Optimizing Local Geoid Undulation Model Using GPS Levelling Measurements and Heuristic Regression Approaches

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Survey Review

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/ysre20

Optimizing Local Geoid Undulation Model using


GPS/Levelling Measurements and Heuristic
Regression Approaches

Mosbeh R. Kaloop , Ahmed Zaki , Hamad Al-Ajami & Mostafa Rabah

To cite this article: Mosbeh R. Kaloop , Ahmed Zaki , Hamad Al-Ajami & Mostafa Rabah (2020)
Optimizing Local Geoid Undulation Model using GPS/Levelling Measurements and Heuristic
Regression Approaches, Survey Review, 52:375, 544-554, DOI: 10.1080/00396265.2019.1665615

To link to this article: https://doi.org/10.1080/00396265.2019.1665615

Published online: 16 Sep 2019.

Submit your article to this journal

Article views: 283

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=ysre20
Optimizing Local Geoid Undulation Model
using GPS/Levelling Measurements and
Heuristic Regression Approaches
∗ 1,2
Mosbeh R. Kaloop , Ahmed Zaki3, Hamad Al-Ajami4,5 and Mostafa Rabah4
This study investigates to use GPS/Levelling measurements of Kuwait and four heuristic regression
methods including Least Square Support Vector Regression (LSSVR), Gaussian Process
Regression (GPR), Kernel Ridge Regression (KRR), and Multivariate Adaptive Regression
Splines (MARS) for modelling local geoid undulation. The accuracy of the models was
compared by geoid undulation of gravitational observations and Global Geopotential Models
(GGMs). The results show that the KRR model is suitable for Kuwait geoid model, its error of
percentage is 0.018 and 0.124% relative to gravity and GPS/Levelling geoid undulation models,
respectively. Furthermore, the comparison of KRR model with GGMs models signifies its accuracy.
Keywords: Geoid, GPS/Levelling, Modelling, Regression

1. Introduction these are broadly divided into three types: Gravimetric,


Astrogeodetic, and Geometric (Becker 2012). The Gravi-
The determination of the orthometric height (H) from metric solution is based on the use of gravity data (Wang
geometric levelling has practical difficulties that, despite et al. 2012, Huang and Véronneau 2013). The Astrogeo-
several scientific and technological advances, passed a detic geoid can be determined by the astronomic obser-
century without substantial modifications or advances vation of latitude, longitude and the deflection of the
(Becker 2012). Currently, the Global Navigation Satellite vertical (Adescu et al. 2016, Kumar and Nath 2016).
System (GNSS) has been used with reasonable success for The Geometric method depended on the combination
ellipsoidal height (h) determination. The ellipsoidal between the GNSS heights and orthometric heights to
height does not have a physical meaning (not linked to determine the geoid undulation then, interpolation the
gravity) so that, the orthometric heights are used in prac- value of geoid undulation at any unknown point. Where-
tice. The geoid undulation (N = H-h) is needed to convert fore with a sufficient number of benchmarks with known
the ellipsoidal heights to orthometric heights. The Geoid ellipsoidal heights and orthometric heights, it is often
is defined as the equipotential surface of the Earth’s grav- possible for using a mathematical expression, or model
ity field which coincides with the Mean Sea Level (MSL) approaches, that allow interpolating of geoid heights
in the absence of disturbing factors like ocean currents, (Yang and Chen 1999, Kao et al. 2014). Herein, data
salinities, wind, etc., and it extends through the continents availability, the accuracy of estimations and simplicity
(Uotila 1971). The geoid is continuous and much of modelling are the main factors to develop the model
smoother than the actual earth surface, unlike the ellip- approaches for estimating the geoid undulation. This
soid, it is still a closed, too complicated to serve as the study was implemented for modelling a soft computing
computational surface on which to solve geometrical pro- gentile local geoid model for Kuwait using GPS/levelling
blems, but it is suitable as a vertical datum (Becker 2012). measurements.
The main important application of the geoid is the deter- Artificial Neural Networks (ANN) have been success-
mination of orthometric heights. The geoid undulation is fully used in modelling the geometric geoid undulation
needed to convert the ellipsoidal heights to orthometric and datum transformation for different regions; this is
heights. There are different methods to determine the because these techniques can be utilized in solving com-
geoid based on methods of development and data used, plicated problems, especially with limitation of GPS/
Levelling data. Seager et al. (1999) utilized the BackPro-
pagation ANN (BPNN) for modelling a local geoid for
1
Department of Civil and Environmental Engineering, Incheon National Uni- area 2°×2°, and they found that it is a suitable tool for
versity, Incheon, Korea
2
Public Works and Civil Engineering Department, Mansoura University, geoid undulation modelling. Kavzoglu and Saka (2005)
Mansoura, Egypt
3
designed an ANN model for modelling the geoid undula-
Civil Engineering Department, The Higher Institute of Engineering in El-
Shorouk City, Cairo, Egypt
tion for Istanbul in Turkey using 190 GPS/Levelling data
4
Civil Engineering Department, Benha University, Benha, Egypt points, and their results showed that the model designed
5
Training Authority, Public Authority for Applied Education and Training, had low distortion from the GPS/levelling data surface.
Adailiyah, Kuwait
Elshambaky et al. (2018) established a three direction

Corresponding author, email: Mosbeh.kaloop@gmail.com

© 2019 Survey Review Ltd


Received 14 July 2019; accepted 3 September 2019
544 DOI 10.1080/00396265.2019.1665615 Survey Review 2020 VOL 52 NO 375
Kaloop et al. Optimizing Local Geoid Undulation Model using GPS/Levelling Measurements

transformation ANN model to transform global to local related reviews and literature that the use of MARS and
coordinates of Egypt, and they summarized that the accu- KRR have been still limited and the use of GPR has
racy of the developed ANN model is an approach to not been studied yet in modelling geoid undulation;
20.16 cm. Further applications for the ANN in the besides, a comparison between these models and
geoid modelling and datum transformation can be LSSVM has not been conducted.
found in (Tierra et al. 2008, Pikridas et al. 2011, Veronez The purposes of this study are (i) to investigate the abil-
et al. 2011, Ziggah et al. 2016, Elshambaky 2018). ity of MARS, GPR and KRR for modelling geoid undu-
Furthermore, the Adaptive Neuro-Fuzzy Inference lation using GPS/Levelling observations, (ii) to compare
System (ANFIS) and regression models were applied for the results with the LSSVM method, and (iii) to estimate
estimating the geoid undulation (Erol and Celik 2004, a novel technique for modelling local geoid model and
Zaletnyik et al. 2008, Akyilmaz et al. 2009, Wang et al. compare the model with local gravity geoid model and
2009, Yilmaz 2010, Erol and Erol 2013, Rabah and GGMs models that assessed in Kuwait. Consequently,
Kaloop 2013, Kao et al. 2014). ANFIS, ANN, Wavelet this paper is structured to achieve these purposes by the
Neural Network (WNN), and Multivariable Polynomial following sections; the input and output datasets of
Regression Equations (MPRE) were used to predict GPS/levelling and the Global Geopotential Models
geoid of Istanbul area using GPS data, and the results (GGMs) of Kuwait are described, and the modelling
showed that ANFIS and WNN models provided a better approaches using the MARS, GPR, KRR, and LSSVM
performance (Erol and Erol 2013). Polynomial, ANFIS, are given in section 2. The comparative results of different
and ANN were applied to model the local geoid of the methods for predictions of Kuwait geoid undulation are
Izmir region, and the authors concluded that the shown in section 3 that the Root Mean Square Error
ANFIS outperformed than other models (Akyilmaz (RMSE), Mean Absolute Error (MAE), Mean Bias
et al. 2009). Erol and Celik (2004) utilized the interp- Error (MBE), agreement index (d) and Nash Sutcliffe
olation methods, Kriging and Inverse Distance Weight Efficiency (NSE) are used to evaluate the accuracy of
(IDW) to estimate a local geoid undulation model, and the four heuristic regression techniques; in addition, the
the results showed that the Kriging was high accurate Taylor diagram and box- whisker plot are used to com-
than IDW to estimate the geoid height. Least Square Sup- pare a regression model with GGMs models. Finally,
port Vector Machine (LSSVM) was compared by ANN the conclusion is presented that the KRR can be provided
and polynomial techniques for estimating a geoid height, accurate predictions compared to other methods and the
and the results concluded that the LSSVM is better than LSSVM is slightly accurate than the GPR. The soft com-
ANN to predict geoid height (Wang et al. 2009). Second- puting can be used to estimate an accurate geoid model
order curve surface method, BPNN and polynomial, lin- for Kuwait comparing by GGMs models.
ear and kernel functions of LSSVM were compared and
assessed into estimating local geoid in Taiwan, and the
results showed that the kernel, Radial Basis Function 2. Data and Methods
(RBF), of LSSVM is better than other methods (Kao
et al. 2014). A shoreline geoid model is predicted using 2.1. Data description
ANN, ANFIS, and LSSVM, and the results noticed Kuwait locates between latitudes 28.5° N and 30.1° N and
that the LSSVM can be used to estimate a missing data longitudes 46.5° E and 48.5° E, as presented in Figure 1.
of the local geoid model (Kaloop et al. 2018). The total area of Kuwait is 17818 Km2. The GPS/level-
Meanwhile, LSSVM algorithm, which is reformulation ling data was collected by Vision International Co.,
of Support Vector Regression (SVR), provide better per- Kuwait in 2016/2017 and made available for this study.
formance than ANN and ANFIS in the geoid estimation; The GPS coordinates of the benchmarks were computed
furthermore, Yuan and Lee (2015) summarized that the from a dual-frequency GPS receiver by Static and
SVR performance is better than ANN under the same Rapid-Static measurement method and refer to
training conditions, and the LSSVM has a shorter com- ITRF2008 datum. The approximate accuracies of GPS
puting time than SVR. Thus, SVM is used for modelling coordinates are ±1.0 and ±1.5 cm in the horizontal and
local geoid in Hungary (Zaletnyik et al. 2008), and vertical directions, respectively (El-Ashquer et al. 2019).
Elshambaky (2019) utilized LSSVM to model Egypt In this manner, the orthometric heights of the benchmarks
geoid model. Herein, it should be noticed that other were determined by high precision spirit levelling from the
regression algorithms, recently, have been developed and Ministry of Defense (MoD) benchmarks network. The
are not used in the geoid modelling. For instance, Multi- MoD networks were referenced to Kuwait Public Work
variate Adaptive Regression Splines (MARS) has been Department vertical datum (Kuwait PWD). Kuwait
used widely in the prediction applications (Zakeri et al. PWD was defined by Kuwait Oil Company as the Mean
2010, Zhang and Goh 2016, Keshtegar et al. 2018). Low Low Water (approximately 1.03 m below the mean
MARS was utilized for two directions coordinate trans- sea level) at Kuwait City (Rabah and Hattab 2009). The
formation and estimating orthometric height in Ghana, absolute accuracy of the orthometric heights is approxi-
and the results showed that the MARS accuracy is higher mately ±1.0 cm. The GPS/Levelling-based geoid undula-
than ANN results (Peprah and Kumi 2017, Ziggah and tion ranges between −18.54 and −10.95 m with an
Laari 2018). Nowadays, Gaussian Process Regression average value of −15.45 m and a standard deviation of
(GPR) and Kernel Ridge Regression (KRR) have been ±1.51 m (El-Ashquer et al. 2019). The available GPS/
developed into using to increase the accuracy of the pre- levelling data are presented in Fig. 1a; 78 stations are
diction models (Volgyesi et al. 2005, Kopsiaftis et al. divided into 63 and 15 as a training and testing datasets.
2019). Volgyesi et al. (2005) presented an application for The geoid undulation of Kuwait was calculated using
calculating the orthometric height by KRR; the standard the whole GPS/levelling data and presented in Fig. 1b.
deviation of model error was ±9 mm. It is evident from Approximately smooth geoid undulation with slope

Survey Review 2020 VOL 52 NO 375 545


Kaloop et al. Optimizing Local Geoid Undulation Model using GPS/Levelling Measurements

1 Kuwait (a) GPS/levelling points’ distributions and (b) contour map of geoid undulation

from west to east shows in Fig. 1b. Herein, an integration over the world EGM2008 is used. In the computing of
between land and ocean observation is recommended to geoid model by the GGMs models, the first-degree term
improve the local geoid model of Kuwait, especially on is set to zero by setting the centre of the reference ellipsoid
the shore region. to the centre of the Earth (Yakubu et al. 2017). The bias
In this study, the geodetic latitudes (w) and longitudes between the GNSS/Levelling stations and GGMs may
(l) of stations are used to estimate the geoid undulation be occurred due to the definition of the vertical datum
(N). The basic statistical properties of the datasets are of Kuwait which defined by Kuwait Oil Company
provided in Table 1 that M is the mean, STD is the stan- depends on the Mean Low Low Water (approximately
dard deviation of variables, CV is coefficient of variances 1.03 m below the mean sea level) at Kuwait City (El-Ash-
as STD/M, SK is skewness of data, Min and Max are quer et al. 2019). Figure 2 illustrates the variation of geoid
respectively the minimum and maximum of data points. undulation that calculated by GPS/Levelling data and
From the table, it can be seen that the skewness of data gravity data; also, the variation of EIGEN-6C4 GGM
is small, and the geoid undulation has the highest skewed model is presented. Here, the whole data available are
distribution followed by the longitude and latitude in both used to estimate the variation of GPS/Levelling geoid
data sets; this indicates that the degree of distortion of model. Table 2 illustrates the variation of the gravity
geoid undulation from the normal distribution is higher model and GGM’s models relative to GPS/levelling
than the longitude and latitude in Kuwait. The CV is geoid model at the whole points.
almost the same for training and testing data sets. Longi- The gravimetric geoid model of Kuwait has been com-
tudes have the highest correlation with geoid undulation puted using the SPW_R5 up to SH d/o 250 with
in both data sets. EGM2008 up to SH d/o 2190 and ground gravity stations
(93 points) (El-Ashquer et al. 2019). The strategy followed
for modelling of the gravimetric geoid is based on the
2.2. GGMs and gravimetric geoid of Kuwait RCR with RTM reduction and the one dimensional
El-Ashquer et al. (2019) have been studied the GGMs of Fast Fourier Transform (1D-FFT) technique with
Kuwait. According to their study, the best combined Wong-Gore modification. From Fig. 2 and Table 2, it
GGMs model can be used in Kuwait are EIGEN-6C4 can be seen that the geoid models that estimated by
and XGM2016 models at their maximum degree and GPS/Levelling and gravity data are very close. Therefore,
there found that the most suitable degree-and-order (d/ the geoid model of GPS/levelling of Kuwait is accurate
o) for GOCE GGMs in Kuwait is 250 from SPace Wise and can be used to model the geoid undulation. The vari-
model (SPW_R5). In this study, these models are used ation of EIGEN-6C4 model is seen high comparing by
to evaluate the best fitting model that will be estimated the gravity model, as presented in Fig. 2. Furthermore,
using soft computing technique; also, a common model the accuracy of EIGEN-6C4 and XGM2016 models are

Table 1 Statistical parameters of using data sets

Dataset Parameter Unit M STD CV SK Min Max Correlation with N

Training w o
29.46 0.37 0.01 −0.41 28.59 30.07 −0.02
l o
47.73 0.37 0.01 −0.64 46.61 48.40 −0.97
N m −15.52 1.35 −0.09 1.04 −17.96 −10.96 1.00
Testing w o
29.35 0.38 0.01 0.18 28.85 30.01 0.31
l o
47.79 0.40 0.01 −0.70 46.96 48.28 −0.99
N m −15.71 1.53 −0.10 0.88 −17.56 −12.60 1.00

546 Survey Review 2020 VOL 52 NO 375


Kaloop et al. Optimizing Local Geoid Undulation Model using GPS/Levelling Measurements

2 GPS/Levelling Geoid model variation of Kuwait relative to (a) gravity model, (b) EIGEN-6C4 model.

also seen higher than EGM2008 when they are compared be optimized based on the optimality condition as fol-
by GPS/levelling geoid model. However, we can be lows:
assumed that the three GGM’s models can be used to esti-
mate the geoid model in Kuwait, whereas the variation 1 g n
min (w, b, e) = wT w + ei (2)
between the three models is small and can be neglected. 2 2 i=1
In the current study, the three GGMs models and gravity
geoid model are used to evaluate the best soft computing where g is regulation parameter, and n is the number of
model that will be estimated by GPS/Levelling data. observations; herein, the lagrange multipliers is applied
to estimate these parameters (Kaloop et al. 2018, Elsham-
baky 2019). By eliminating both (w, e) and solving for
2.3. Heuristic methods for modelling geoid (b, a), where a is the lagrange multiplier, the geoid undu-
undulation lation can be represented as follows:
2.3.1. Least square support vector regression (LSSVR) 
n

Least square support vector machine/regression (LSSVR) N= ai k(x, xi ) + b (3)


i=1
and support vector machine/regression (SVR) are almost
the same with slight differences. Elshambaky (2019) and, the RBF kernel function can be presented as follows:
demonstrated the differences between SVR and LSSVR, −x − x 
i
and he found the performances of LSSVR is better in k(x, xi ) = exp (4)
2 s2
the datum transformation. Furthermore, LSSVR is used
widely to optimize geoid undulation in different areas where, s is the kernel function bandwidth.
(Zaletnyik et al. 2008, Wang et al. 2009, Kao et al. The structure diagram for predicting Kuwait geoid
2014, Kaloop et al. 2018, Elshambaky 2019). Here, the undulation based on LSSVR model is represented in
LSSVR is used to estimate Kuwait geoid undulation Fig. 3a. Thus, from above, it can be seen that the main
and evaluate the accuracy of other models, so a summary parameters that affected the accuracy of LSSVR are g,
of this method is presented. In LSSVR, The geoid undu- a, b, and s. In the current study, Matlab code has been
lation can be calculated as follows: designed to estimate these parameters and geoid undula-
tion. The cross-validation method is applied to estimate
Ni = wT c(xi ) + b + ei (1) the LSSVR parameters, and the obtained parameters g,
where, c(xi ) is a mapping function to transfer the input b and s are 14.65, −0.83 and 0.08, respectively. Figure
variables xi = [wi , li ] from its primal space to feature 3b shows the a values of training points.
space. w is input weight vector; b represents the biase of
the linear model, and e is the model errors. Herein, a com- 2.3.2. Multivariate Adaptive Regression Splines (MARS)
mon mapping function utilized is the kernel radial basis MARS is a nonlinear and non-parametric regression
function (RBF) (k(x, xi )), and the other parameters can method; it can simulate nonlinear relationship between
input and output variables of models using piecewise lin-
ear splines. The basis functions (BFs) are used in MARS
Table 2 Evaluation of global and gravity geoid models which are defined in pairs forms based on a knot to define
relative to GPS/Levelling geoid model of Kuwait
(units in m)
an inflection region. The geoid undulation can be esti-
mated by MARS as follows:
Model Min Max M STD

m

EIGEN-6C4 −1.23 −0.72 −0.90 0.09


N = b0 + bi BFi (5)
i=1
XGM2016 −1.27 −0.70 −0.90 0.09
EGM08 −1.06 −0.63 −0.79 0.10 where, bi are unknown parameters which are calculated
Gravity model −0.05 0.03 0.00 0.01
using the least square method, and m is the number of

Survey Review 2020 VOL 52 NO 375 547


Kaloop et al. Optimizing Local Geoid Undulation Model using GPS/Levelling Measurements

3 (a) LSSVR model diagram, (b) The a parameter of LSSVR


model

terms in the final model that is found in a forward back-


ward stepwise process. BFi is the i th basis function which
is defined based on knot C from piecewise linear or non-
linear basis function from the following set function,
BFi = {|x − Ci |+ , |Ci − x|+ }, where, |x − Ci |+ =
max (0, x − Ci ) and |Ci − x|+ = max(0, Ci − x). Finally, 4 (a) MARS geoid model diagram, (b) MARS model of Kuwait
the cross validation is used to estimate the best model geoid undulation
can be used to estimate the geoid undulation; the lowest
generalized cross validation (GCV) model was selected.
The GCV can be calculated as follows: linear model. The GCV of this model is 0.038 m.
N = −15.504 + 1.7169∗BF1 − 1.5416∗BF2
SSE
GCV = (6) − 2.9524∗BF3 + 2.7805∗BF4
(1 − enp/n)2
− 0.85102∗BF5 (7)
Where, SSE is the mean square error of the model in the
training stage; enp is the effective number of par- where, BF1 = 2(l |−1,46.981,47.347,47.642); BF2 = 2(w |
ameters, the default enp is between 2 and 4 (Keshtegar +1,28.843,29.092,29.409); BF3 = 2(l |+1,47.642,47.937,
et al. 2018), and n is the number of observations. 47.997); BF4 = 2(l |−1,47.997,48.057,48.226); BF5 =
Herein, no assumption for the relationship between 2(w |−1,29.409,29.727,29.898)
input and output variables; high flexibility is provided
by MARS to obtain piecewise curves. It generates
2.3.3. Gaussian process regression (GPR)
BF’s by exploring in a stepwise manner. Two stages
are used in modelling by MARS, the first is the forward GPR principals and theory are presented in Rasmussen
stage which adds functions and finds potential knots to and Williams (2006). To simplify this method to estimate
advance the accuracy, to produce an overfit model. The the geoid undulation, Richardson et al. (2017) defined the
second one is the backward stage which contains elim- GPR as a non-parametric model that is a probability dis-
inating the terms having the lowest effect (Keshtegar tribution over function, and the model in the training
et al. 2018). Figure 4a illustrates MARS model flow- stage can be built as follows:
chart. More details for MARS model design and evalu- N = f (x) + e (8)
ation can be found in Friedman (1993). In this study,
nonlinear and linear piecewise models are evaluated; where,
the results show that the two models can be used in f (x) gP(m(x), K(x, x′ )) (9)
our case, but the GCV of the nonlinear model of the
estimated model using training points and presented where, x is any finite collection of input points
in Fig. 4b and Eq.7 is shown lower than that for the x = x1 , . . . ., xn . m(x) and K(x, x′ ) are the mean and

548 Survey Review 2020 VOL 52 NO 375


Kaloop et al. Optimizing Local Geoid Undulation Model using GPS/Levelling Measurements

covariance, commonly is a kernel, functions, respect- be computed as follows:


ively. Thus, the prediction value of geoid undulation
for the inputs x∗ can be estimated by computing the a = (K + lIn )−1 N (12)
conditional distribution p(N ∗ |x∗ , x, N), and this can
where, K is a kernel function between training and testing
be calculated analytically by the standard rules for con-
data. Thus, the prediction new geoid undulation (Nx ) can
ditioning Gaussians (Murphy 2012, Richardson et al.
be estimated as follow:
2017).
In GPR model optimal values of the hyper-parameters Nx = N ′ (K + lIn )−1 K (13)
(kernel, sigma variance noise, and termination criteria)
were tuned. The RBF is tune by maximizing the marginal Here, as presented above, the kernel function and ridge
likelihood based on the trial and error basis. The pro- parameter are vital variables in this method. Algorithm
cedure of GPR geoid model is given in algorithm 2 illustrates the design KRR model for estimating Kuwait
1. The final configurations of tuned parameters sigma geoid undulation. In this study, the RBF kernel function
(0.08) and variance noise of 0.1 with a tolerance of ter- is used, the s of RBF is 0.08, and cross validation is
mination criterion of 0.001 with three-fold cross-vali- used to calculate the ridge parameter and we estimated
dation were chosen in order to obtain their higher the best value in Kuwait is 1.0e−5.
prediction with the lowest error.

Algorithm 2: KRR geoid model


Algorithm 1: GPR geoid model
Input: n labeled data input (xi , Ni ) for training; k labeled data
Input: Training set of n observatuins {(xi , Ni ), i = 1, 2, . . . , n}, points (x̂j , N̂j ) for testing; xi and x̂j are d dimensional vectors;
RBF kernel function was chosen, test inputs x∗ assuming intial MSE, i [ {1, 2, . . . , n}, j [ {1, 2, . . . , k}
Output: Test outputs N ∗ |x ∗ , x, N Output: mean square error (MSE) of prediction; best paramters
1. Estimating the optimal hyperparamter (uopt ) by maximizing the l, s.
likelihood, in each iterative step of which one need to 1. If rank =0 then
2. From a covariance matrix K(x, x ′ ) 2. Store data points (xi , Ni ) as (X , N)
3. Calculate a vector a such that Ka = N 3. M=[X,N] is a n-by-d+1 matrix
4. Calculate the likelihood 4. Shuffle M by rows
n 5. Scatter M to all the nodes evenly by rows
log p(N|x) = −0.5Na − 0.5 log |K| − log(2p) (Guo and
2 6. End if
Hesthaven, 2018) Creat a m-by-m kernel matrix K, m=n
5. Calculate the gradient of the likelihood with respect to the 7. For i=1to m do
hyperparameters 8. xi , Ni = M[i][1:d + 1]
6. Set K = K(uopt ) and a = a(uopt ) for the optimal 9. End for
hyperparameters  
10. For l = L and s = do, L and are assuming set
7. From correlation matrices K ∗∗ = K(x ∗ , x ∗ ) and K ∗ = K(x ∗ , x) paramters of l and s, respectively.
for the optimal hyperparamteres 11. For i=1to m do
8. Calculate a matrix A∗ such that KA∗ =K∗ 12. For j=1to m do

9. From the conditioning mean value vector m∗ =K Ta and the 13. K[i][j]  select RBF function
∗ ∗∗ ∗ T ∗
corresponding covariance matrix C =K -K A 14. End for
10. Define N ∗ |x ∗ , x, N  N (m ∗ , C ∗ ) 15. End for
Solve linear equation (12) for a
16. For j = 1 to k do
2.3.4. Kernel ridge regression (KRR) 17. Solve equation (13) for Nx
18. End for
KRR is a technique used to perform ridge regression
19. If rank=0 then
with a potentially infinite number of nonlinear trans- 20. Calculate MSE of the prediction geoid undulation
formations of the independent variables as regressors 21. If estimated MSE less than assumed MSE then
(Exterkate 2013). This method is used SVM and Gaus- 22. Select best paramteres l and s
sian processing to improve prediction results, and it is 23. End if
called a regularization network in the ANN (Cawley 24. End if
25. End for
et al. 2005, Exterkate 2013, Haworth et al. 2014). the
Nonlinear predicting KRR of geoid undulation is
designed in this study. The nonlinear algorithm can From above, it can be seen that there are small differ-
be estimated from linear KRR by rearranging in ences between LSSVR, KRR, and GPR methods in the
term of the weight vector w as follows (Haworth processing to estimate the geoid model; that can be sum-
et al. 2014): marized as follows (Rasmussen and Williams 2006, Well-
ing 2013): the KRR learns a linear function by the
N = Xw + e (10) respective kernel which corresponds to a non-linear func-
tion in the original space. The linear function in the kernel
w = l−1 X ′ (N − Xw) = aX ′ (11)
space is chosen based on MSE loss with ridge regulariz-
where, X represents the input variables (X = [w, l]), ation. GPR uses the kernel to define the covariance of a
and l ≥ 0 is called the ridge parameter and it is calcu- prior distribution over the target function and uses the
lated empirically using cross validation method. e is observed training data to define a likelihood function. A
the model errors, which independent of X with mean major difference is that GPR can choose the kernel’s
zero (E(e) = 0), and variance s2 (E(ee’)=s2In). hyperparameters based on gradient-ascent on the mar-
In 
training, w can be written as a linear combination as: ginal likelihood function while KRR needs to perform a
w = ni=1 ai Xi , with a = l−1 (N − Xw). Hence, a can grid search on a cross-validated loss function (MSE

Survey Review 2020 VOL 52 NO 375 549


Kaloop et al. Optimizing Local Geoid Undulation Model using GPS/Levelling Measurements

5 Observed and estimated geoid undulation of training dataset

loss). A further difference is that GPR learns a generative, n


(N pi − Noi )2
probabilistic model of the target function and can thus d = 1 − n i=1
(17)
provide meaningful confidence intervals and posterior (|N pi − Nm | + |N pi − Nm |)2
i=1
samples along with the predictions while KRR only pro- n
(N pi − Noi )2
vides predictions. Whereas, the main difference between NSE = 1 − i=1
n (18)
i=1 (Noi − Nm )
2
KRR and SVR is the different of loss functions that
used in both methods, in KRR uses MSE loss while in where, n is the number of observation and Nm, No, and Np
SVR uses linear epsilon-insensitive loss; besides, SVR is are average, observed, and predicted values of geoid
much faster in the learned model (Welling 2013). undulation, respectively.

2.4. Performance evaluation of models


Different comparison statistics are used to assess the per-
3. Results and Discussions
formance of the four models. The Root Mean Square In the training stage, the scatterplot based comparison of
Error (RMSE), Mean Absolute Error (MAE), Mean the four models is plotted in Fig. 5; in addition, the com-
Bias Error (MBE), agreement index (d) and Nash Sut- parative statistical results of different models are tabu-
cliffe Efficiency (NSE) are calculated and compared. lated in Table 3. It is clear from the figure and table that
The equations of the used criteria are given as follows: the LSSVM, GPR, and KRR perform superior accuracy
 compared to the MARS model in the training stage. In
n addition, the linear fitting equation, which presented in
i=1 (N pi − Noi )
2
RMS = (14) Fig. 5, of four models shows that the four models can
n
be used to detect the geoid undulation of Kuwait. From
n these results, it can be found that the GPR and KRR
|N pi − Noi |
MAE = i=1 (15) accuracy is acceptable to detect the geoid undulation of
n
our case, while the weakest model is MARS. Herein, it
1 n
N pi − Noi should be mentioned that the area of Kuwait is small;
MBE = (16) therefore, the accuracy of fitting models assumed to be
n i=1 Noi
high. Consequently, it can be concluded that the MARS

550 Survey Review 2020 VOL 52 NO 375


Kaloop et al. Optimizing Local Geoid Undulation Model using GPS/Levelling Measurements

Table 3 Comparison of statistical errors for models in evaluation of these models. The linear fitting equation
training stage of the models is presented in Fig. 6.
Model MAE (m) RMSE (m) MBE d NSE From Table 4, it can be seen that the MAE, RMSE, and
MBE of KRR model are 0.019, 0.024 m, and −5.19e−4,
LSSVM 0.049 0.066 −3.53e−11 0.975 0.998 respectively, and these values are the lowest for the tested
GPR 0.054 0.074 −5.30e−11 0.973 0.997 models. In addition, the d and NSE of KRR model are
KRR 0.035 0.051 −1.82e−05 0.982 0.999 0.992 and 0.999, respectively, and these values are the
MARS 0.114 0.155 −2.65e−11 0.942 0.987 highest for the other models. Furthermore, the slope of
the linear equation of KRR model is an approach to
one; it means that the accuracy of this model is high.
model may not suitable for fitting geoid undulation with a The accuracy of this model is improved by 45.45% relative
large area. to LSSVM model in term of RMSE. In contrast, the
Furthermore, it can be observed from Table 3 that the MARS model is found to be the worst model compared
five statistical parameters achieved for the training data by LSSVM model in this stage.
portray the extent at which the optimum trained models From training and testing stages, it is clearly shown that
predicted output agreed to the observed data. The com- the accuracy of Kuwait geoid model is improved by using
parative LSSVM and other models show that the devel- KRR model comparing by that estimated by LSSVM
oped KRR model is better to use in Kuwait. In model; therefore, it can be concluded that the KRR
addition, the lowest MAE, RMSE, and MBE and higher model is the best that can be used to estimate the geoid
agreement index (d ) and NSE is the KRR model com- undulation of Kuwait.
pared by other models. The accuracy of this model is To more validate the accuracy of the developed KRR
improved by 22.7% in RMSE term relative to LSSVM model, the gravity geoid model and EGM2008, EIGN-
model. Therefore, the best model can be used to estimate 6C4, and XGM-2016 GGM’s of Kuwait are used to assess
the geoid undulation in Kuwait is KRR model. the model. In this phase, the training data is used to build
In the testing stage, the four models are also compared the KRR model and the whole data are used to compare
and evaluated. Figure 6 presents the scatterplot of the the model with gravity observation model and GGM’s
four fitting models, and Table 4 shows the statistical models. Table 5 demonstrates the geoid undulation of

6 Observed and estimated geoid undulation of testing dataset

Survey Review 2020 VOL 52 NO 375 551


Kaloop et al. Optimizing Local Geoid Undulation Model using GPS/Levelling Measurements

Table 4 Comparison of statistical errors for models in Figure 7a clearly shows that the four models’ corre-
testing stage lations are high and the RMSD of them are small; the per-
Model MAE (m) RMSE (m) MBE d NSE formances of XGM2016 and EIGN-6C4 are
approximately the same into estimating Kuwait geoid
LSSVM 0.037 0.044 5.31e−6 0.984 0.999 undulation. The KRR geoid undulation has performed
GPR 0.043 0.052 −3.28e−4 0.982 0.999 significantly and its standard deviation is very close to
KRR 0.019 0.024 −5.19e−4 0.992 0.999 the gravity geoid undulation values. Figure 7b shows the
MARS 0.098 0.134 1.40e−3 0.959 0.992 error and their distribution produced by models through
five number summaries (minimum value, lower quartile
(Q1), median value (Q2), upper quartile (Q3), maximum
Table 5 Comparison of Geoid undulation model for value). From the analysis of the result, it is evident that
observed data and models (units in m)
KRR has performed significantly as its Q2, Q1, Q3 are
Model Min Max M STD very closed showing the quantile range (Q1, Q3) and
median are equal distances and following approximately
Gravity −17.96 −10.96 −15.56 1.38 normally distributed. This result signifies that KRR
GPS/Levelling −17.96 −10.96 −15.56 1.38 model is consist and produced normally distributed
EGM2008 −17.14 −10.11 −14.77 1.36
error. In addition, the whisker length and outlier points
EIGN-6C4 −17.16 −10.03 −14.66 1.40
XGM2016 −17.11 −9.94 −14.66 1.41 for KRR model is also comparatively less than the
KRR −17.96 −10.96 −15.56 1.38 GGM’s models. Furthermore, the median of KRR
model errors is close to zero. Therefore, based on the
above evidence, it can be concluded that KRR model
can be considered as a better option as a soft-computing
gravity and GPS/Levelling data and KRR and GGM’s technique for the modelling of Kuwait geoid undulation.
models. In addition, Fig. 7 presents the statistical Taylor
diagram of the models and box–whisker plot of the errors
of GGM’s and KRR models relative to gravity geoid 4. Conclusions
model. Taylor diagram is statistical summaries to evaluate The accuracy of novel LSSVR, GPR, KRR, and MARS
the degree of correspondence between values of geoid models based on geodetic coordinates inputs in modelling
undulation of gravity model and that for KRR model in geoid undulation of Kuwait were investigated in this
terms of correlation coefficient and standard deviation study. These model’s errors were compared through sev-
and Root Mean Square Difference (RMSD) (Taylor eral comparative statistics including root mean square
2001). The Pearson correlation is represented by an azi- error (RMSE), mean absolute error (MAE), mean bias
muthal angle (dash-blue contours); RMSD is pro- error (MBE), agreement index (d) and Nash Sutcliffe effi-
portional to the radial distance from the point of ciency (NSE). In the training stage, the comparative
correlation equal one (dash-green contours) and radial LSSVM and other models show that the KRR model
distance from the origin (dot-black contours) is the stan- was better to use in Kuwait. In addition, the accuracy
dard deviation of the simulated pattern. The box–whisker of this model is improved by 22.7% in terms of RMSE
plot is used to check the model’s error and its centre and relative to LSSVM model. The MAE, RMSE, and
spread of a distribution of error. MBE of KRR model were 0.019, 0.024 m, and −5.19e
From Table 5, it can be seen that the KRR model accu- −4, respectively, and d and NSE of the model were
racy is high to estimate the geoid undulation of Kuwait 0.992 and 0.999, respectively, in the testing stage. The
compared by gravity and GPS/Levelling geoid models. accuracy of this model was improved by 45.45% relative
The percentage of error accuracy of KRR model is to LSSVM model in term of RMSE. Furthermore, the
0.018 and 0.124% relative to gravity and GPS/Levelling KRR model is compared by GGMs models
geoid models, respectively, in terms of standard deviation. (EGM2008, EIGN-6C4, and XGM-2016) to validate
This indicates that the KRR model is suitable to detect the developed KRR model for using in Kuwait, and the
Kuwait geoid model. results show that KRR model can be considered as a

7 Evaluation models and models errors: (a) Taylor diagram displaying the model performance, and (b) Error distributions of
different models for geoid undulation estimation

552 Survey Review 2020 VOL 52 NO 375


Kaloop et al. Optimizing Local Geoid Undulation Model using GPS/Levelling Measurements

better option as a soft-computing technique for the mod- Huang, J. and Véronneau, M., 2013. Canadian gravimetric geoid model 2010.
elling of Kuwait geometric geoid undulation. Finally, for Journal of geodesy, 87 (8), 771–790. doi:10.1007/s00190-013-0645-0.
Kaloop, M.R., et al., 2018. Using advanced soft computing techniques
future study, an integration for land and ocean data is rec- for regional shoreline geoid model estimation and evaluation.
ommended for improving the geoid model, especially on Marine georesources and geotechnology, 36 (6), 688–697. doi:10.
the shore region. 1080/1064119X.2017.1370622.
Kao, S.-P., et al., 2014. Using a least squares support vector machine
to estimate a local geometric geoid model. Boletim de Ciências
Acknowledgment Geodésicas, 20 (2), 427–43. doi:10.1590/s1982-21702014000200025.
Kavzoglu, T. and Saka, M.H., 2005. Modelling local GPS/levelling geoid
This work was supported by Basic Science Research Pro- undulations using artificial neural networks. Journal of geodesy, 78,
gram through the National Research Foundation of 520–27. doi:10.1007/s00190-004-0420-3.
Keshtegar, B., Mert, C., and Kisi, O., 2018. Comparison of four heuristic
Korea (NRF) funded by the Ministry of Science, ICT & regression techniques in solar radiation modeling: kriging method
Future Planning (2019R1I1A1A01062202). vs RSM, MARS and M5 model tree. Renewable and sustainable
energy reviews. doi:10.1016/j.rser.2017.07.054.
Kopsiaftis, G., et al., 2019. Gaussian process regression tuned by
Funding Bayesian optimization for seawater intrusion prediction.
Computational intelligence and neuroscience, 2019, 1–12. doi:10.
This work was supported by Ministry of Science, ICT and Future Plan- 1155/2019/2859429.
ning: [Grant Number 2019R1I1A1A01062202]. Kumar, J. and Nath, U., 2016. Determination of geoid undulation by
astro-geodetic method. Journal of surveying engineering, 142 (3),
05015007. doi:10.1061/(asce)su.1943-5428.0000152.
ORCID Murphy, K.P., 2012. Machine learning: A probabilistic perspective (adap-
tive computation and machine learning series). Cambridge, MA:
Mosbeh R. Kaloop http://orcid.org/0000-0002-8449- MIT Press.ISBN 9780262018029.
8883 Peprah, M.S. and Kumi, S.A., 2017. Appraisal of methods for estimating
orthometric heights – A case study in a mine. Journal of geosciences
and geomatics, 5 (3), 96–108. doi:10.12691/jgg-5-3-1.
References Pikridas, C., et al., 2011. Estimation and evaluation of GPS geoid heights
using an artificial neural network model. Applied geomatics, 3 (3),
Adescu, O.B., et al., 2016. Astro-Geodetic platform for high accuracy 183–87. doi:10.1007/s12518-011-0052-2.
geoid. Romanian astron. J., 26 (2), 57–65. Rabah, M. and Hattab, A., 2009. Can the New develop earth gravita-
Akyilmaz, M.O., et al., 2009. Soft computing methods for geoidal height tional model 2008 Be used as a geoid model for Kuwait? ESRI-
transformation. Earth, planets and space, 61 (7), 825–33. Faculty of engineering at shoubra, 11, 38–51.
Becker, M. 2012. Geodesy. Springer handbook of geographic information. Rabah, M. and Kaloop, M., 2013. The Use of minimum curvature sur-
Springer Handbook of Geographic Information. doi:10.1007/978- face technique in geoid computation processing of Egypt.
3-540-72680-7_8. Arabian journal of geosciences, 6 (4), 1263–72. doi:10.1007/
Cawley, G.C., Talbot, N.L.C., and Chapelle, O. 2005. Estimating predic- s12517-011-0418-0.
tive variances with kernel ridge tegression. In MLCW 2005: Rasmussen, C.E. and Williams, C.K.I., 2006. Regression. In: Gaussian
machine learning challenges. Evaluating predictive uncertainty, processes for machine learning. MIT Press, 7–31. doi:10.1093/
visual object classification, and recognising tectual entailment, 56– bioinformatics/btq657.
77. https://doi.org/10.1007/11736790_5. Richardson, R.R., Osborne, M.A., and Howey, D.A., 2017. Gaussian
El-Ashquer, M., et al., 2019. Study on the selection of optimal global geo- process regression for forecasting battery state of health. Journal
potential models for geoid determination in Kuwait. Survey review. of power sources, 357, 209–19. doi:10.1016/j.jpowsour.2017.05.
doi:10.1080/00396265.2019.1611256 To. https://doi.org/10.1080/ 004.
00396265.2019.1611256. Seager, J., Collier, P., and Kirby, J. 1999. Modelling Geoid Undulations
Elshambaky, H., 2018. Application of neural network technique to deter- with an Artificial Neural Network. In IJCNN’99. International
mine a corrector surface for global geopotential model using GPS/ Joint Conference on Neural Networks, Washington, DC, USA,
levelling measurements in Egypt. Journal of applied geodesy, 12 (1), USA. doi:10.1109/ijcnn.1999.836195.
29–43. doi:10.1515/jag-2017-0017. Taylor, K.E., 2001. Summarizing multiple aspects of model performance
Elshambaky, H., 2019. Enhancing the predictability of least-squares col- in a single diagram. Journal of geophysical research: atmospheres,
location through the integration with least-squares-support vector 106 (D7), 7183–92. doi:10.1029/2000JD900719.
machine. Journal of applied geodesy, 13 (1), 1–15. doi:10.1515/jag- Tierra, A., Dalazoana, R., and De Freitas, S., 2008. Using an artificial
2018-0017. neural network to improve the transformation of coordinates
Elshambaky, H.T., Kaloop, M.R., and Hu, J.W., 2018. A novel three- between classical geodetic reference frames. Computers and geos-
direction datum transformation of geodetic coordinates for Egypt ciences, 34 (3), 181–89. doi:10.1016/j.cageo.2007.03.011.
using artificial neural network approach. Arabian journal of geos- Uotila, U.A., 1971. Physical geodesy. Eos, transactions American geophy-
ciences, 11, 110. doi:10.1007/s12517-018-3441-6. sical union. Eos, Transactions American Geophysical Union.
Erol, B. and Celik, R.N. 2004. Modelling local GPS/Levelling geoid with doi:10.1029/EO052i003pIU012.
the assesstment of inverse distance weighting and geostatistical kri- Veronez, M.R., et al., 2011. Regional mapping of the geoid using GNSS
ging methods. Proceedings of XXth ISPRS congress, technical com- (GPS) measurements and an artificial neural network. Remote sen-
mission IV, Vol. XXXV, 76, Istanbul, Turkey. http://cartesia.org/ sing, 3 (4), 668–83. doi:10.3390/rs3040668.
geodoc/isprs2004/comm4/papers/319.pdf. Volgyesi, L., et al., 2005. Application of kernel ridge regression to net-
Erol, B. and Erol, S., 2013. Learning-Based computing techniques in work levelling via mathematica. Baztech, 2005, http://yadda.icm.
geoid modeling for precise height transformation. Computers & edu.pl/baztech/element/bwmeta1.element.baztech-article-PWA6-
geosciences, 52, 95–107. 0023-0018.
Exterkate, P., 2013. Model selection in kernel ridge regression. Wang, Y., et al., 2012. The US gravimetric geoid of 2009 (USGG2009):
Computational statistics and data analysis, 68, 1–16. doi:10.1016/j. model development and evaluation. Journal of geodesy, 86 (3),
csda.2013.06.006. 165–180. doi:10.1007/s00190-011-0506-7.
Friedman, J.H. 1993. Fast MARS, department of statistics, Stanford Wang, J., Hu, Y., and Zhou, J., 2009. Combining model for regional GPS
University, Tech. Report LCS110. Department of statistics, height conversion based on least squares support vector machines.
Stanford University, Tech. Report LCS110. Proceedings - 2009 international conference on environmental science
Guo, M. and Hesthaven, J.S., 2018. Reduced order modeling for nonlinear and information application technology, ESIAT 2009, 2 (2), 639–41.
structural analysis using Gaussian process regression. Computer doi:10.1109/ESIAT.2009.182.
methods in applied mechanics and engineering, 341, 807–826. Welling, M. 2013. Kernel ridge regression. Max Welling’s Classnotes in
Haworth, J., et al., 2014. Local online kernel ridge regression for forecast- Machine Learning, 1–3, available online (August 28, 2019):
ing of urban travel times. Transportation research part C: emerging https://www.ics.uci.edu/~welling/classnotes/papers_class/Kernel-
technologies, 46, 151–78. doi:10.1016/j.trc.2014.05.015. Ridge.pdf.

Survey Review 2020 VOL 52 NO 375 553


Kaloop et al. Optimizing Local Geoid Undulation Model using GPS/Levelling Measurements

Yakubu, C.I., Ferreira, V.G., and Asante, C.Y., 2017. Towards the selec- Journal of applied physiology, 108 (1), 128–36. doi:10.1152/
tion of an optimal global geopotential model for the computation japplphysiol.00729.2009.
of the long-wavelength contribution: A case study of Ghana. Zaletnyik, P., Völgyesi, L., and Paláncz, B., 2008. Modelling local GPS/
Geosciences, 7 (4), 113 (1–12). levelling geoid undulations using support vector machines.
Yang, Z.J. and Chen, Y.Q., 1999. Determination of local geoid with geo- Periodica polytechnica civil engineering, 52 (1), 39–43. doi:10.
metric method: case study. Journal of surveying engineering, 125 (3), 3311/pp.ci.2008-1.06.
136–46. doi:10.1061/(ASCE)0733-9453(1999)125:3(136). Zhang, W. and Goh, A.T.C., 2016. Multivariate adaptive regression
Yilmaz, M., 2010. Adaptive network based on fuzzy inference system splines and neural network models for prediction of pile drivability.
estimates of geoid heights interpolation. Scientific research and Geoscience frontiers, 7 (1), 45–52. doi:10.1016/j.gsf.2014.10.003.
essays, 5 (16), 2148–54. Ziggah, Y., et al., 2016. Capability of artificial neural network for forward
Yuan, F.C. and Lee, C.H., 2015. Using least square support vector conversion of geodetic coordinates (f, λ, h) to cartesian coordinates
regression with genetic algorithm to forecast beta systematic risk. (X, Y, Z). Mathematical geosciences, 48 (6), 687–721. doi:10.1007/
Journal of computational science, 11, 26–33. doi:10.1016/j.jocs. s11004-016-9638-x.
2015.08.004. Ziggah, Y.Y. and Laari, P.B., 2018. Application of multivariate adaptive
Zakeri, I.F., et al., 2010. Multivariate adaptive regression splines models regression spline approach for 2D coordinate transformation ∗ .
for the prediction of energy expenditure in children and adolescents. Ghana journal of technology, 2 (2), 50–62.

554 Survey Review 2020 VOL 52 NO 375

You might also like