Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
https://doi.org/10.1007/s10291-020-0964-6
ORIGINAL ARTICLE
Abstract
The total electron content (TEC) is one of the most important parameters for studying the behavior of the ionosphere. The
global ionosphere maps (GIMs) can be used to study the TEC time series variations. The time resolution of the GIM-TEC
is 2 h, whereas the frequency of the ionospheric temporal behavior can be less than 2 h. To solve this problem, we present
a new method for ionosphere time series modeling and prediction in Iran. The adaptive neuro-fuzzy inference system
(ANFIS) and principal component analysis are combined to model the TEC of the ionosphere. In fact, the observations are
decomposed into principal components before entering to the ANFIS network and only a few main components are used
for training the network. The main advantage of this combination is to increase the accuracy of the results and reduce the
time of convergence to achieve an optimal solution. To evaluate the proposed method, we used observations of a Tehran
GNSS station in 2016 and 2017. The root-mean-square error, correlation coefficient, and dVTEC = |VTECGPS − VTECmodel|
were used to assess the accuracy of the proposed method. Also, all results are compared with the International Reference
Ionosphere 2016 (IRI2016), GIM-TEC, and artificial neural networks (ANNs) ionosphere models. The results indicate a 1 to
4.72 TECU improvement in the temporal resolution of TEC modeling with the proposed method, compared to the IRI2016,
GIM, and ANNs in the Iranian region.
13
Vol.:(0123456789)
51 Page 2 of 13 GPS Solutions (2020) 24:51
training algorithms (Ma et al. 2005; Habarulema et al. 2009; optimize its algorithm, so in some cases, ANFIS might be
Hirooka et al. 2011; Ghaffari Razin et al. 2015; Ghaffari better than ANN.
Razin and Voosoghi 2017). The main objective is to model the time series of the total
The ionosphere working group of the international electron content of the ionosphere in Iran using ANFIS.
GNSS service (IGS) has been a reliable source of global Because in some cases, like in ionospheric time series mod-
ionospheric maps (GIMs) since 1998 (Hernández-Pajares eling, the number of input observations to the algorithm is
et al. 2009). This group consists of several ionosphere analy- very high, the use of a principal component analysis (PCA)
sis centers, such as the Center of Orbit Determination in method is proposed. Using the PCA, we can use the prin-
Europe (CODE), the Jet Propulsion Laboratory (JPL), and cipal frequencies in the ionosphere time series as inputs to
the Technical University of Catalonia (UPC) (Mannucci the ANFIS network. Therefore, it is possible to model the
et al. 1998; Hernández-Pajares et al. 1999). The GIM model ionospheric behavior with high precision and high tempo-
covers ± 87.5° latitude and ±180° longitude ranges. It also ral resolution using several principal components. Also, the
has 2.5° and 5° spatial resolution in latitude and longitude, convergence speed to the optimal solution increases in the
respectively. The accuracy of GIM-TEC was evaluated to ANFIS network. In order to evaluate the proposed method,
be about 2–7 TECU (Sekido et al. 2003). The IGS track- observations of Tehran GNSS station in 2016 and 2017 are
ing network has over 500 active GNSS stations worldwide. used. Also, all results are compared with the International
But the distribution of these stations is not homogeneous. Reference Ionosphere 2016 (IRI2016), global ionosphere
Large gaps of this network in the Middle East, e.g., the Iran map (GIM) TEC and ANN model. The main focus is to
plateau, reduce the accuracy of the IGS GIMs in these areas. compare the results of the proposed method with those of
The time resolution of the IGS’s ionosphere maps is 2 h. It the GIM model.
should be noted that temporal variations in the ionosphere
can be less than 2 h. Therefore, it is not possible to investi-
gate the time series of high-resolution ionosphere variations
VTEC modeling using spherical harmonic
using the products of this network. To overcome this limita-
(SH) expansion
tion, we propose modeling using an adaptive neuro-fuzzy
inference system (ANFIS).
Global ionosphere maps TEC (GIMs-TEC) are provided
Recently, it has become clear that the techniques derived
by the IGS at a 2-h time interval (ftp://cddis.nasa.gov/gnss/
from artificial intelligence research and modern computer
products/ionex/). Spherical harmonic (SH) functions were
science provide a number of system aids to analyze and
used for modeling VTEC with reference to a solar geomag-
predict the behavior of complex solar-terrestrial dynamic
netic reference frame according to the following equation
systems. Methods of artificial intelligence have provided
(Schaer 1999):
tools that potentially make the task of ionospheric mod-
eling possible (Cander 1998). Soft computing includes ∑ ∑
nmax n
( )
components such as genetic algorithms (GA), ANNs, fuzzy VTEC(𝜑, 𝜆) = P̃ mn (sin 𝜑) anm cos (m𝜆) + bnm sin (m𝜆)
n=0 m=0
inference system (FIS), reinforcement learning (RL), and (1)
ant colony methods, often referred to as computational intel-
where φ is the geomagnetic latitude of the ionospheric pierce
ligence. Typically, soft computing has a special interest in
point (IPP), λin is the sun-fixed longitude of the IPP, n and
the combination of different methods, whereas each of the
m are the degree and order of the model, respectively, P̃ mn is
methods is usually used separately in computational intel-
the normalized associated Legendre function of degree n and
ligence. In other words, in soft computing when combining
order m, and anm and bnm are the unknown SH coefficients
neural networks with fuzzy algorithms or combining neural
and GIM parameters, respectively. The temporal resolution
networks with genetic algorithms, it is possible to achieve
of the TEC calculated with the GIM model is 2 h. However,
results that, when used separately, are impossible to achieve
research has shown that ionospheric temporal variations
or may have lower accuracy and computational speed (Zadeh
have frequencies of less than 2 h. These variations increase
1996). Therefore, the advantage of combining ANNs and
over the solar activity periods. Therefore, understanding the
fuzzy logic is used to model the ionosphere TEC.
temporal behavior of the ionosphere requires a higher tem-
The theory of fuzzy sets and fuzzy logic was first intro-
poral resolution model.
duced in Zadeh (1996). One of the best computations in
soft computing is combining ANNs with fuzzy logic. This
combination is called the ANFIS network. The ANFIS Fuzzy inference system (FIS)
has some advantages, including the ability to capture the
nonlinear structure of a process, adaptation capability, and Fuzzy inference systems take inputs and process them based
rapid learning capacity. ANFIS applies neural networks to on prespecified rules to produce the outputs. Both the inputs
13
GPS Solutions (2020) 24:51 Page 3 of 13 51
and outputs are real-valued, whereas the internal processing Adaptive neuro‑fuzzy inference system (ANFIS)
is based on fuzzy rules and fuzzy arithmetic (Zadeh 1996).
FIS is the key unit of a fuzzy logic system having decision ANFIS is an integration system in which neural networks
making as its primary task. It uses the “IF…THEN” rules are applied to optimize the fuzzy inference system. ANFIS
along with connectors “OR” or “AND” for drawing essential constructs a series of fuzzy if–then rules with appropriate
decision rules. A FIS is defined according to the following membership functions to produce the stipulated input–out-
five main sections: put pairs. The initial fuzzy rules and membership func-
tions are first set by using human expertise about the out-
• Rule base: It contains fuzzy if–then rules. puts to be modeled. Then, ANFIS can modify these fuzzy
• Database: It defines the membership functions of fuzzy if–then rules and membership functions to minimize the
sets used in fuzzy rules. error measure or explain the input–output relationship of
• Decision-making unit: It performs an operation on rules. a complex system. Figure 1 depicts the structure of an
• Fuzzification interface unit: It converts the crisp quanti- ANFIS network in five layers.
ties into fuzzy quantities. In Fig. 1, the first layer performs fuzzification, the sec-
• Defuzzification interface unit: It converts the fuzzy quan- ond layer performs the fuzzy T-norm operation for the
tities into crisp quantities. premise part of fuzzy rules, the third layer is used for nor-
malization, the fourth layer creates the fuzzy rules of the
Artificial neural networks (ANNs) consequent part, and finally, the fifth layer calculates the
final output of the system. The forward relations for the
An artificial neural network is based on a collection of con- network structure shown in Fig. 1 are as follows:
nected units or nodes called artificial neurons, which loosely ( ) ( )
model the neurons in a biological brain. Each connection, wjk = 𝜇M1 x1 .𝜇M2 x2
j k
j, k = 1, 2, … , m (3)
like the synapses in a biological brain, can transmit a signal
from one artificial neuron to another. An artificial neuron wjk
that receives a signal can process it and then send it to the wjk = j, k = 1, 2, … , m
∑ ∑
mm
(4)
next layer (Schalkoff 1997). A neuron or a nerve cell is actu- wi1i2
ally a function with n input and an output which the neuron i1=1 i2=1
13
51 Page 4 of 13 GPS Solutions (2020) 24:51
∑m ∑m � � � �
� � 7. If the Pg parameter is improved based on the fitting func-
k=1 fjk 𝜇Mj
x1 𝜇M2 x21 m m
j=1
y = ∑m ∑m � �
j
� � = fjk wjk (6) tion, the final one is selected. Otherwise, the search con-
j=1
𝜇 1 x1 𝜇M2 x2
k=1 M j
j=1 k=1
j tinues.
13
GPS Solutions (2020) 24:51 Page 5 of 13 51
cannot be considered to investigate the behavior of a phe- Using PCA, we can detect dominant frequencies in iono-
nomenon because it requires a lot of input observations. spheric behavior. The main purpose of using this method in
In order to overcome this problem, the idea of using the modeling the time series of the ionosphere is to identify the
principal component analysis to reduce the dimensions of first few components that are most effective in ionospheric
input space is presented. Reducing the dimensions of input behavior. Based on the PCA, the principal components that
space before using the soft computing system increases have the most information about the input observations are
the speed of convergence of algorithms and reduces the used to train the ANFIS network.
computational time.
The multivariable analysis of geophysical data is essen- Observations
tial for determining the inherent characteristics of the data
(Jackson 2003). The PCA method provides a multivariate In order to analyze the new theory, data from the GNSS
analysis based on the transformation of the space of cor- station of Tehran (35.69° E, 51.33° N), for 2016 and 2017
related variables into the space of uncorrelated variables have been used. All RINEX observation files have been
(Mallika et al. 2019). In order to model the time series of downloaded from the IGS Web site. The algorithm of the
the ionospheric variations, the VTEC values derived from previous part is used to calculate the TEC values and subse-
GPS stations are included in the VTEC(t, d)m×n matrix. In quent VTEC values at 24 h. The diagram of the new theory
this matrix, the parameter t indicates the observing time is shown in Fig. 2.
expressed in hours for day d, and d is the day of the year. The results were analyzed by computing root-mean-
As a result, it can be written as: square error and dVTEC:
√
⎡ VTEC1,1 … VTEC1,n ⎤ √
√1 ∑ N
( )2
VTEC(t, d) = ⎢ ⋮ ⋱ ⋮ ⎥ (11) RMSE = √ (17)
⎢ ⎥ VTECGPS − VTECM
⎣ VTECm,1 … VTECm,n ⎦ N i=1
13
51 Page 6 of 13 GPS Solutions (2020) 24:51
Table 1 Eigenvalue of VTEC matrix and its corresponding principal Table 2 Comparison of RMSE for training and testing stage and time
components of convergence (s) for PCA-ANFIS methods
PCA number Eigenvalue PCA (%) Method Training RMSE Testing RMSE Time of
(TECU) (TECU) conver-
1 1.082E + 06 97.968 gence
2 1.877E + 04 1.708 (s)
3 1.991E + 03 0.181
PCA1-ANFIS 1.856 2.454 148.32
PCA1,2-ANFIS 1.794 2.346 201.19
PCA1,2,3-ANFIS 1.756 2.297 246.95
13
GPS Solutions (2020) 24:51 Page 7 of 13 51
network have been assessed in the training stage. If the value time interval 1, 3,…, 23 UT and the prediction of the TEC
of correlation coefficient is closer to one, it indicates higher was performed for the time interval 2, 4,…, 24 UT. Since
correlation and accuracy of the model. Residual histogram GIM’s products are offered in 2-h time frames (2, 4,…, 24),
plot gives the accuracy level of the predicted data. It is a a comparison has been made for these times.
graph drawn between the value of residuals on horizontal According to the results of Fig. 5, the RMSE values for
axis and its frequency on vertical axis. The accumulation of the PCA-ANFIS, PCA-ANN, GIM, and IRI2016 models in
bins around zero with high frequency indicates a high level the testing stage are 2.454, 4.395, 4.775 and 6.895 TECU,
of accuracy of the estimated parameters. Figure 4 shows this respectively, with respect to the correlation coefficients
analysis. In both cases, the results of GPS-TEC are consid- as 0.881, 0.783, 0.762, and 0.734. The results of both the
ered as reference. Based on the explanations presented in RMSE and correlation coefficient indicate the superiority of
the previous sections, the training of both networks is based the PCA-ANFIS method. The IRI2016 model is less accu-
on the largest principal component in the 2-h period (1, 3, rate than the other three models in predicting the TEC value.
5,…, 23) UT. The GIM is a traditional model for calculating the TEC vari-
Based on Fig. 4, the RMSE values for the ANFIS and ations on a global scale. The temporal resolution of the TEC
ANN networks are 1.8562 and 2.1935 TECU, respectively. modeling with this model is 2 h. The comparison in Fig. 5
Also, the correlation coefficient between the ANFIS-TEC shows that this model is not very accurate in time series
and the GPS-TEC is 0.898, while for the ANN-TEC with modeling of TEC on a local scale. Another comparison was
the GPS-TEC, it is equal to 0.858. Examination of these made using the dVTEC (|VTECGPS − VTECmodel|) between
two indices (RMSE and correlation coefficient) for the train- the four models. Again, in this comparison, the TEC values
ing stage indicates a more accurate training in case of the obtained from the GPS are considered as reference. Figure 6
ANFIS method than the ANN method. Also, in the residual shows the results of this comparison.
histogram, the error curve of the ANFIS method is narrower From Fig. 6, it can be seen clearly that the results of the
than the ANN method. This means that the parameters esti- ANFIS method are much closer to the results of the GPS-
mated by the ANFIS model have smaller errors. A similar TEC observations. In other words, the numerical value of the
analysis has been made for the testing stage. The results of dVTEC in the ANFIS method is much less than those of the
this analysis are shown in Fig. 5. In this figure, the results other methods. The maximum dVTEC value is 4.31 TECU
of the analysis of the IRI2016 model and GIM-TEC are also for the IRI2016 model, which is calculated in December
shown. In all cases, the results of GPS-TEC are considered a 2016. In January 2017, the dVTEC value for the IRI2016
reference. At this stage, the modeling was carried out for the model is 4.14 TECU, which is the highest value compared to
13
51 Page 8 of 13 GPS Solutions (2020) 24:51
other months. In both years, the minimum value of dVTEC The dVTEC of the GIM is approximately twice that of the
is obtained for the ANFIS model. The dVTEC of the GIM ANFIS model. According to the results of Table 4, even the
is higher than the ANFIS and ANN models and lower than ANN is more accurate than the GIM model. Since there are
the IRI2016 model. In April 2016, the maximum dVTEC only two IGS stations in Iran, the calculation of the VTEC
in the GIM model is 3.58 TECU, while it is 3.05 TECU over Iran is performed using interpolation methods. The low
for September of 2017. Table 4 shows the average error of accuracy of the GIM in time series modeling of the VTEC
all models in 2016 and 2017. The average is calculated for can be due to the low number of IGS stations and the use of
12 months each year. interpolation methods. To compare the modeled VTEC more
13
GPS Solutions (2020) 24:51 Page 9 of 13 51
Fig. 6 Comparison of the
dVTEC for the ANFIS, ANN,
GIM and IRI2016 models for
2016 (top) and 2017 (bottom)
with consideration of GPS-TEC
as reference
Table 4 Comparison of average dVTEC for ANFIS, ANN, GIM and GPS-TEC data than the GIM model. On July 2, 2016, and
IRI2016 models in 2016 and 2017 March 11, 2017, the disagreement between GPS and the
dVTEC (TECU) GIM model can be as large as 6 TECU. To demonstrate the
performance of each ionospheric model, the monthly mean
ANFIS ANN GIM IRI2016
TEC calculated over a 1-month period (February and Octo-
2016 1.33 1.67 2.29 2.39 ber 2016, April and November 2017) using the GPS, ANFIS
2017 1.08 1.34 2.11 1.75 and GIM ionospheric models over 2 years (2016 and 2017)
is shown in Fig. 8.
It can be seen that the ANFIS model has a better agree-
clearly with the ANFIS, we plotted the profiles of daily vari- ment with the GPS data than GIM models. In February
ations for 4 arbitrary days. Figure 7 represents these profiles. 2016, ANFIS underestimates TEC during all hours by a few
Also, the VTEC profiles of the GPS and GIM are shown in TECU with respect to the GPS-TEC. For this month, the
these figures. GIM model overestimates the TEC with 0–3 TECU with
Figure 7 shows that the TEC derived from the ANFIS respect to the GPS-TEC. On October 2016, ANFIS overes-
model has a shape very similar to that of GPS-TEC. timates the TEC during 6–16 UT but underestimates TEC
The maximum TEC value occurs around 10:00 UT during 2–6 UT and 16–24 UT, by a few TECU with respect
(LT = UT + 3.30 h), and the minimum one is around 02:00 to the GPS-TEC. In April and November 2017, the disagree-
UT. The ANFIS model has a better agreement with the ment between GPS-TEC and GIM models can be as large as
13
51 Page 10 of 13 GPS Solutions (2020) 24:51
Fig. 8 Diurnal variation of
monthly mean TEC derived
from the two ionospheric
models in February and October
2016 and April and November
2017
3 TECU. These overestimations might be attributed to the September, and October, and in the interval 10–16 UT. In
large TEC data gaps in the Iranian region. There is a low both years and for the GIM model, the maximum difference
error value of 0–2 TECU for the ANFIS model compared is 4 to 6 TECU.
to the GPS-TEC in April and November 2017. In Fig. 9, the The absolute accuracies of ANFIS and GIM ionospheric
difference between the GPS-TEC and the TEC of the ANFIS models need further evaluation through precise point posi-
and GIM models is shown for all months of 2016 and 2017. tioning (PPP). In the above analyses, the TEC obtained from
This comparison is made for 2-h intervals (2,4,…,24). the GPS is used as a reference and the other two models
In 2016, the largest differences between the GPS-TEC are compared. To evaluate more precisely the ANFIS and
and TEC of the ANFIS model (dVTEC = VTECGPS − VTE- GIM ionospheric models, we used a single-frequency PPP
CANFIS) occurred in April and October, during 6–14 UT. approach. The main purpose of using single-frequency data
The maximum difference is 1–3 TECU. In 2017, most of is to correct the ionospheric delay using the ANFIS and GIM
the disagreement between GPS and ANFIS occurred in models. Since the dominant error in the single-frequency
April and at 6–14 UT. For the GIM model and in 2016, the PPP is the ionospheric delay, the accuracy of each model
maximum dVTEC occurred between April and May at 10-15 can be evaluated. To calculate the true coordinates of the
UT. In 2017, the largest differences occurred in the April, Tehran GNSS station, the observations of this station were
13
GPS Solutions (2020) 24:51 Page 11 of 13 51
Conclusion
13
51 Page 12 of 13 GPS Solutions (2020) 24:51
Fig. 10 RMSE of monthly
single-frequency PPP differ-
ences between GPS and other
two models
observational time and for various solar activities can be Komjathy A, Langley R. B (1996) An assessment of predicted and
considered. measured ionospheric total electron content using a regional
GPS network. In: Proceedings of ION NTM 1996, Institute of
Navigation, Santa Monica, CA, January 22–24, pp 615–624
Acknowledgements The International GNSS Service is acknowledged Leandro RF, Santos MC (2007) A neural network approach for
for providing the data used in this study. The anonymous reviewers are regional vertical total electron content modeling. Stud Geophys
thanked for their constructive comments. Geod 51(2):279–292
Liu Z (2004) Ionospheric tomographic modeling, UCGE Reports,
Number 20198, University of Calgary
Liu Z, Gao Y (2003) Ionospheric TEC predictions over a local area
References GPS reference network. GPS Solut 8(1):23–29
Ma XF, Maruyama T, Ma G, Takeda T (2005) Three dimensional
Cander R (1998) Artificial neural network applications in ionospheric ionospheric tomography using observation data of GPS ground
studies. Ann Geofis 41(5–6):757–766 receivers and ionosonde by neural network. J Geophys Res
Ciraolo L, Azpilicueta F, Brunini C, Meza A, Radicella SM (2007) 110(A05308):1–12
Calibration errors on experimental Slant Total Electron Content Mallika I, Ratnam D, Sivavaraprasad G, Raman S (2019) Implemen-
(TEC) determined with GPS. J Geodesy 81(2):111–120 tation of hybrid ionospheric TEC forecasting algorithm using
Ghaffari Razin MR (2015) Development and analysis of 3D ionosphere PCA-NN method. IEEE J Sel Top Appl Earth Observ Rem Sens
modeling using base functions and GPS data over Iran. Acta Geod 12(1):371–381
Geophys 51(1):95–111 Mannucci A, Wilson B, Yuan D, Ho C, Lindqwister U (1998)
Ghaffari Razin MR, Voosoghi B (2017) Ionosphere tomography using Runge T (1998) Global mapping technique for GPS-derived
wavelet neural network and particle swarm optimization training ionospheric total electron content measurements. Radio Sci
algorithm in Iranian case study. GPS Solut 21(3):1301–1314 33:565–582
Ghaffari Razin MR, Voosoghi B, Mohammadzadeh A (2015) Effi- Saito A, Teraishi S, Ueno G, Fujita N, Tsugawa T (2007) GPS iono-
ciency of artificial neural networks in map of total electron content spheric tomography over Japan with constrained least‐squares
over Iran. Acta Geod Geophys 51(3):541–555 method. Eos Trans AGU 88(52), Fall Meet. Suppl., Abstract
Habarulema JB, McKinnell L-A, Opperman BDL (2009) A recurrent SA13A-1061
neural network approach to quantitatively studying solar wind Schaer S (1999) Mapping and predicting the earths ionosphere using
effects on TEC derived from GPS; preliminary results. Ann Geo- the global positioning system. Ph.D. dissertation, Astronomical
phys 27(11):2111–2125 Institute, University of Berne, Switzerland, 205
Hernández-Pajares M, Juan J, Sanz J (1999) New approaches in global Schalkoff RJ (1997) Artificial neural networks, vol 1. McGraw-Hill,
ionospheric determination using ground GPS data. J Atmos Sol New York
Terr Phys 61:1237–1247 Sekido M, Kondo T, Kawai E, Imae M (2003) Evaluation of GPS based
Hernández-Pajares M, Juan J, Sanz J, Orus R, Garcia-Rigo A, Feltens ionospheric TEC map by comparing with VLBI data. Radio Sci
J, Komjathy A, Schaer S, Krankowski A (2009) The IGS VTEC 38(4):1069. https://doi.org/10.1029/2000RS002620
maps: a reliable source of ionospheric information since 1998. J Strangeways HJ et al (2009) Near-earth space plasma modeling and
Geodesy 83:263–275 forecasting. Ann Geophys 52(3–4):255–271
Hirooka S, Hattori K, Takeda T (2011) Numerical validations of neu- Walker JK (1989) Spherical cap harmonic modeling of high latitude
ral-network-based ionospheric tomography for disturbed iono- magnetic activity and equivalent sources with sparse observations.
spheric conditions and sparse data. Radio Sci 46(05):1–13 J Atmos Terr Phys 51(2):67–80
Jackson JE (2003) A users’ guide to principal components. Wiley, Wen D, Wang Y, Norman R (2012) A new two-step algorithm for iono-
London spheric tomography solution. GPS Solut 16(1):89–94
13
GPS Solutions (2020) 24:51 Page 13 of 13 51
Yao Y, Chen P, Zhang S, Chen J (2013) A new ionospheric tomography Behzad Voosoghi received his
model combining pixel-based and function-based models. Adv Ph.D. in geodesy and geomatics
Space Res 52(4):614–621 engineering, from the Depart-
Yao Y, Tang J, Kong J (2015) New ionosphere tomography algo- ment of Geodesy and Geomatics
rithm with two-grid virtual observations constraints and three- Engineering, University of Stutt-
dimensional velocity profile. IEEE Trans Geosci Remote Sens gart, Stuttgart, German (1995–
53(5):2373–2383 2000). Now he is an associate
Yin P, Mitchell CN, Spencer PSJ, Foster JC (2004) Ionospheric elec- professor in the Faculty of Geod-
tron concentration imaging using GPS over the USA during the esy and Geomatics Engineering,
storm of July 2000. Geophys Res Lett 31:L12806 K. N. Toosi University of Tech-
Yizengaw E, Moldwina MB, Dysonb PL, Essexb EA (2006) Using nology. His research focuses on
tomography of GPS TEC to routinely determine ionospheric atmospheric modeling, earth
average electron density profile. J Atmos Solar Terr Phys surface deformation analysis,
69(3):314–321 and tensor calculations.
Zadeh LA (1996) Fuzzy sets. Inf Control 8:338–353
Zhang JR, Zhang J, Lok T, Lyu M (2007) A hybrid particle swarm
optimization–back-propagation algorithm for feed forward neural
network training. Appl Math Comput 185(2007):1026–1037
Zheng D, Hu W, Nie W (2015) Multiscale ionospheric tomography.
GPS Solut 19:579–588
13