Professional Documents
Culture Documents
Kriging and Moving Window Kriging On A Sphere in Geometric GNSS Levelling Geoid Modelling
Kriging and Moving Window Kriging On A Sphere in Geometric GNSS Levelling Geoid Modelling
Kriging and Moving Window Kriging On A Sphere in Geometric GNSS Levelling Geoid Modelling
To cite this article: M. Ligas & M. Kulczycki (2018) Kriging and moving window kriging on a
sphere in geometric (GNSS/levelling) geoid modelling, Survey Review, 50:359, 155-162, DOI:
10.1080/00396265.2016.1247131
(Stein 1999). Owing to its versatility it has found a promi- Finally, we minimise equation (10) subject to the unbia-
nent place among prediction methods used in different sedness condition in equation (9) and this leads to the sys-
fields of science. There are two basic principles behind tem of equations to be solved for optimal coefficients λ
kriging, i.e. unbiasedness and minimisation of mean and a scalar Lagrange multiplier κ:
squared error of prediction, and thus it is equivalent to
V 1 l v
the minimum error variance prediction. Here, we briefly = (11)
recall the basic facts concerning ordinary and universal 1T 0 k 1
kriging-based prediction (since these methods have been On the basis of optimised coefficients λ and a scalar κ, we
used in this study). We consider the following general obtain the ordinary kriging variance (prediction variance)
decomposition of a random field (collection of random of the form
variables, also a random function or a spatial stochastic
process): s2OK = s2 − lT v − k (12)
Z(s) = m(s) + 1(s) for observables Ordinary kriging predictor may be formulated explicitly
(3)
Z(s0 ) = m(s0 ) + 1(s0 ) for unobservables by using the inverse of the partitioned matrix on the
left-hand side of equation (11) and solving the system of
where Z(s) is a vector of observed quantities at locations s, equations for λ and κ. Hence, after some manipulations
μ(s) is a vector of means, either a constant (ordinary kri- one arrives at the explicit form of the predictor, i.e.
ging) or a location-dependent trend/drift (universal kri-
ging), ε(s) is a vector of spatially correlated disturbances, Ẑ(s0 ) = m̂ + vT V−1 [Z − 1m̂ ] with
Z(s0) is a single quantity to be predicted at a location s0, −1 −1 T −1
(13)
μ(s0) is mean value at a location s0, either a constant or m̂ = (1 V 1) 1 V Z
T
resulting from the adopted trend and ε(s0) is a disturbance and then the prediction variance reads
at a location s0, spatially correlated with disturbances ε(s).
We shall also assume the following stochastic charac- s2OK = s2 − vT V−1 v + (1 − vT V−1 1)
terisation, e.g. Goldberger (1962) (we drop a denotation (14)
of location ‘s’ that should not lead to any confusion): × (1T V−1 1)−1 (1 − vT V−1 1)T
where λ is the vector of coefficients of a linear combi- Unbiasedness condition is obtained from the same con-
nation of data – ‘kriging weights’. dition – zero prediction error on average, i.e.
Unbiasedness condition is obtained from equating a E(e) = E(Ẑ0 − Z0 ) = 0 (17)
mean prediction error to zero, i.e.
This leads to the condition
E(e) = E(Ẑ0 − Z0 ) = 0 (8)
l T X = xT (18)
what leads to the condition
Further, we construct a mean squared error of prediction
lT 1 = 1 (9) which reads
Further, we construct a mean squared error of prediction E(eeT ) = lT Vl − 2lT v + s2 (19)
which is expressed as
Now, we minimise equation (19) subject to the unbiased-
E(eeT ) = lT Vl − 2lT v + s2 (10) ness condition in equation (18) and this leads to a system
Results
4 Spatial distribution of ‘GPS on benchmarks’ training data
for the TF2 region The quality of kriging prediction for both a classical
approach (denoted as cK in the following tables) and a
moving window approach (denoted as MWK in the fol-
training set and a test set. Each training set consists of lowing tables) has been assessed for two variants of data.
points on which kriging models have been constructed. The first one concerned raw undulations (without the
Test sets, on the other hand, served only for the true vali- use of EGM08 model) and the second one used residual
dation purposes, to obtain reasonable prediction error undulations (after removal of undulations derived from
characteristics. The main characteristics of the mentioned EGM2008 model and adding them back after the pre-
regions with respect to the spatial extent and number of diction procedure). Always a better variant either
N N RES
- root mean square prediction error (RMSPE) Max [m] 1.0359 0.7453 0.1315 0.1256
MAPE [m] 0.1463 0.1062 0.0295 0.0281
n p 2 RMSPE [m] 0.2953 0.2033 0.0427 0.0396
i=1 (Ni − Ni )
RMSPE = (37) MeAPE [m] 0.0434 0.0314 0.0191 0.0183
n Corr. Coef. 0.994118 0.997237 0.999879 0.999895
Table 2 Summary statistics for geoid undulation data for three regions
Statistics N WGS [m] N RES [m] N WGS [m] N RES [m] N WGS [m] N RES [m]
N N RES
cK MWK cK MWK
Interval [cm] n na % n na % n na % n na %
[0, 1] 211 211 55.4 216 216 56.7 235 235 61.7 236 236 61.9
(1, 2] 94 305 80.1 93 309 81.1 85 320 84.0 81 317 83.2
(2, 3] 40 345 90.6 35 344 90.3 39 359 94.2 39 356 93.4
(3, 4] 16 361 94.8 15 359 94.2 12 371 97.4 15 371 97.4
(4, 5] 7 368 96.6 10 369 96.9 4 375 98.4 6 377 99.0
(5, 10] 13 381 100.0 12 381 100.0 6 381 100.0 4 381 100.0
N N RES
cK MWK cK MWK
Interval [cm] n na % n na % n na % n na %
(calculation time) is not rewarded with a significant leave no illusions (see Tables 4 and 7 for TF2 and Tables
improvement in the result (see Table 3). 5 and 8 for TF3). All summary statistics and the distri-
However, when interpolating raw geoid heights N the bution of absolute errors within particular intervals advo-
results demonstrate the superiority of MWK over cK, cate for MWK. The only exception is the maximum value
but differences at particular statistics are counted in tenths of differences when interpolating N RES (see Table 4)
of millimetres. When interpolating residual undulations which is 1 mm better for cK than MWK. In the presented
N RES a vestigial deterioration of summary statistics is tables, attention should be paid to significant improve-
noticeable. Similarly, when analysing individual intervals ment in interpolation results obtained using residual
(see Table 6) it is difficult to confidently identify MWK as undulations based on EGM2008. For TF2, one immedi-
giving better results. In contrast, for less numerous data- ately notices approximately, 50% improvement in accu-
sets where the data points are additionally arranged at racy in relation to raw undulations interpolation almost
greater distances (a much larger spatial extent of a on all error characteristics. Also the classification into
domain), as is in the cases of TF2 and TF3, the results particular error intervals changes significantly. For TF3,
N N RES
cK MWK cK MWK
Interval [cm] n na % n na % n na % n na %
the gain in accuracy resulting from EGM2008 incorpor- 11.11.150.006 in the Department of Geomatics, AGH
ation is even better and varies, roughly speaking, from University of Science and Technology, Krakow.
three to six times depending on the statistics.
Conclusions References
In this paper, we investigated prediction capabilities of Chiles, J.P. and Delfiner, P., 1999. Geostatistics – modeling spatial uncer-
spherical (performed on a sphere) kriging method in tainty. New York: John Wiley & Sons.
two variants: cK (with a global semi-variance function/ Featherstone, W.E., Dentith, M.C., and Kirby, J.F., 1998. Strategies for
covariance function) and an MWK (with a local semi- the accurate determination of orthometric heights from GPS.
Survey review, 34 (267), 278–296.
variance function/covariance function). The research Gneiting, T., 2013. Strictly and non-strictly positive definite functions on
has been conducted on three test fields (being on the ter- spheres. Bernoulli, 19 (4), 1327–1349.
ritory of the conterminous USA) characterised by differ- Goldberger, A.S., 1962. Best linear unbiased prediction in the generalized
ent spatial extent and density of sampling. The linear regression model. Journal of the American statistical associ-
ation, 57 (298), 369–375.
comparison of two variants of kriging was performed
Haas, T.C., 1990. Kriging and automated variogram modeling within a
on two kinds of input data. We used raw undulation mowing window. Atmospheric environment, 24A (7), 1759–1769.
data and residual undulation data based on EGM2008 Heiskanen, W.A. and Moritz, H., 1967. Physical geodesy. San Francisco:
geopotential model. Numerical tests conducted in this W.H. Freeman and Co.
study revealed that in case of high-sampling density, as Huang, C., Zhang, H. and Robeson, S.M., 2011. On the validity of com-
monly used covariance functions and variogram functions on the
it was in the case of test field 1, there was no gain when sphere. Math geoscience, 43, 721–733.
using MWK instead of cK. On the other hand, in case Kavzoglu, T. and Saka, M.H., 2005. Modelling local GPS/levelling geoid
of TF1 we obtained a visible accuracy gain for both cK undulations using artificial neural networks. Journal of geodesy, 78,
and MWK when we incorporated EGM2008-based undu- 520–527.
Kotsakis, C. and Katsambalos, K., 2010. Quality analysis of global geo-
lations into the interpolation procedure. In contrast, for
potential models at 1542 GPS/leveling benchmarks over the
less numerous datasets and a much larger spatial extent, Hellenic mainland. Survey review, 42 (318), 327–344.
as is in the cases of TF2 and TF3, the results leave no illu- Moritz, H., 1976. Covariance functions in least – squares collocation.
sions. MWK adapts itself to data much better than cK. Report no. 240, The Ohio State University, Department of
Incorporation of EGM2008, i.e. working with residual Geodetic Science.
Moritz, H., 1980. Advanced physical geodesy. Karlsruhe: Herbert
geoid heights gives a significant improvement in compari- Wichmann Verlag.
son to the raw undulations interpolation. Thus, the Pavlis, N., Holmes, S.A., Kenyon, S.C. and Factor, J.K., 2012. The devel-
research proves that the stochastic combination of opment and evaluation of the Earth Gravitational Model 2008
MWK and global geopotential model EGM2008 with (EGM2008). Journal of geophysical research: solid earth, 117
(B04406), 1–38.
raw undulation data has a potential to significantly
Roy, R., 1973. Estimation of the covariance function of a homogeneous
improve the final geometric geoid model even for large process on the sphere. The annals of statistics, 1 (4), 780–785.
areas and a sparse distribution of points. Schlather, M., 2015. Package ‘random fields’. In: Simulation and analysis
of random fields. cran.r-project.org/web/packages/RandomFields/
RandomFields.pdf.
Acknowledgements Schoenberg, I.J., 1942. Positive definite functions on spheres. Duke math
journal, 9 (1), 96–108.
The paper is the result of research on geospatial methods Stein, M., 1999. Interpolation of spatial data – some theory for kriging.
carried out within statutory research grant No. New York: Springer.