Feed forward neural network based ionospheric
model for the East African region
A. Tebabal1, S.M. Radicella2, B. Damtie1, Y. Migoya-Orue’2, M.Nigussie1, B. Nava2
Beacon Satellite Symposium 2019, University of Mazury
19–23 August 2019, Olsztyn, Poland
1Bahir Dar University, Bahir Dar, Ethiopia, 2Abdus Salam International Center for Theoretical Physics, Trieste, Italy.
Outline 2
1 IntroductionThe Earth’s ionosphere
2 Data Source and Model ApproachesNeural Networks (NN)
3 ResultsThe Working Principle of Our Model
4 Summary
Introduction Data Source and Model Approaches Results Summary
The Earth’s Ionosphere 3
I The ionosphere is the ionized part of the Earth’s upper atmosphere.
The composition of theatmosphere changes withheight, the ion production ratealso changes.
This leads to the formation ofthree main ionized peaks: D, E,and F regions.
Earth’s ionosphere is a highlyvariable in space and time.
source:https://scied.ucar.edu/ionosphere.
Introduction Data Source and Model Approaches Results Summary
Cont. . . 4
Variability in the solar irradiance leads tovariability in the Earth’s atmosphere.
1 UV variationsmodify
ozone and the middleatmosphere structure.
2 Heating of the upper atmosphere (EUV;30-120 nm) satellite drag
3 Formation of the ionosphere (XUV-EUV,1-120 nm) satellite communications
IEUV is the primary driver of ionospheric variability butgeomagnetic activity and lower atmosphere meteorology alsocontribute.
Introduction Data Source and Model Approaches Results Summary
Cont. . . 5
Understanding the variations of the ionosphereis crucial to mitigate ionospheric effects and thedrivers behind these variations.
Ionospheric measurements are limited by theirincomplete spatial and temporal coverage.Therefore, ionospheric models are employed.
Developing a data driven model, to investigatethe ionospheric variability, is the goal of thisstudy.
Introduction Data Source and Model Approaches Results Summary
Data Source 6
32 34 36 38 40 42 44 46 48Geographic longitude (degrees)
2
4
6
8
10
12
14
16G
eogr
aphi
c la
titud
e (d
egre
es)
Bdmt
Armi
Aboo
Asos
Debk
Ginr
Nege
Adis
Nazr
Serb
Robe
Shis
Metu
√Hourly Total Electron Content (TEC) data from GPS stations
marked with blue circles used for training and red stars for testing.
Introduction Data Source and Model Approaches Results Summary
Model Approaches 7
Neural Networks
are models that attempt to mimic some of thebasic information processing methods found inthe brain [Samarasinghem, 2007].
Benefits of Neural Networks
its computing power,its ability to learn and generalization,
generalization means producing reasonable outputs for inputsnot encountered during the training(learning).
have useful properties such as nonlinearity,adaptivity and fault tolerance.
Introduction Data Source and Model Approaches Results Summary
Model Approaches 8
I We employed a feed-forward neural network (i.e. fully connectedlayers)
Σ
z1
y1
Σ
zk
yk... ...
...
x1
Input layer Output layer
Σ
Hidden layer
h1
a1 w(2)11
x2 Σ
h2
xi Σ
hj
ajw
(2)kj
bias w(1)j0
bias
w(2)k0
w(1)11
w(1)ji
A schematic diagram of feed-forward neural network composed ofthree layers.
Introduction Data Source and Model Approaches Results Summary
Model Approaches 9
The basic algorithm for NN
yk = ϕ
∑
j
w(2)kj ϕ
(∑
i
w(1)ji xi + w
(1)jo
)
j
+ w(2)ko
k
(1)
Phases of application of NN modeling technique
Training data
Model parameters estimation
Testing data
Model validation
Introduction Data Source and Model Approaches Results Summary
The Working Principle of Our Model (Tebabal et al, 2019) 10
Model Input Parameters
.
∑
∑
∑
∑
∑
∑
∑
∑
∑∑
∑
∑
∑
∑
∑
∑
∑
∑
∑
∑
∑TEC
BiasBias
Bias
Inputlayer
Hiddenlayer
Hiddenlayer
Outputlayer
DOY
HR
F10.7
ap
Gp
Introduction Data Source and Model Approaches Results Summary
Model validation and comparison 11
One hour ahead prediction
TEC data from 2012 to 2014 used to develop the NNmodel parameters (weights and biases).Observations in the year 2015 used for validation/testingof the model.
model prediction RMS error ranges from 5.44 – 6.45 TECU.correlation coefficients between model prediction andGPS–TEC ranges from 0.925 to 0.96.
GPS-TEC observations from 2012 to 2015 used forfurther validation
model was able to explain more than 93% of the variability ofGPS–TEC.RMS error ranges from 3.9 – 6 TECU.
Introduction Data Source and Model Approaches Results Summary
Model validation and comparison 12
One day ahead prediction
The response of ionosphere for solar and geomagnetic activityranges 1–2 days (e.g., Kutiev et al. 2012).
Model inputs are F10.7 and ap index values of the previous day
Table: One-day ahead forecasting over for different GPS stations inthe year 2015.
GPS station R RMS (TECU) STD (TECU)1 Adis 0.945 6.166 6.1092 Aboo 0.942 5.989 5.9803 Armi 0.955 5.270 5.2654 Asos 0.955 5.653 5.4765 Ginr 0.947 5.670 5.6696 Nege 0.949 5.557 5.5557 Debk 0.938 6.504 6.4068 Bdmt 0.948 5.879 5.8549 Shis 0.962 5.485 5.345
Introduction Data Source and Model Approaches Results Summary
Diurnal Variation analysis 13
0
40
80 DOY 75
a) DOY 68 DOY 76 DOY 66 DOY 67
Q1
Q2
Q3
Q4
Q5
Robe 2014
0
20
40
60
TE
C/T
EC
U DOY 152
b)
DOY 163 DOY 178 DOY 166 DOY 173
GPS TEC NN TEC NeQuick 2 TEC
0
40
80 DOY 257c)
DOY 258 DOY 251 DOY 260 DOY 263
0 8 16 0 8 16 0 8 16 0 8 16 0 8 16Time,hrs (UT)
0
40
80 DOY 345d)
DOY 352 DOY 351 DOY 350 DOY 361
Quiet days
a) March
b) June
c) September
d) December
Introduction Data Source and Model Approaches Results Summary
Quiet days RMS error 14
March June September December
0
5
10
15R
MS
/TE
CUNN TEC NeQuick 2 TEC
Introduction Data Source and Model Approaches Results Summary
Disturbed conditions 15
IPerformance of the NN model to disturbed conditions
Four intense geomagnetic storms (with peak Dst< -110nT) during 2012-2015 selected.
These storm events occurred on:
315 July 2012 with Dst index value of -139 nT,
319 February 2014 with Dst index value of -124nT,
301 June 2013 with Dst index value of -119nT, and
317 march 2015 with Dst index value of -223nT.
Introduction Data Source and Model Approaches Results Summary
Disturbed conditions 16
Metu: July 2012
-150
-100
-50
0
50
Dst
(nT
)
DOY 195 DOY 196 DOY 197 DOY 198 DOY 199
0
70
140
ap in
dex
Jul-13 Jul-14 Jul-15 Jul-16 Jul-17Date
0
40
80
TE
C/T
EC
U
GPS TEC NN TEC
Serb: June 2013
-140
-100
-60
-20
20
Dst
(nT
)
DOY 150 DOY 151 DOY 152 DOY 153 DOY 154
0
40
80
120
ap in
dex
May-30 May-31 Jun-01 Jun-02 Jun-03Date
0
20
40
60
TE
C/T
EC
U
GPS TEC NN TEC
Robe: February 2014
-140
-80
-20
40
Dst
(nT
)
DOY 48 DOY 49 DOY 50 DOY 51 DOY 52
0
50
100
ap in
dex
Feb-17 Feb-18 Feb-19 Feb-20 Feb-21Date
0
40
80
TE
C/T
EC
U
GPS TEC NN TEC
Nazr: March 2015
-250
-150
-50
50
150
Dst
(nT
)
DOY 74 DOY 75 DOY 76 DOY 77 DOY 78
0
100
200
ap in
dex
Mar-15 Mar-16 Mar-17 Mar-18 Mar-19Date
0
40
80
TE
C/T
EC
U
GPS TEC NN TEC
Dst and ap index with predicted and GPS TEC for five days period, centered at the time when Dst reaches
minimum.
Introduction Data Source and Model Approaches Results Summary
Disturbed conditions 17
RMS Error between GPS-TEC and predicted values
Station Year Month DOYNNRMS error
Metu 2012 July
195 4.3198196 5.4912197 3.8134198 8.1569199 4.4319
Serb 2013 June
150 3.1788151 3.3501152 5.1312153 2.4064154 3.5089
Robe 2014 February
48 7.200249 3.641150 9.111151 7.477452 –
Narz 2015 March
74 2.886275 1.787676 12.095377 6.088678 5.4806
Introduction Data Source and Model Approaches Results Summary
Summary 18
We present an NN model for the East African regionionosphere.
The model inputs are geographic locations, DOY, HR, F10.7,and ap index and produce time dependent TEC.
Our NN models result indicate
the overall RMS error between GPS TEC and the modelsprediction lies in the range of 3 to 6.05 TECU.one-hour and one-day ahead prediction are in good agreementwith the observed GPS-TEC values.
Low latitude ionosphere is highly variable at different timescales.
Knowledge of the ionospheric response to other disturbancesources and corresponding observations required to increasethe performance of the model.
Introduction Data Source and Model Approaches Results Summary
Acknowledgments 19
I gratefully acknowledgeOrganizers of 20th International Beacon Satellite Symposium
Introduction Data Source and Model Approaches Results Summary
Thank you for your attention!!
1Bahir Dar University, Bahir Dar, Ethiopia, 2Abdus Salam International Center for Theoretical Physics, Trieste, Italy.