Search

Minimal Detectable Biases (MDBs) or Minimal Detectable Outliers for the Expectation Maximization (EM) algorithm based on the variance-inflation and the mean-shift model are determined for an example. A Monte Carlo method is applied with no outlier and with one, two and three randomly chosen outliers. The outliers introduced are recovered and the corresponding MDBs are almost independent from the number of outliers. The results are compared to the MDB derived earlier by the author. This MDB approximately agrees with the MDB for one outlier of the EM algorithm. The MDBs for two and three outliers are considerably larger than MDBs of the EM algorithm.

Artificial intelligence is interpreted by a machine learning algorithm. Its realization is applied for a two-dimensional grid of points and depends on six parameters which determine the limits of loops. The outer loop defines the width of the grid, the most inner loop the number of scans, which result from the three-dimensional coordinate system for the $$x_i$$-, $$y_i$$-, $$z_i$$-coordinates of a laser scanner. The $$y_i$$-coordinates approximate the distances measured by the laser scanner. The minimal standard deviations of the measurements distorted by systematic effects for the $$y_i$$-coordinates are computed by the Monte Carlo estimate of Sect. 6. The minimum of these minimal standard deviations is found in the grid of points by the machine learning algorithm and used to judge the outcome. Two results are given in Sect. 7. They differ by the widths of the grid and show that only for precise applications the systematic effects of the laser scanner have to be taken care of. Instead of assuming a standard deviation for the systematic effects from prior information as mentioned in Sect. 1, the $$x_i$$-, $$y_i$$-, $$z_i$$-coordinates are repeatedly measured by the laser scanner. However, there are too few repetitions to fulfill the conditions of the multivariate model of Sect. 2 for all measured coordinates. The variances of the measurements plus systematic effects computed by the Monte Carlo estimate of Sect. 6 can therefore be obtained for a restricted number of points only. This number is computed by random variates. For two numbers, the variations of the standard deviations of the $$y_i$$-coordinates, the variations of the standard deviations of the $$x_i$$-, $$y_i$$-, $$z_i$$-coordinates from the multivariate model, the variations of the standard deviations of the systematic effects and the variations of the confidence intervals are presented. The repeated measurements define time series whose auto- and cross-correlation functions are applied as correlations for the systematic effects of the measurements. The ergodicity of the time series is shown.

The concept of reliability was introduced into geodesy by Baarda (A testing procedure for use in geodetic networks. Publications on Geodesy, vol. 2. Netherlands Geodetic Commission, Delft, 1968). It gives a measure for the ability of a parameter estimation to detect outliers and leads in case of one outlier to the MDB, the minimal detectable bias or outlier. The MDB depends on the non-centrality parameter of the \(\chi ^2\)-distribution, as the variance factor of the linear model is assumed to be known, on the size of the outlier test of an individual observation which is set to 0.001 and on the power of the test which is generally chosen to be 0.80. Starting from an estimated variance factor, the \(F\)-distribution is applied here. Furthermore, the size of the test of the individual observation is a function of the number of outliers to keep the size of the test of all observations constant, say 0.05. The power of the test is set to 0.80. The MDBs for multiple outliers are derived here under these assumptions. The method is applied to the reconstruction of a bell-shaped surface measured by a laser scanner. The MDBs are introduced as outliers for the alternative hypotheses of the outlier tests. A Monte Carlo method reveals that due to the way of introducing the outliers, the false null hypotheses cannot be rejected on the average with a power of 0.80 if the MDBs are not enlarged by a factor.

For deriving the robust estimation by the EM (expectation maximization) algorithm for a model, which is more general than the linear model, the nonlinear Gauss Helmert (GH) model is chosen. It contains the errors-in-variables model as a special case. The nonlinear GH model is difficult to handle because of the linearization and the Gauss Newton iterations. Approximate values for the observations have to be introduced for the linearization. Robust estimates by the EM algorithm based on the variance-inflation model and the mean-shift model have been derived for the linear model in case of homoscedasticity. To derive these two EM algorithms for the GH model, different variances are introduced for the observations and the expectations of the measurements defined by the linear model are replaced by the ones of the GH model. The two robust methods are applied to fit by the GH model a polynomial surface of second degree to the measured three-dimensional coordinates of a laser scanner. This results in detecting more outliers than by the linear model.

A mixture of normal distributions is assumed for the observations of a linear model. The first component of the mixture represents the measurements without gross errors, while each of the remaining components gives the distribution for an outlier. Missing data are introduced to deliver the information as to which observation belongs to which component. The unknown location parameters and the unknown scale parameter of the linear model are estimated by the EM algorithm, which is iteratively applied. The E (expectation) step of the algorithm determines the expected value of the likelihood function given the observations and the current estimate of the unknown parameters, while the M (maximization) step computes new estimates by maximizing the expectation of the likelihood function. In comparison to Huber’s M-estimation, the EM algorithm does not only identify outliers by introducing small weights for large residuals but also estimates the outliers. They can be corrected by the parameters of the linear model freed from the distortions by gross errors. Monte Carlo methods with random variates from the normal distribution then give expectations, variances, covariances and confidence regions for functions of the parameters estimated by taking care of the outliers. The method is demonstrated by the analysis of measurements with gross errors of a laser scanner.

GOCO02S is a combined satellite-only geopotential model, regularized from degrees 180 to 250 of the expansion into spherical harmonics. To investigate the start of the regularization, the normal equations of GOCO02S have been used to compute additional geopotential models by regularizations beginning at degrees 160, 200, 220 and with no regularization. Three different methods are applied to determine where to start the regularization. The simplest one considers the decrease of the degree variances of the not regularized solution. The second one tests for the same solution the hypothesis that the square root of the degree variance is equal to the value computed by the estimated harmonic coefficients. If the hypothesis has to be rejected for a certain degree, the error degree variance is so large that the estimated harmonic coefficients cannot be trusted anymore so that the regularization has to start at that degree. The third method uses the density anomalies by which the disturbing potential is caused resulting from the geopotential model. The density anomalies are well suited to visualize the effects of the higher degree harmonics. In contrast to the base functions of the harmonic coefficients with global support, the density anomalies are expressed by a B-spline surface with local support. Multi-scale representations were applied and the hypotheses tested that the wavelet coefficients are equal to zero. Accepting the hypotheses means that nonsignificant wavelet coefficients were determined which lead to nonsignificant density anomalies. By comparing these anomalies for different regularizations, the degree where to start the regularization is determined. It turns out that beginning the regularization at degree 180, as was done for GOCO02S, is a correct choice.

Quantities like tropospheric zenith delays or station coordinates are repeatedly measured at permanent VLBI or GPS stations so that time series for the quantities at each station are obtained. The covariances of these quantities can be estimated in a multivariate linear model. The covariances are needed for computing uncertainties of results derived from these quantities. The covariance matrix for many permanent stations becomes large, the need for simplifying it may therefore arise under the condition that the uncertainties of derived results still agree. This is accomplished by assuming that the different time series of a quantity like the station height for each permanent station can be combined to obtain one time series. The covariance matrix then follows from the estimates of the auto- and cross-covariance functions of the combined time series. A further approximation is found, if compactly supported covariance functions are fitted to an estimated autocovariance function in order to obtain a covariance matrix which is representative of different kinds of measurements. The simplification of a covariance matrix estimated in a multivariate model is investigated here for the coordinates of points of a grid measured repeatedly by a laserscanner. The approximations are checked by determining the uncertainty of the sum of distances to the points of the grid. To obtain a realistic value for this uncertainty, the covariances of the measured coordinates have to be considered. Three different setups of measurements are analyzed and a covariance matrix is found which is representative for all three setups. Covariance matrices for the measurements of laserscanners can therefore be determined in advance without estimating them for each application.

For reverse engineering, nonuniform rational B-spline surfaces are fitted to measured coordinates of points. Two methods are often applied. The first one consists of simultaneously estimating the unknown control points of the surface. The second method is a special case of the lofting or skinning method where cross-sectional curves are not interpolated but fitted to the measurements, thus obtaining the control points. This approach is considered to give an approximate solution of the first one. However, it is shown here by an analytical proof and confirmed by a numerical example that both methods give identical results. Since the computational complexity of the first approach is the square of the second one, the simultaneous estimation of a large number of control points should be avoided.

Among the Markov chain Monte Carlo methods, the Gibbs sampler has the advantage that it samples from the conditional distributions for each unknown parameter, thus decomposing the sample space. In the case the conditional distributions are not tractable, the Gibbs sampler by means of sampling-importance-resampling is presented here. It uses the prior density function of a Bayesian analysis as the importance sampling distribution. This leads to a fast convergence of the Gibbs sampler as demonstrated by the smoothing with preserving the edges of 3D images of emission tomography.

The use of sampling-based Monte Carlo methods for the computation and propagation of large covariance matrices in geodetic applications is investigated. In particular, the so-called Gibbs sampler, and its use in deriving covariance matrices by Monte Carlo integration, and in linear and nonlinear error propagation studies, is discussed. Modifications of this technique are given which improve in efficiency in situations where estimated parameters are highly correlated and normal matrices appear as ill-conditioned. This is a situation frequently encountered in satellite gravity field modelling. A synthetic experiment, where covariance matrices for spherical harmonic coefficients are estimated and propagated to geoid height covariance matrices, is described. In this case, the generated samples correspond to random realizations of errors of a gravity field model.

To visualize the surface of an object, laser scanners determine the rectangular coordinates of points of a grid on the surface of the object in a local coordinate system. Vertical angles, horizontal angles and distances of a polar coordinate system are measured with the scanning. Outliers generally occur as gross errors in the distances. It is therefore investigated here whether rectangular or polar coordinates are better suited for the detection of outliers. The parameters of a surface represented by a polynomial are estimated in the nonlinear Gauss Helmert (GH) model and in a linear model. Rectangular and polar coordinates are used, and it is shown that the results for both coordinate systems are identical. It turns out that the linear model is sufficient to estimate the parameters of the polynomial surface. Outliers are therefore identified in the linear model by the expectation maximization (EM) algorithm for the variance-inflation model and are confirmed by the EM algorithm for the mean-shift model. Again, rectangular and polar coordinates are used. The same outliers are identified in both coordinate systems.

A robust Kalman filter is derived for rank deficient observation models. The datum for the Kalman filter is introduced at the zero epoch by the choice of a generalized inverse. The robust filter is obtained by Bayesian statistics and by applying a robust M-estimate. Outliers are not only looked for in the observations but also in the updated parameters. The ability of the robust Kalman filter to detect outliers is demonstrated by an example.

Best invariant quadratic unbiased estimates (BIQUE) of the variance and covariance components for a nonlinear Gauss Helmert (GH) model are derived. To detect outliers, the expectation maximization (EM) algorithm based on the variance-inflation model and the mean-shift model is applied, which results in an iterative reweighting least squares. Each step of the iterations for the EM algorithm therefore includes first the iterations for linearizing the GH model and then the iterations for estimating the variance components. The method is applied to fit a surface in three-dimensional space to the three coordinates of points measured, for instance, by a laser scanner. The surface is represented by a polynomial of second degree and the variance components of the three coordinates are estimated. Outliers are detected by the EM algorithm based on the variance-inflation model and identified by the EM algorithm for the mean-shift model.

In three basic points Bayesian statistics differs from traditional statistics. First Bayesian statistics is founded on Bayes’theorem. By this theorem unknown parameters are estimated, confidence regions for the unknown parameters are established and hypotheses for the parameters are tested. Furthermore, Bayesian statistics extends the notion of probability by defining the probability for statements or propositions. The probability is a measure for the plausibility of a statement. Finally, the unknown parameters of Bayesian statistics are random variables. But nevertheless, the unknown parameters can represent constants. There are numerous applications of Bayesian statistics for the analysis of geodetic data, some of them are pointed out.

Different types of present or future satellite data have to be combined by applying appropriate weighting for the determination of the gravity field of the Earth, for instance GPS observations for CHAMP with satellite to satellite tracking for the coming mission GRACE as well as gradiometer measurements for GOCE. In addition, the estimate of the geopotential has to be smoothed or regularized because of the inversion problem. It is proposed to solve these two tasks by Bayesian inference on variance components. The estimates of the variance components are computed by a stochastic estimator of the traces of matrices connected with the inverse of the matrix of normal equations, thus leading to a new method for determining variance components for large linear systems. The posterior density function for the variance components, weighting factors and regularization parameters are given in order to compute the confidence intervals for these quantities. Test computations with simulated gradiometer observations for GOCE and satellite to satellite tracking for GRACE show the validity of the approach.

Chapter 3 dealt with the estimation of the unknown parameters. By means of the variances and covariances of the estimators, measures of the dispersion of the estimators about their expected values and measures of the linear dependency of the estimators were determined With respect to the variances and covariances, additional information shall be obtained in the following by estimating for several parameters the region and for one parameter the interval in which the values of the parameters can be expected with a given probability. This problem is called an interval estimation.

The statistical inference on parameters will be derived in linear models. A linear relation can be treated in a compact and lucid form by vectors and matrices, so that in the following, the definitions and theorems of linear algebra are introduced which will be needed later. The methods of vector spaces will also be discussed. They allow one to use geometric conceptions even then, when the spaces being used are of higher dimensions than the three-dimensional space we are familiar with. Finally, generalized inverses are discussed, by which one can easily change models of full rank for the estimation of parameters to models which are not full rank.

The matrixt distribution is derived and its properties are shown which are applied for the Bayesian inference in the multivariate linear model. By this approach hypothesis tests for the multivariate model are derived which are less sensitive than the tests of the sampling theory. Examples of their application in the analysis of data for the detection of deformations are given.

Using the orthogonal complement likehood function, an iterative procedure for the maximum likelihood estimates of the variance and covariance components is derived. It is shown that these estimates are identical with the reproducing estimates of the locally best invariant quadratic unbiased estimation of variance and covariance components. Successive approximations of the maximum likelihood estimates are given in addition.

For the estimation of parameters in linear models best linear unbiased estimates are derived in case the parameters are random variables. If their expected values are unknown, the well known formulas of least squares adjustment are obtained. If the expected values of the parameters are known, least squares collocation, prediction and filtering are derived. Hence in case of the determination of parameters, a least squares adjustment must precede a collocation because otherwise the collocation gives biased estimates. Since the collocation can be shown to be equivalent to a special case of the least squares adjustment, the variance of unit weight can be estimated for the collocation also. This estimate gives the scale factor for the covariance matrices being used in the collocation. In addition, the methods of testing hypotheses and establishing confidence intervals for the parameters of the least squares adjustment may be applied to the collocation.

Deviations from the null hypothesis, which can be detected by tests in an univariate model, are derived by means of the expected value of the test criterion. As shown, the boundary of the detectable deviations is represented by an hyperellipsoid. The variations of the hyperellipsoid due to the variance of the test criterion are also derived. If each equation of the hypothesis involves only one parameter, the deviations detectable in a plane can be obtained from ellipses which are found by multiplying the axes of the confidence ellipse for two parameters by a certain factor. This method is applied to compute movements to be detected in an analysis of data for recent crustal movements.

The geodetic boundary value problem using the known surface of the earth is defined and shown to have at most one solution. Furthermore it is proved that the solution exists and that its harmonic part can be represented by the potential of a simple layer under the sufficient condition that at the surface of the earth directions are known which lie differentially close to the gradients of the gravity field. The advantages of this boundary value problem are outlined in comparison to the boundary value problem formulated by Molodensky.

The linear models for estimating parameters are so composed that the expected values of the observations, which are carried out for the estimation of the parameters and which represent random variables, are expressed as linear functions of the unknown parameters. The coefficients of the linear functions are assumed to be known. The estimation of parameters in linear models therefore means essentially the estimation of the expected values of the observations.