Optimization Model for GPS Height Fitting Considering of AIC

Lingli Zhao,Shuai Liu , Junsheng Li, Haicheng Xu School of Engineering, Honghe University, Mengzi, P.R.China ,661100 e-mail:zll_csu@126.com Lingli Zhao, Shuai Liu School of Info-Physics and Geomatics Engineering, Central South University, Changsha , P.R.China,410083

Abstract—Polynomial fitting method is widely used in GPS height conversion into normal heights in survey engineering and whcih has become a hotspot. Different models used in different area of interpolation have different precisions, therefore, it is very important to optimize model before GPS height conversion in order to obtain high precision. Lots of engineering practice has proven that The conicoid model could satisfied forth degree demand, but there is no proof why the conicoid model is better. So, the paper introduce AIC to choose the better model for transforming GPS height, we separate the value of AIC into two parts, AICin and AICout, then add them up to obtain the total value, AICtotal, which could testify which model is the best model comparing with the relative information. The experiments show that the optimization model for GPS height fitting considering of AIC criterion is valid and something useful is concluded. Keywords- AIC; GPS height; optimization model; polynomial fitting

and N of these known points, N=f(X,Y) , using this model we could calculate the N of other GPS points ,namely, the points of interpolation by their plan coordinates X and Y. Then, we could get the H by the formula H=h-N. [2][3][4][5][6] There are many mathematical models in fitting methods, different models used in different area of interpolation have different precisions, therefore, it is very important to optimize model before GPS height conversion in order to obtain high precision. Lots of engineering practice has proven that The conicoid model could satisfied forth degree demand, but there is no proof why the conicoid model is better. Akaike's information criterion [7], developed by Hirotsugu Akaike under the name of "an information criterion" (AIC) in 1971 and proposed in Akaike (1974), is a measure of the goodness of fit of an estimated statistical model. So, the paper introduce AIC to choose the better model for transforming GPS height, we separate the value of AIC into two parts, AICin and AICout, then add them up to obtain the total value, AICtotal, which could testify which model is the best model comparing with the relative information. The experiments show that the optimization model for GPS height fitting considering of AIC criterion is valid and something useful is concluded.

I.

INTRODUCTION

Knowledge of the position of the geoid with respect to the WGS84 ellipsoid is required to transform GPS (Globe Position System)-derived ellipsoidal heights (h) to physically meaningful orthometric heights, in the application of survey engineering, regardless of the [small] curvature of the plumblines, adopting normal heights (H) to stand for orthometric heights while quasi-geoid (N) stand for geoid, we could get the formula H = h-N (1) However, the Quasi-geoid (N) could not be obtained directly, and we could not transfer GPS heights (h) to normal heights (H) accurately. Conventional geometry leveling is the main method to determine H, but this method has poor efficiency. GPS positioning has the quality of high precision in position, rapid speed, and simple manipulation etc. Therefore, how to utilize GPS positioning to deal with the heights problem has become a hotspot in survey engineering. GIS fitting method is adopted in GPS heights conversion in a small certain area. A solution is sought the points where colocated by GPS (getting h) and geometry leveling (getting H) in certain area, and we could get the values of Quasi-geoid of these points (N=h-H) [1]. We call these points GPS/leveling points, namely the known points. In this method, we consider the Quasi-geoid in a certain area of interpolation to be a polynomial surface or other mathematical model, then we can construct the mathematical model by the plan coordinates X, Y

II.

POLYNOMIAL FITTING MODELS

Supposed that one point A(x, y) in an area of interpolation, its Quasi-geoid value is N, they relationship could be expressed:

N = f ( x, y ) + ε f ( x, y ) is fitting function, ε is residual error.
f (x, y) = a0 + a1x + a2 y + a3xy+ a4x2 + a5 y2 + a6x2 y + a7 y2x + a8x3 + a9 y3 +
If there are n points, we could get the below matrix

(2)

(3)

N = AX + ε

4

978-1-4244-4994-1/09/$25.00 ©2009 IEEE

[8] In the general case.Figure 2 and Figure 3 respectively. Figure 3. the value of AIC is lager and the model is better. the area is about 500 km2. m=10 if adopting three degree model. . The plane model IV. if unknown number is one degree term. rather it is a tool for model selection. called three degree model fitting. and however. called the conicoid model fitting. Akaike's information criterion. It is grounded in the concept of entropy. then back to formula (3). in effect offering a relative measure of the information lost when a given model is used to describe reality and can be said to describe the tradeoff between bias and variance in model construction. There are two kinds of errors in GPS height conversion. From the AIC value one may infer that e. Figure 2. we could get any points’ Quasi-geoid value and obtain normal heights. the AIC is AIC = log σ 2 + (2m + 3 / n) (5 Where m is the number of parameters in the statistical model and n is the number of the calculation for the estimated model. if unknown number is quadratic term. or loosely speaking that of precision and complexity of the model [7]. with the one having the lowest AIC being the best. the other is from the checking points. 2 Figure 1.1 x1 1 x2 A = 1 x3 1 xn a0 a1 X = a2 an y1 y2 y3 yn x1 y1 x2 y2 x3 y3 xn yn Where ε0 ε1 ε = ε2 εn III. the one is from the known points. In formula (3). but one should not assign a value above which a given model is 'rejected'. Three degree model AIC AND GPS HEIGHT FITTING MODELS We could get the values of X under least square . The conicoid model EXPERIMENTS A city area made a E-class GPS control net. the number of the control points is 153 and the number of GPS/leveling points of these control points is 58. For every chosen model. they are AICin and AICout respectively. σ is residual sum of squares. In GPS height conversion m=3 if adopting plane model.g the top three models are in a tie and the rest are far worse. we could calculate two values of AIC by formula (5). they geometrical models are showed in Figure 1 . n=20 for the three models. The AIC is not a test on the model in the sense of hypothesis testing. several competing models may be ranked according to their AIC. called plane model fitting. Given a data set. If we choose 20 GPS/leveling points. developed by Hirotsugu Akaike under the name of "an information criterion" (AIC) in 1971 and proposed in Akaike (1974). Add both of AIC values to justify which model is better. if unknown number is three degree term. is a measure of the goodness of fit of an estimated statistical model. m=6 if adopting conicoid model. we should take both of them into account to optimize a model and obtain which model is better.

17 3. V.04 -2.36 1. and the three degree model needs more GPS/leveling points . conicoid model and three degree model. conicoid model and three degree model.03 -.P C..34 -.69 1.29 1.18 8.76 13.88 13.P and C.36 17.09 21. and the results are showed in table .30 38. max .07 .P stands for the known points and the checking points respectively. we could see that the best model is the three degree model. the other four points as checking points and the results are showed in table . From the precision comparability. choosing 11 of them distributed equably in area of interpolation as known points to optimize plane model. Table and table give the known points’ and the checking points’ mean square errors.P Mean square errors (cm) K.58 3. Model Mean square errors (cm) K.69 11. Figure 4.21 16.51 .78 12.P Max errors (cm) K.We pick up 13 GPS/leveling points showed in Figure 4.06 17. errors and AIC values.12 . Mean square errors are AIC values are calculated by formula (5).03 -. K.09 Table II. the three degree model has better precision than the conicoid model’ in known points. while the conicoid model has better precision than the three degree model’ in checking points. which one is the best model. the other two points as checking points.97 .P C.66 25. we could see that the AIC as optimization model for GPS height fitting is valid.80 5.P Max errors (cm) K.P AICin AIC values AICout AICtotal plane conicoid three degree 3.45 24.P AICin AIC values AICout AICtotal plane conicoid 3. What we should pay attention is that both the three . which one is the best model.7 . Test 2: There are 13 GPS/leveling points showed in Figure 4. which has more workflow.72 . Table I. the precision is a little better.37 4. but we need to have more GPS/leveling points and increase workflow. minimum errors.21 -5. The distribution of 13 GPS/leveling points V T PV r Test 1: There are 13 GPS/leveling points showed in Figure 4. CONCLUSIONS To Sum up.69 -5. if we choose the three degree model .P C. From table . Moreover.30 From table .24 8. we could see that the conicoid model is better that the plane model.30 8.51 1.P C.21 -2.73 8. Model Three models’ precisions comparability and AIC values in test 1 Minimum errors (cm) K.44 -.61 -. the fitting precisions is much related with fitting models.01 . and make two tests to optimize model. AIC testifies it is valid to optimize model for GPS height fitting.25 . then the conicoid model and the plane model.03 -. choosing 9 of them distributed equably in area of interpolation as known points to optimize plane model.P Two models’ precisions comparability and AIC values in test 2 Minimum errors (cm) K. and it is not obvious to improve precision compared with the conicoid model.P C.67 -2.31 .P C.

1109/TAC. IEEE Transactions on Automatic Control 19 (6): 716– 723. 5. M. Hinton.1100705. McCulloch. The MIT Press. The paper introduce AIC criterion to choose the better model for transforming GPS height. doi:10. Österreichische Beiträge zu Meteorologie und Geophysik. A logical calculus of the ideas immanent in nervous activity. & Papert. 115-133. (1986). Hirotugu (1974).. AICin and AICout. Rumelhart. pp. Minsky. E. Beyond Regression: New Tools for Prediction and Analysis in the Behav-ioral Sciences. & Pitts. separating the value of AIC into two parts. G. Perceptrons: An Introduction to Computational Geometry.Survey Review 38(301): 573-582.degree model and the conicoid model have one convexity or concave.1974. S. E. Proceedings of the 7th International Meeting on Alphine Gravimetry. Kalmár J 1996: Interpretation of local geoid undulations in the Pannonian Basin. http://en. Unpublished doctoral dissertation. The experiments show that the optimization model for GPS height fitting considering of AIC is valid and we get a conclusion the conicoid model is better model which has widely used in survey engineering. then add them up to obtain the total value.. Harvard University. (1943). Werbos. AICtotal. W. 95-96.. R. (1974). [2] [3] [4] [5] [6] REFERENCES [1] Featherstone WE.wikipedia. Bulletin of Mathematical Biophysics. (1969). Papp G. we should guarantee that there be one convexity or concave in fitting as possible otherwise the fitting model may be fuzzy and its result will not optimize the fitting model by AIC. S. J. Learning representations by backpropagating errors. 323. which could testify which model is the best model comparing with the relative information in the experiments. & Williams. D. P. Sproule DM (2006). MR0423716. . "A new look at the statistical model identification". 533-536. Nature.org/wiki/Akaike_information_criterion Akaike. Fitting AUSGeoid98 to the Australian Height Datum using GPS data and least squares collocation: [7] [8] application of a cross-validation technique.. J. W.