithm has very close value to that of Algo,
whereas for Max_Err both IDWSQU and IDW have nearly twice the value of that
of Algo. Again for RMSE both IDWSQU and IDW have nearly one and a half of

Page 15 of 18
that of Algo. For RELMSE it is apparent that both well-known algorithms
IDWSQU and IDW have nearly three times that of Algo.

For AEEP the two algorithms IDWSQU and IDW values have even more than
twice that of our proposed algorithm Algo.

3.14.5 From 3.10, it is clear that the same discussion as in 3.14.1 could be repeated
here; however the operating frequency is 600 MHz instead of 263.2148 MHz.

3.14.6 Interpreting the graph in section 3.11, again the same comments in section
3.14.2 could be repeated, taking in mind that the frequency is 600MHz.

3.14.7 Investigating the graph in section 3.12, we can repeat the same comments as
of section 3.14.3, however here even for the far distant points, similarity of pattern
is pretty apparent. Also even for odd values of powers, the similarity in values is
clear.

3.14.8 For the table in setion3.13 we noticed the following:

For the AEEP, although the value for Algo is the lowest value, the other two values
for IDWSQU and IDW are very close to it.

Also for RELMSE the same above comment is valid.

One more time the same above comment could be repeated for RMSE.

However for, for Max_Err, the value for both IDWSQU and IDW appears to be
nearly six times that of Algo.

Again, for the Av_Err, the value for both algorithms IDWSQU and IDW reaches
twenty five multiples of that of Algo.

4. CONCLUSION AND RECOMMENDATIONS

The overall conclusion from the above discussion is the clarity of superiority of our
proposed algorithm over the traditional algorithms (as in this paper IDWSQU and

Page 16 of 18
IDW) in so many ways, as pattern, values, evaluation indices, and also in dealing
with odd points in both axes location and power.

Apparently in our proposed algorithm, we introduced a new weighted algorithm

depending on histogram concept, with an eye on traditional algorithms’ concepts
that take distance contribution into account in order to estimate the missing
parameter, but in a new way by taking another weight of the average power of the
nearest points to the point of interest , and the result is pretty good.

As a recommendation, different values for a and b could be taken (instead of a= b

0.5 as taken in this paper) and studied carefully in different frequencies and
circumstances.

Also it is recommended to use different techniques for the regression of the

powers’ curve and estimated value, for example machine learning is a powerful
tool that could be used extensively in this matter, and this may lead to more
accurate results.

Finally and importantly, our proposed algorithm could be used in so many

applications beside power parameter estimation in Radio Environment Map.

REFERENCES

[1]. Zhao, L.Morales , J.Gaaeddert, K.KBae , J.Sum and J.H.Reed , “applying

radio environment map to cognitive wireless regional area network”, 2007 , 2nd
IEE international symposuimon new Frontiers in dynamic spectrum access
network . Dublin, 2007 , pp115-118 .

[2] Ojaniemi, J.Kalliovaara,A.Alam,J.Poikonen and R.Wichman,”optimalfield

measurement design for radio environment mapping,”2013 47th Annual Conference
on Information Sciences and Systems (CISS),Bltimore,MD,2013, PP.1-6.

[3]. M.Pesko , T.Jovorink , A.Kosir , M.Sutlar , M.Mohorc ic , “radio environment

maps: survey of construction methods” ,ks11 transaction on internet and
information systems , vol -8 , no.11 , 2014 , pp 3789-3809.

Page 17 of 18
[4]. J.Riihijarvi, P.Mahonen, W.Wellens and M. Gordziel,“characterization and
modeling of spectrum for Dynamic spectrum Access with spatial statistics and
random field,”. In IEEE 19th international symposium on personal, indoor and
mobile Radio communications, Cannes PP.1-6.

[5]. D.Denkovski , V.Atanasovski, L.Gavrilovska, J.Riihijarive and P.mahonen “

Reliability of Radio Environment Map (REM) : case of spatial interpolation
technique ,” in the 7th international ICI conference On Cognitive Radio Oriented
Wireless Networks And Communications ( CROWNCOM),in Stockholm ,
pp.248-253.

[6]. C. Phillips, M.Ton, D. Sicker and D.Grunwald, “practical Radio environment

mapping with geo statistics,” in IEEE international symposium on Dynamic
spectrum Access Network, in Bellevue, wa, PP.422- 433.

[7]. S. Ulaganathan, D.Deschrijver, M. Pakparvar, I. CouKuyt, W.Liu, D. Plets, W.

Josph, T. Dhaene, L. Martens and I. Moerman “ Building accurate radio
environment map from Multi – Fidelity spectrum sensing data,” in wireless
networks, vol. 22, no.8, pp. 2551- 2562.

[8]. K.Sato, and T. Fujii, proposed a paper titled “Kriging – based interference
power constraint: integrated Design of the Radio Environment Map (REM) and
transmission power,” in IEEE transactions on cognitive communications and
Networking, pp (99), pp. 1-1.

Page 18 of 18